WO2001022270A2 - Method for indexing and retrieving manufacturing-specific digital imagery based on image content - Google Patents

Method for indexing and retrieving manufacturing-specific digital imagery based on image content Download PDF

Info

Publication number
WO2001022270A2
WO2001022270A2 PCT/US2000/040959 US0040959W WO0122270A2 WO 2001022270 A2 WO2001022270 A2 WO 2001022270A2 US 0040959 W US0040959 W US 0040959W WO 0122270 A2 WO0122270 A2 WO 0122270A2
Authority
WO
WIPO (PCT)
Prior art keywords
vector
manufacturing
image
feature
feature vectors
Prior art date
Application number
PCT/US2000/040959
Other languages
French (fr)
Other versions
WO2001022270A3 (en
Inventor
Regina K. Ferrell
Thomas P. Karnowski
Kenneth W. Tobin, Jr.
Original Assignee
Ut-Battelle, Llc
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Ut-Battelle, Llc filed Critical Ut-Battelle, Llc
Priority to AU13656/01A priority Critical patent/AU1365601A/en
Publication of WO2001022270A2 publication Critical patent/WO2001022270A2/en
Publication of WO2001022270A3 publication Critical patent/WO2001022270A3/en

Links

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/50Information retrieval; Database structures therefor; File system structures therefor of still image data
    • G06F16/58Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually
    • G06F16/583Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually using metadata automatically derived from the content
    • G06F16/5838Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually using metadata automatically derived from the content using colour

Definitions

  • This invention relates to the field of content-based image retrieval and more particularly to a method and apparatus for indexing and retrieving manufacturing-specific digital imagery based upon image content.
  • Description of the Related Art Images are being generated at an ever-increasing rate by sources such as defense and civilian satellites, military reconnaissance and survellance flights, fingerprinting and mug-shot-capturing devices, scientific experimentation, biomedical imaging, architectural and engineering design, and industrial manufacturing.
  • image knowledge structure and an access method for retrieving images are two significant problems in the design of large intelligent image database systems.
  • Content-based image retrieval [CBIR] represents a promising and cutting-edge technology useful in addressing the problem of high-speed image storage and retrieval.
  • CBIR refers to techniques used to index and retrieve images from databases based on their pictorial content.
  • pictorial content is defined by a set of features extracted from an image that describes the color, texture, and/or shape of the entire image or of specific objects.
  • This feature description is used in CBIR to index a database through various means such as distance-based techniques, rule-based decision making, and fuzzy inferencing.
  • the manufacturing environment represents an application area where CBIR technologies have not been extensively studied.
  • the low cost of computer systems, memory, and storage media have resulted in manufacturers collecting and storing more information about the manufacturing process.
  • Much of the data being stored is product imagery collected from automated inspection tools. This imagery contains an historical record of manufacturing events that cause a reduction in product quality. Under the proper circumstances this data can be used to rapidly source product quality issues and improve product yield.
  • Semiconductor manufacturing is representative of an industry that has a mature computer vision component for the inspection of product.
  • Digital imagery for failure analysis is generated between process steps from optical microscopy and laser scattering systems and from confocal, SEM, atomic force microscope and focused ion beam imaging modalities.
  • This data is maintained in a yield management database and used by fabrication engineers to diagnose and source manufacturing problems, verify human assertions regarding the state of the manufacturing process, and to train inexperienced personnel on the wide variety of failure mechanisms observed.
  • the semiconductor industry currently has no direct means of searching the yield management database using image-based queries.
  • the ability to query the fabrication process is based primarily on date, lot, and wafer identification number. Although this approach can be useful, it limits the user's ability to quickly locate historical information.
  • a method for manufacturing-specific CBIR that addresses defect analysis, product quality control, and process understanding in the manufacturing environment.
  • a method and apparatus for indexing and retrieving manufacturing- specific digital imagery based on image content in accordance with the inventive arrangement satisfies the long-felt need of the prior art by providing manufacturing-specific, context based image retrieval in an industrial environment.
  • the inventive method can afford fast access to historical image-based records of similar industrial events so that a corrective action can be quickly taken.
  • the inventive arrangements provide a method and apparatus for employing an image-based query-by-example method to locate and retrieve similar imagry from a database of digital imagery.
  • the inventive arrangements have advantages over all content-based image retrieval systems, and provides a novel and nonobvious system and method for indexing and storing content-based, manufacturing-specific digital imagery for subsequent fast retrieval.
  • a method for indexing and retrieving manufacturing-specific digital images based on image content comprises three steps.
  • First, at least one feature vector can be extracted from a manufacturing-specific digital image stored in an image database.
  • each extracted feature vector corresponds to a particular characteristic of the manufacturing-specific digital image.
  • the extracting step can comprise extracting three independent feature vectors, the three independent feature vectors corresponding to a digital image modality and overall characteristic, a substrate/background characteristic, and an anomaly/defect characteristic, respectively.
  • the extracting step includes generating a defect mask using a detection process selected from the group consisting of thresholding the manufacturing-specific digital image, comparing the manufacturing-specific digital image with a golden template, and comparing the manufacturing-specific digital image with a digital image of a neighboring product; and, extracting a feature vector for the substrate/background characteristic or the anomaly/defect characteristic using the defect mask.
  • a detection process selected from the group consisting of thresholding the manufacturing-specific digital image, comparing the manufacturing-specific digital image with a golden template, and comparing the manufacturing-specific digital image with a digital image of a neighboring product.
  • each extracted feature vector can be indexed in a hierarchical search tree.
  • the hierarchical search tree includes data encapsulating nodes.
  • the nodes can be leaf nodes or level nodes.
  • Each leaf node encapsulates a feature vector.
  • each level node references at least one additional node.
  • each level node encapsulates a vector average defined by the average value of all feature vectors and vector averages encapsulated by nodes referenced by the level node.
  • the using step can comprise the steps of: forming a subset of feature vectors, the subset including at least one feature vector; establishing a main branch in the hierarchical search tree, the main branch having at least one leaf node encapsulating a feature vector included in the subset of feature vectors; and, adding leaf nodes to the hierarchical search tree using a top-down algorithm.
  • each additional leaf node encapsulates a feature vector exclusive of the subset of feature vectors.
  • a manufacturing-specific digital image associated with a feature vector stored in the hierarchicial search tree can be retrieved, wherein the manufacturing-specific digital image has image content comparably related to the image content of the query image. More particularly, the retrieving step includes several steps. First, the query image is converted into at least one query vector corresponding to a particular characteristic of the manufacturing image. Subsequently, a first-level data reduction of feature vectors stored in the hierarchical data structure can be performed based upon the query vector. The first-level data reduction preferably constructs a subset of the feature vectors comparable to the query vector.
  • the relevance feedback can include a user-chosen selection of manufacturing- specific digital images corresponding to the subset of feature vectors.
  • the relevance feedback can include a user-chosen selection of manufacturing- specific digital images corresponding to the subset of feature vectors.
  • three independent feature vectors of manufacturing-based digital imagery can be extracted, the three independent feature vectors corresponding to a digital image modality and overall characteristic, a substrate/background characteristic, and an anomaly/defect characteristic, respectively.
  • at least one prototype vector can be calculated, the prototype vector corresponding to the particular characteristic of the manufacturing-specific digital image.
  • each independent feature vector for each manufacturing-specific digital image in the selection can be logically combined, the logical combination forming the prototype vector for each independent feature vector.
  • a second-level data reduction based upon the prototype vector can be performed, the second-level data reduction resuting in a subset of the feature vectors comparable to the prototype vector, and further comparable to the query vector.
  • the retrieving step can further comprise the step of fetching from the image database a manufacturing-specific digital image defined by an intersection of the three independent feature vectors corresponding to the prototype vector.
  • the inventive method can include a fourth step.
  • the fourth step can include managing the hierarchical search tree.
  • the managing step comprises the steps of: identifying a level node referencing redundant nodes having redundant feature vectors; equating the vector average encapsulated by the level node with all feature vectors and vector averages encapsulated by nodes referenced by the level node; and, purging the hierarchical search tree of the redundant nodes referenced by the level node.
  • FIG. 1 is a schematic representation of a manufacturing-specific CBIR system in accordance with the inventive arrangement.
  • Fig. 2 is a schematic representation of the decomposition and description of a manufacturing image.
  • Figs. 3A - 3B taken together, are a flow chart illustrating a method for indexing manufacturing-specific digital imagery.
  • Fig. 4 is a table listing features used to describe background, defect and global characteristics of an image.
  • Fig. 5 is a schematic illustration of a typical computer vision process for detecting a defect between a test and reference image.
  • Fig. 6 is a tree diagram depicting a hierarchical search tree for use with the inventive method.
  • Fig. 7 is a flow chart illustrating a method for retrieving manufacturing- specific digital imagery from an indexed image database.
  • Fig. 8 is a Venn diagram showing the region of best match between a test query image and relevance images for all image descriptions.
  • FIG. 1 illustrates a manufacturing-specific CBIR system 1 in accordance with the inventive arrangements.
  • a manufacturing-specific CBIR system 1 consists of an image database 5 having a collection of manufacturing- 5 specific images and descriptive data corresponding to each stored image, as well as three basic modules, an image feature extraction module 2, an indexing module 3, and a querying module 4, each performing a required CBIR function.
  • the image database 5 can contain, in addition to the collection of images, other process information that describes the manufacturing process, individual process ⁇ o tools and equipment, and the defect category or type.
  • the image feature extraction module 2 can represent query and database images 8 in terms of a small number of numerical descriptors. Specifically, the image feature extraction module 2 can receive as an input, image 8. The image feature extraction module 2 can survey the image 8 i s deriving a vector of numerical descriptors corresponding to the image 8. Unlike prior CBIR systems, in the preferred embodiment, the manufacturing imagery can be described in terms of three independent sets of characteristics shown in Figure 2. Specifically, the manufacturing image can be characterized in terms of image modality and overall characteristics, substrate-background characteristics,
  • the image feature extraction module 2 pre-processes every image in the image database 5 to generate a series of vectors having these descriptive set of features, each vector
  • the image feature extraction module 2 can store each of the series of vectors in a corresponding feature vector list 7, contained as part of the image database 5.
  • the second module forming the manufacturing-specific CBIR system 1 an indexing module 3, can generate a series of hierarchical search 6, each binary
  • the indexing module 3 can read a vector of numerical descriptors contained in a particular feature vector list 7, the vector corresponding to an image 8 stored in the image database 5. Subsequently, using an unsupervised clustering method, the indexing module 3 can create and insert a node containing the vector into a hierarchical search tree 6 keyed on the same image characteristic as the feature vector list 7. The indexing module 3 can perform the node insertion operation for each feature vector list 7 stored in the image database 5. Thus, each resulting hierarchical search tree 6 can provide for the rapid location of candidate imagery in the image database 5, each hierarchical search tree 6 weighted to a particular image characteristic.
  • the third module forming the manufacturing-specific CBIR system 1 can accept a query image from a user and can return to the user, a collection of similar images stored in the image database 5.
  • the querying module 4 can perform an appropriate first level data reduction based upon the query image's associated vector of numerical descriptors.
  • the image feature extraction module 2 using the query image as an input, can generate the associated vector of numerical descriptors.
  • a very rapid traversal of indexing tree 6 in the first-level data reduction routine can produce a preliminary selection of matching images from the image database 5.
  • a relevance feedback routine contained within the querying module 4 can receive input from the user to further focus the image search to the most relevant images.
  • the user can select several images contained in the preliminary selection of matching images, the selected images having similar charactistics to the query image.
  • a second level data reduction can be performed using the relevance feedback.
  • each image can be combined to provide the user with a vastly reduced set of images having similar characteristics to the query image.
  • a method for indexing and retrieving manufacturing-specific digital imagery based on image content can be further subdivided into two processes: a method for indexing manufacturing-specific digital imagery; and, a method for retrieving manufacturing-specific digital imagery from an indexed image database.
  • FIGS. 3A - 3B taken together, are a flow 5 chart illustrating a method for indexing manufacturing-specific digital imagery.
  • a method for indexing manufacturing-specific digital imagery begins in step 1 0 of figure 3A.
  • the image extraction module 2 can load an image 8 stored in the image database 5.
  • the image 8 needn't be stored in any particular database. Rather, in the preferred embodiment, all images 8 stored in ⁇ o the image database merely are organized in one file directory on a computer.
  • image feature extraction module 2 can extract image features pertinent to the manufacturing process. Specifically, the image feature extraction module 2 can extract modality features, background features and defect features. More particularly, the image feature extraction module 2 can i s define three unique and approximately independent feature vectors of manufacturing-specific digital imagery. The three feature vectors include an anomaly/defect characteristic, a substrate/background characteristic, and an image modality and global characteristic. The defect region can undergo additional masking of morphological features which include an inner mask, which
  • a solid white-on-black mask identifying the defect; an annular mask, consisting of one pixel inside the inner mask border, one pixel on the border, and one pixel outside the border; and, an outer mask, surrounding the outside of the annular mask by a specified number of pixels.
  • the inner mask is required for the feature extraction.
  • the other masks can be generated as 5 needed.
  • the planes of these images for instance red, green, blue, or gray, can be specified.
  • an edge-enhanced image for instance as performed by a Sobel edge detection operation, can be specified.
  • a special set of features can be generated solely from the inner mask.
  • each feature vector as extracted by the image 0 feature extraction module 2 can be grouped into descriptive categories: color, texture, and shape.
  • Each feature contributing to the vector description can be generated by masking the image 8 and plane with an appropriate mask and measuring the feature according to appropriate equations.
  • the table shown in figure 4 lists features measured from manufacturing imagery for CBIR analysis to describe defect, background, and global characteristics. These features can be described in terms of the color, shape, and texture of the area of interest. Still, the architecture is flexible and new descriptive features under continued investigation can be easily added.
  • the manufacturing-specific CBIR system 1 implements a series of mathematical equations to derive each feature vector.
  • One particular feature, the area of the inner mask can be calculated using the following equation:
  • This area feature is useful for measuring the relative size of defects.
  • the perimeter feature is useful in measuring the relative size of defects.
  • the perimeter feature can form a component of other features.
  • the perimeter of the defect can be measured by counting the pixels in the border of the mask.
  • Compactness measures the roundness of a defect. A perfect disc has minimal compactness of 1 . In the manufacturing-specific CBIR system 1 , the compactness of a feature can be measured by the formula
  • Elongation is a measure of the mask morphology. Whereas, a line has a high elongation, a disk has a low elongation. Elongation can be determined by fitting an ellipse to the data object, working with the major and minor axis. Elongation is computed using the following equation:
  • Geometric moments can be computed using the well-known relationship
  • f(x,y) can be replaced with m(x,y), the mask of the defect.
  • this feature assists in defining the defect location. Still, this feature can be of little use if the defect image is always centered or is always randomly located.
  • central moments can be computed from which other descriptive families of moments can be generated. Central moments can be computed by calculating:
  • Invariant moments are features derived from the central moments which can be used as image descriptors. Invariant moments can be particularly useful, for instance in describing the shape of an object while ignoring scale, position, and rotation.
  • the manufacturing-specific CBIR system 1 uses the following formulas to derive invariant moments.
  • ⁇ 4 ( ⁇ 30 + ⁇ i2 ) + ( ⁇ 21 + ⁇ 03 ) 2
  • the system 1 capitalizes on widely-used and automatically generated defect masks which are an inherent component of the defect detection process.
  • a defect mask is a representation of the segmentation of a defect or anomaly residing in a digital image of a manufactured product.
  • the image feature extraction module 2 can develop each feature vector in step 1 6 by isolating the corresponding characteristic with the defect mask.
  • figure 5 illustrates a typical computer vision process used to detect a defect between a test image 70 and a reference image 72.
  • the reference image 72 is compared to the test image 70, the test image 70 first undergoing spatial and intensity registration 74.
  • the resulting product can be further subjected to threshold and filter processing 76.
  • the result is a binary defect mask 78 that segments the anomaly 80 from the expected image components.
  • numerous methods are employed for applying a defect detection mask to detect defects in a manufacturing-specific image, each which can be effectively employed by the system 1 . Examples of known defect detection mask techniques follow.
  • the threshold defect detection masks employs a simple threshold to detect deviations in an expected uniform background.
  • the threshold process usually requires post filtering to remove small nuisance anomalies that arise due to image noise.
  • the simple threshold method has been applied to continuously manufactured web products such as paper and film products, and unpatterned thick and thin-film ceramics.
  • the golden template defect detection technique utilizes a single reference image of a non-defective product which is maintained as a template for subtractive comparison followed by thresholding and filtering .
  • the golden template method has been applied to the production of printed textiles, printed graphics, semiconductors, printed circuit boards, and patterned technical ceramics.
  • Another defect masking technique, the neighborhood template defect detection mask method can use a dynamic reference image which is updated and maintained based on a neighboring product. This dynamic reference image is applied in the same manner as a golden template.
  • the neighborhood template defect detecion mask can be applied to the production of printed textiles, printed graphics, semiconductors, printed circuit boards, and patterned technical ceramics.
  • the CAD specification defect detection mask applies a grey-scale template derived from a CAD design specification. This is used in a similar manner to the golden template. This technique is especially useful for verifying dimensional specifications.
  • the CAD specification technique finds particular application to the production of printed circuit boards, flexible printed circuit product, patterned technical ceramics, and multi-layer packages.
  • non-spatial techniques can be used to measure variations in characteristics such as texture or color to locate and segment anomalous conditions.
  • These non-spatial techniques can be applied to products that exhibit non-geometric but statistically expected structure, for instance, non-woven textiles, unpatterned woven and knit textiles, and specialty paper products.
  • the image feature extraction module 2 can develop each feature vector in step 1 6 by isolating the corresponding characteristic.
  • the inventive method can further employ a defect filling technique to facilitate description of the background region of an image by removing the defect region based on an estimate derived from the surrounding regions.
  • This defect filling method allows the background to be mathematically described so that other similar background geometries and structures and be retrieved.
  • each resulting feature vector can be added to a feature vector list 7.
  • step 20 additional images remain to be characterized by the image extraction module 2, the process can repeat in step 12, wherein the image extraction module 2 can load another image 8 stored in the image database 5.
  • the manufacturing- specific CBIR system 1 can form a hierarchical search tree for sorting and storing each feature vector contained in each feature vector list 7.
  • the manufacturing-specific CBIR system 1 can initiate an indexing process which organizes the data into a hierarchical tree structure 6 5 allowing for rapid retrieval of imagery during the query process.
  • the manufacturing-specific CBIR system 1 employes feature indexing ⁇ a process by which the manufacturing-specific CBIR system 1 organizes feature vectors contained in each feature vector list 7 to facilitate rapid access and retrieval of similar imagery.
  • the indexing technique is critical to the efficient retrieval of ⁇ o similar data from the software system.
  • the manufacturing-specific CBIR system 1 utilizes a sequential, agglomerative, hierarchical, non-overlapping [SAHN] algorithm for sorting the features contained in the feature vector list 7.
  • SAHN sequential, agglomerative, hierarchical, non-overlapping
  • the SAHN algorithm has been used as an investigative tool for unsupervised clustering in pattern recognition problems. i s
  • SAHN algorithms are used to quickly reduce the number of feature vectors, v ⁇ r that must be compared to the query vector, Q v during a retrieval operation.
  • vector pair (v 3 , v,) have the most similar features to one another.
  • the features represented by vector pair (v 2 , v 5 ) also have the most similar features to one another.
  • the next closest pair of vectors (L v 7 ) maintain a lesser degree of similarity than vector pair (v 3 , v ), where L 1 02 is the prototype of (l ⁇ , ⁇ ) at level 1 defined by the vector average, ⁇ v ,v > . 5
  • the illustrated hierarchy when populated with a large number of vectors corresponding to a large image database can exhibit redundancy among similar images.
  • the redundancy measure can be used to drive the short-term usage of new image data or to facilitate data archiving for long-term storage, or data purging of data whose redundancy does not add value to the 0 query process.
  • the Z., prototype can be used to represent redundant images without loss of database value.
  • Each node in the indexing tree 6 is inserted into the indexing tree 6 according this hybrid tree-filling algorithm, unique to a manufacturing-specific CBIR.
  • tree filling is defined as the placement of each descriptive image vector in a location in the indexing tree 6.
  • bottom up filling where each vector in the feature vector list 7 is compared to every other vector to determine the closest match
  • top-down filling where a vector is added to the tree by comparison to the level structure, L compress from the top filling in the tree downward.
  • the bottom up approach is equivalent to a nearest-neighbor algorithm where closest neighbors are joined based on proximity within the n-dimensional feature space of the data.
  • the top down approach can be analogous to a furthest-neighbor algorithm that can result in a similar clustering of the data but also vastly different clustering dependent on the distributions of points in the n-dimensional feature space, for instance elongated groups versus spherical, well separated groups.
  • a nearest-neighbor method the growth of elongated clusters is encouraged while a furthest neighbor approach results in discouraging elongated clusters.
  • the tree in a top-down filling, the tree can be constructed and appended rapidly whereas in bottom-up filling the computational complexity is much greater.
  • operations on an NxN similarity matrix can rapidly become enormous for even a moderately sized image database of N images.
  • appending the indexing tree 6 with new data requires a recalculation of the indexing tree 6.
  • the manufacturing-specific CBIR system 1 employs a strategy for tree filling as represented in figure 3B.
  • the initial tree is built around a sub-set of the image feature database to establish the main branch structure in a nearest-neighbor sense. Tree construction is then completed by rapidly filling in the structure from the top-down with the remaining data.
  • a randomly chosen subset of feature vectors from feature vector list 7 preferably are loaded.
  • the main branch structure of the hierarchical search tree 6 is established using leaf nodes to encapsulate each feature vector in the subset of feature vectors.
  • step 26 the process continues if, in decision step 26, additional feature vectors in the feature vector list 7 remain to be inserted into the hierarchical search tree 6.
  • step 28 the next unprocessed feature vector in the feature vector list 7 is loaded, and, in step 30, a node encapsulating the feature vector is created.
  • step 32 the newly created node can be top-down inserted into the hierarchical search tree 6.
  • step 34 seeding the tree from the bottom up establishes a minimum structure that models the desired nearest-neighbor method. Top-down filling then fills in the branches in rapid order around this basic structure.
  • a query vector, Q v could traverse a branch of the tree that does not lead to the most similar match.
  • a query vector Q v could be more similar to Z. 4 than L 7 , but more similar to Z. 5 than L 4 .
  • a top down search could lead the query down a branch, or partition, that is not the most similar.
  • the manufacturing-specific CBIR system 1 utilizes a multi-level tree searching technique as shown in figure 6.
  • the SAHN algorithm partitions, or clusters, the image data into non- overlapping groups of vectors, for example, each branch of the tree can only refer to a feature vector one time.
  • the vector average stored in a level node 102 at level L is representative of the partitioned vectors 1 06 below that level.
  • the criterion includes at least the number of levels between the top-most branching nodes, for example L 4 through L ⁇ , and possibly more nodes as specified by the user, for example the same number of nodes below the bottom-most node. The result is a higher likelihood of finding the most similar imagery associated with a query.
  • FIG. 7 is a flow chart illustrating a method for retrieving manufacturing- specific digital imagery from an indexed image database. Beginning with step 40, once a feature vector list 7 and hierarchical search tree 6 are established for each relevant feature, a user can perform a query by presenting the manufacturing-specific CBIR system 1 with a query image.
  • the image retrieval process is in actuality a feature vector retrieval process.
  • the image can be submitted as a query after, in step 42, conversion to a series of feature vectors analogous to those defined in the feature vector list 7.
  • a subset of the images can be rapidly obtained, the query image sharing similar feature characteristics.
  • the Euclidean distance definition is applied where
  • an exhaustive algorithmic comparison of the query vector, Q v , to all feature vectors, v sacrifice in the set has an algorithmic complexity of O(N), which would be time consuming at best considering the potential size of the image database and the broad range of queries that a user could perform.
  • SAHN algorithms are used to quickly reduce the number of feature vectors, vdonating that must be compared to the query vector, Q v .
  • a query search is performed by comparing the query vector, Q v , to the topmost prototype pair in the tree, and traversing the most similar branch until the number of vectors below the node equals the number of similar matches requested. In the example of figure 6, this equates to a comparison to L 7 and L A . If Z. 4 is the closest prototype then the next comparison will be to L 2 and L 3 , until the desired number of images is represented below the most similar node. These images can be retrieved for further visualization and analysis during the relevance feedback and analysis stage.
  • step 46 the number of images stored below the present level node 1 02 exceeds a desired number of similar matches requested by the user, in step 48, an additional level of the hierarchical search tree 6 can be traversed. The process can repeat until the number of images stored below the present level node 102 equals the number of similar matches requested.
  • the subset can presented to the user so that other relevant images can be selected.
  • Relevance images are those that are considered by the user to strengthen the query by providing stronger examples of the desired characteristics, for example defect type or substrate structure.
  • the relevance imagery can be used to weight the query process more heavily towards those characteristics that the user is attempting to locate in the image database 5.
  • the relevance process is iterative and the user, in step 52, can select relevant imagery from which a new prototype vector can be calculated.
  • a new initial traversal of the indexing tree 6 is performed.
  • decision step 58 if the number of images stored below the present level exceeds a desired number of similar matches requested by the user, in step 60, an additional level of the indexing tree 6 can be traversed.
  • the process can repeat until the number of images stored below the present level node 1 02 equals the number of similar matches requested.
  • One skilled in the art will recognize that a more accurate search can be performed, at the expense of maximum expediency, by allowing the user to continue to select new relevant imagery after each pass through the hierarchical search tree 6.
  • the manufacturing-specific CBIR system 1 returns all image vectors below some level node L, where the closest match occurs.
  • the images that are returned may contain several other images that have the potential to strengthen, or better describe, the initial query.
  • the search vector, Q v based on these similar images, the user is assigning relevance to the new data through manual association.
  • This relevance information is used to modify Q v , to give a new query vector, Q v ', based on a prototype composed of the original and relevance data, and it is used to modify the tree search criteria.
  • the manufacturing-specific CBIR system 1 in step 62 can present to the user a list of images that approximately match the user's query and relevance requirements. Moreover, the list of images can be displayed with a similarity metric indicating how similar the results are to the original query image. Subsequently, the method can exit in step 64.
  • a second feedback technique treats the relevance data problem as a binary decision by comparing the relevance data set, ⁇ Q v , v a , v b , v c , ... ⁇ to all other data in the L, group as "correct” versus "incorrect” data.
  • This approach is used to reduce the number of decision tree groups required to describe the relevance data set.
  • Figure 8 shows the decision space associated with the indexing process for defect, background, and global descriptions.
  • the retrieval process is considered a logical combination of the eight decision trees shown: three for defect (texture, shape, and color), three for background (texture, shape, and color), and two for global (texture and color).
  • the number of hierarchical search trees required to implement a query is made to be dependent on the relevance data.
  • the inventive method can assist with off-line review and analysis of unclassifiable defects, provide assisted defect library generation for supervised automatic defect classification systems, provide unsupervised classification of defects during early yield learning, and, assist in training yield management personnel.
  • the three unique and approximately independent feature vectors allow the user to perform a broad variety of focused, detailed searches based on specific image queries.
  • inventive method can capitalize on defect detection masks which are an inherent component of the defect detection process in current commercial systems.
  • inventive method can further employ a defect filling technique to facilitate description of the background region of an image by removing the defect region based on an estimate derived from the surrounding regions. This defect filling method allows the background to be mathematically described so that other similar background geometries and structures and be retrieved.
  • the inventive method can apply an unsupervised clustering method such as the hierarchical search tree to achieve a first-level reduction of the image data based on the unique image descriptions of modality, background, and defect.
  • the hierarchical search tree allows for rapid and reliable data reduction based on the query image and provides data for logical combination.
  • a relevance mechanism can be applied for a second-level data reduction which uses a nearest-prototype method that incorporates the user- selected relevant imagery to further reduce the number of images in the subset.
  • the hierarchical search tree 6 structure permits the reduction of image redundancy by controlling the storage, archiving, and purging of image data in a large image database as a function of the manufacturing-specific CBIR system 1 indexing method.

Abstract

A method for indexing and retrieving manufacturing-specific digital images based on image content comprises three steps. First, at least one feature vector can be extracted from a manufacturing-specific digital image stored in an image database. In particular, each extracted feature vector corresponds to a particular characteristic of the manufacturing-specific digital image, for instance, a digital image modality and overall characteristic, a substrate/background characteristic, and an anormaly/defect characteristic. Notably, the extracting step includes generating a defect mask using a detection process. Second, using an unsupervised clustering method, each extracted feature vector can be indexed in a hierarchical search tree. Third, a manufacturing-specific digital image associated with a feature vector stored in the hierarchical search tree can be retrieved, wherein the manufacturing-specific digital image has image content comparably related to the image content of the query image. More particularly, it can include two data reductions, the first performed based upon a query vector extracted from a query image. Subsequently, a user can select relevant images resulting from the first data reduction. From the selection, a prototype vector can be calculated, from which a second-level data reduction can be performed. The second-level data reduction can result in a subset of feature vectors comparable to the prototype vector, and further comparable to the query vector. An additional fourth step can include managing the hierarchical search tree by substituting a vector average for several redundant feature vectors encapsulated by nodes in the hierarchical search tree.

Description

METHOD FOR INDEXING AND RETRIEVING MANUFACTURING-SPECIFIC DIGITAL IMAGERY BASED ON IMAGE CONTENT
CROSS REFERENCE TO RELATED APPLICATIONS
(Not Applicable)
STATEMENT REGARDING FEDERALLY SPONSORED RESEARCH OR
DEVELOPMENT This invention was made with government support under contract DE-
AC05-96OR22464, awarded by the United States Department of Energy to Lockheed Martin Energy Research Corporation. The United States government has certain rights in this invention.
BACKGROUND OF THE INVENTION Technical Field
This invention relates to the field of content-based image retrieval and more particularly to a method and apparatus for indexing and retrieving manufacturing-specific digital imagery based upon image content. Description of the Related Art Images are being generated at an ever-increasing rate by sources such as defense and civilian satellites, military reconnaissance and survellance flights, fingerprinting and mug-shot-capturing devices, scientific experimentation, biomedical imaging, architectural and engineering design, and industrial manufacturing. In consequence, image knowledge structure and an access method for retrieving images are two significant problems in the design of large intelligent image database systems. Content-based image retrieval [CBIR] represents a promising and cutting-edge technology useful in addressing the problem of high-speed image storage and retrieval. Specifically, CBIR refers to techniques used to index and retrieve images from databases based on their pictorial content.
Typically, pictorial content is defined by a set of features extracted from an image that describes the color, texture, and/or shape of the entire image or of specific objects. This feature description is used in CBIR to index a database through various means such as distance-based techniques, rule-based decision making, and fuzzy inferencing. Yet, to date, no significant work has been accomplished to apply these technologies to the manufacturing environment. Notwithstanding, imagery collected from the manufacturing processes have unique characteristics that can be taken advantage of in developing a manufacturing-specific CBIR approach.
The manufacturing environment represents an application area where CBIR technologies have not been extensively studied. The low cost of computer systems, memory, and storage media have resulted in manufacturers collecting and storing more information about the manufacturing process. Much of the data being stored is product imagery collected from automated inspection tools. This imagery contains an historical record of manufacturing events that cause a reduction in product quality. Under the proper circumstances this data can be used to rapidly source product quality issues and improve product yield.
Semiconductor manufacturing is representative of an industry that has a mature computer vision component for the inspection of product. Digital imagery for failure analysis is generated between process steps from optical microscopy and laser scattering systems and from confocal, SEM, atomic force microscope and focused ion beam imaging modalities. This data is maintained in a yield management database and used by fabrication engineers to diagnose and source manufacturing problems, verify human assertions regarding the state of the manufacturing process, and to train inexperienced personnel on the wide variety of failure mechanisms observed. Yet, the semiconductor industry currently has no direct means of searching the yield management database using image-based queries. The ability to query the fabrication process is based primarily on date, lot, and wafer identification number. Although this approach can be useful, it limits the user's ability to quickly locate historical information. For example, if SEM review has determined that a particular defect or pattern problem exists on a wafer, the yield engineer must query on dates, lots, and wafers to find similar historical instances. Although roughly 70% of all space occupied in the database consists of imagery, queries to locate imagery are manual, indirect, tedious, and inefficient. Therefore, this becomes an iterative and slow process that can prove unwieldy in the modern semiconductor environment where a single manufacturing campus having multiple fabrication facilities at one site can generate thousands of images daily. If a query method can be designed that allows the user to look for similar informational content, a faster and more focused result can be achieved. A process for locating similar imagery based on image content, for example the image structure rather than lot number, wafer identification, and date, would result in a reduced time-to-source. Hence, what is needed is a method for manufacturing-specific CBIR that addresses defect analysis, product quality control, and process understanding in the manufacturing environment.
SUMMARY OF THE INVENTION
A method and apparatus for indexing and retrieving manufacturing- specific digital imagery based on image content in accordance with the inventive arrangement satisfies the long-felt need of the prior art by providing manufacturing-specific, context based image retrieval in an industrial environment. In response to an industrial event, the inventive method can afford fast access to historical image-based records of similar industrial events so that a corrective action can be quickly taken. Thus, the inventive arrangements provide a method and apparatus for employing an image-based query-by-example method to locate and retrieve similar imagry from a database of digital imagery. The inventive arrangements have advantages over all content-based image retrieval systems, and provides a novel and nonobvious system and method for indexing and storing content-based, manufacturing-specific digital imagery for subsequent fast retrieval.
A method for indexing and retrieving manufacturing-specific digital images based on image content comprises three steps. First, at least one feature vector can be extracted from a manufacturing-specific digital image stored in an image database. In particular, each extracted feature vector corresponds to a particular characteristic of the manufacturing-specific digital image. The extracting step can comprise extracting three independent feature vectors, the three independent feature vectors corresponding to a digital image modality and overall characteristic, a substrate/background characteristic, and an anomaly/defect characteristic, respectively. Notably, the extracting step includes generating a defect mask using a detection process selected from the group consisting of thresholding the manufacturing-specific digital image, comparing the manufacturing-specific digital image with a golden template, and comparing the manufacturing-specific digital image with a digital image of a neighboring product; and, extracting a feature vector for the substrate/background characteristic or the anomaly/defect characteristic using the defect mask. Moreover, the extracting step can comprise the steps of : distinguishing a defect-region from a non-defect region in the manufacturing- specific digital image; rendering the defect-region similar to the non-defect region based on an estimate derived from a region surrounding the defect-region, the removal of which forms a modified manufacturing-specific digital image; and, extracting a feature vector corresponding to the substrate/background characteristic from the modified manufacturing-specific digital image. Second, using an unsupervised clustering method, each extracted feature vector can be indexed in a hierarchical search tree. Specifically, the hierarchical search tree includes data encapsulating nodes. The nodes can be leaf nodes or level nodes. Each leaf node encapsulates a feature vector. In contrast, each level node references at least one additional node. Moreover, each level node encapsulates a vector average defined by the average value of all feature vectors and vector averages encapsulated by nodes referenced by the level node.
The using step can comprise the steps of: forming a subset of feature vectors, the subset including at least one feature vector; establishing a main branch in the hierarchical search tree, the main branch having at least one leaf node encapsulating a feature vector included in the subset of feature vectors; and, adding leaf nodes to the hierarchical search tree using a top-down algorithm. In the adding step, each additional leaf node encapsulates a feature vector exclusive of the subset of feature vectors.
Third, a manufacturing-specific digital image associated with a feature vector stored in the hierarchicial search tree can be retrieved, wherein the manufacturing-specific digital image has image content comparably related to the image content of the query image. More particularly, the retrieving step includes several steps. First, the query image is converted into at least one query vector corresponding to a particular characteristic of the manufacturing image. Subsequently, a first-level data reduction of feature vectors stored in the hierarchical data structure can be performed based upon the query vector. The first-level data reduction preferably constructs a subset of the feature vectors comparable to the query vector.
Thereafter, relevance feedback can be accepted. Specifically, the relevance feedback can include a user-chosen selection of manufacturing- specific digital images corresponding to the subset of feature vectors. For each manufacturing-specific digital image in the selection, three independent feature vectors of manufacturing-based digital imagery can be extracted, the three independent feature vectors corresponding to a digital image modality and overall characteristic, a substrate/background characteristic, and an anomaly/defect characteristic, respectively. From the selection, at least one prototype vector can be calculated, the prototype vector corresponding to the particular characteristic of the manufacturing-specific digital image. Specifically, where three independent feature vectors of manufacturing-based digital imagery are extracted, each independent feature vector for each manufacturing-specific digital image in the selection can be logically combined, the logical combination forming the prototype vector for each independent feature vector.
Finally, a second-level data reduction based upon the prototype vector can be performed, the second-level data reduction resuting in a subset of the feature vectors comparable to the prototype vector, and further comparable to the query vector. Still, the retrieving step can further comprise the step of fetching from the image database a manufacturing-specific digital image defined by an intersection of the three independent feature vectors corresponding to the prototype vector.
In addition to the three steps, the inventive method can include a fourth step. The fourth step can include managing the hierarchical search tree. In particular, the managing step comprises the steps of: identifying a level node referencing redundant nodes having redundant feature vectors; equating the vector average encapsulated by the level node with all feature vectors and vector averages encapsulated by nodes referenced by the level node; and, purging the hierarchical search tree of the redundant nodes referenced by the level node.
BRIEF DESCRIPTION OF THE DRAWINGS
There are presently shown in the drawings embodiments which are presently preferred, it being understood, however, that the invention is not limited to the precise arrangements and instrumentalities shown. Fig. 1 is a schematic representation of a manufacturing-specific CBIR system in accordance with the inventive arrangement.
Fig. 2 is a schematic representation of the decomposition and description of a manufacturing image.
Figs. 3A - 3B, taken together, are a flow chart illustrating a method for indexing manufacturing-specific digital imagery.
Fig. 4 is a table listing features used to describe background, defect and global characteristics of an image.
Fig. 5 is a schematic illustration of a typical computer vision process for detecting a defect between a test and reference image. Fig. 6 is a tree diagram depicting a hierarchical search tree for use with the inventive method.
Fig. 7 is a flow chart illustrating a method for retrieving manufacturing- specific digital imagery from an indexed image database.
Fig. 8 is a Venn diagram showing the region of best match between a test query image and relevance images for all image descriptions. DETAILED DESCRIPTION OF THE INVENTION
Figure 1 illustrates a manufacturing-specific CBIR system 1 in accordance with the inventive arrangements. Specifically, a manufacturing-specific CBIR system 1 consists of an image database 5 having a collection of manufacturing- 5 specific images and descriptive data corresponding to each stored image, as well as three basic modules, an image feature extraction module 2, an indexing module 3, and a querying module 4, each performing a required CBIR function. The image database 5 can contain, in addition to the collection of images, other process information that describes the manufacturing process, individual process ιo tools and equipment, and the defect category or type.
First, the image feature extraction module 2 can represent query and database images 8 in terms of a small number of numerical descriptors. Specifically, the image feature extraction module 2 can receive as an input, image 8. The image feature extraction module 2 can survey the image 8 i s deriving a vector of numerical descriptors corresponding to the image 8. Unlike prior CBIR systems, in the preferred embodiment, the manufacturing imagery can be described in terms of three independent sets of characteristics shown in Figure 2. Specifically, the manufacturing image can be characterized in terms of image modality and overall characteristics, substrate-background characteristics,
20 and anomoly-defect charcteristics. Moreover, the characteristics used to describe the modality, background, and defect are based on the texture, color, and shape of the image. In the preferred embodiment, the image feature extraction module 2 pre-processes every image in the image database 5 to generate a series of vectors having these descriptive set of features, each vector
25 weighted to a particular characteristic of the stored image. Subsequently, the image feature extraction module 2 can store each of the series of vectors in a corresponding feature vector list 7, contained as part of the image database 5.
The second module forming the manufacturing-specific CBIR system 1 , an indexing module 3, can generate a series of hierarchical search 6, each binary
30 search hierarchical search tree 6 corresponding to a particular characteristic of a stored image. Specifically, the indexing module 3 can read a vector of numerical descriptors contained in a particular feature vector list 7, the vector corresponding to an image 8 stored in the image database 5. Subsequently, using an unsupervised clustering method, the indexing module 3 can create and insert a node containing the vector into a hierarchical search tree 6 keyed on the same image characteristic as the feature vector list 7. The indexing module 3 can perform the node insertion operation for each feature vector list 7 stored in the image database 5. Thus, each resulting hierarchical search tree 6 can provide for the rapid location of candidate imagery in the image database 5, each hierarchical search tree 6 weighted to a particular image characteristic. The third module forming the manufacturing-specific CBIR system 1 , a querying module 4, can accept a query image from a user and can return to the user, a collection of similar images stored in the image database 5. Specifically, the querying module 4 can perform an appropriate first level data reduction based upon the query image's associated vector of numerical descriptors. Significantly, the image feature extraction module 2, using the query image as an input, can generate the associated vector of numerical descriptors. Using the vector of numerical descriptors as a guideline, a very rapid traversal of indexing tree 6 in the first-level data reduction routine can produce a preliminary selection of matching images from the image database 5. Subsequently, a relevance feedback routine contained within the querying module 4 can receive input from the user to further focus the image search to the most relevant images. In particular, the user can select several images contained in the preliminary selection of matching images, the selected images having similar charactistics to the query image. Following the relevance feedback procedure, a second level data reduction can be performed using the relevance feedback. Once the system has produced a reduced set of image descriptions, each image can be combined to provide the user with a vastly reduced set of images having similar characteristics to the query image. A method for indexing and retrieving manufacturing-specific digital imagery based on image content can be further subdivided into two processes: a method for indexing manufacturing-specific digital imagery; and, a method for retrieving manufacturing-specific digital imagery from an indexed image database. Both inventive methods can be developed for use on the Microsoft® Windows® 95/98/NT platform using widely-available development tools, for instance Microsoft Visual C + + ®. Figures 3A - 3B, taken together, are a flow 5 chart illustrating a method for indexing manufacturing-specific digital imagery. A method for indexing manufacturing-specific digital imagery begins in step 1 0 of figure 3A. In step 1 2, the image extraction module 2 can load an image 8 stored in the image database 5. Notably, the image 8 needn't be stored in any particular database. Rather, in the preferred embodiment, all images 8 stored in ιo the image database merely are organized in one file directory on a computer.
In step 1 4, image feature extraction module 2 can extract image features pertinent to the manufacturing process. Specifically, the image feature extraction module 2 can extract modality features, background features and defect features. More particularly, the image feature extraction module 2 can i s define three unique and approximately independent feature vectors of manufacturing-specific digital imagery. The three feature vectors include an anomaly/defect characteristic, a substrate/background characteristic, and an image modality and global characteristic. The defect region can undergo additional masking of morphological features which include an inner mask, which
20 is a solid white-on-black mask identifying the defect; an annular mask, consisting of one pixel inside the inner mask border, one pixel on the border, and one pixel outside the border; and, an outer mask, surrounding the outside of the annular mask by a specified number of pixels. Generally, only the inner mask is required for the feature extraction. Still, the other masks can be generated as 5 needed. In addition, the planes of these images, for instance red, green, blue, or gray, can be specified. Alternatively, an edge-enhanced image, for instance as performed by a Sobel edge detection operation, can be specified. Finally, a special set of features can be generated solely from the inner mask.
The characteristics forming each feature vector as extracted by the image 0 feature extraction module 2 can be grouped into descriptive categories: color, texture, and shape. Each feature contributing to the vector description can be generated by masking the image 8 and plane with an appropriate mask and measuring the feature according to appropriate equations. As an example, the table shown in figure 4 lists features measured from manufacturing imagery for CBIR analysis to describe defect, background, and global characteristics. These features can be described in terms of the color, shape, and texture of the area of interest. Still, the architecture is flexible and new descriptive features under continued investigation can be easily added.
The manufacturing-specific CBIR system 1 implements a series of mathematical equations to derive each feature vector. In these equations, described in detail herein, f(x, y) is the region of the original image masked by m(x, y), where m(x,y) = 1 where the mask is white, for example when overlaying the object of interest. In all other cases, m(x,y) = 0. One particular feature, the area of the inner mask, can be calculated using the following equation:
Area=∑ x Σ y m(x,y)
This area feature is useful for measuring the relative size of defects. Similarly, the perimeter feature is useful in measuring the relative size of defects. Moreover, the perimeter feature can form a component of other features. The perimeter of the defect can be measured by counting the pixels in the border of the mask. Compactness measures the roundness of a defect. A perfect disc has minimal compactness of 1 . In the manufacturing-specific CBIR system 1 , the compactness of a feature can be measured by the formula
(perimeter)1 compactness =
4π(area) Elongation is a measure of the mask morphology. Whereas, a line has a high elongation, a disk has a low elongation. Elongation can be determined by fitting an ellipse to the data object, working with the major and minor axis. Elongation is computed using the following equation:
, Mai or Axis - Minor Axis elongation =
Major Axis + Minor Axis
Geometric moments can be computed using the well-known relationship,
m =∑ ∑ x py qf(x,y)
P
where p and q are non-negative integers. The X centroid, Xc, and Y centroid, Yc, are given by the values
c m c m
OToo '00
In the manufacturing-specific CBIR system 1 , f(x,y) can be replaced with m(x,y), the mask of the defect. Thus, this feature assists in defining the defect location. Still, this feature can be of little use if the defect image is always centered or is always randomly located. To provide more information from the moment features, central moments can be computed from which other descriptive families of moments can be generated. Central moments can be computed by calculating:
m nι μ = ∑ ∑ (x—± l γ(y—SLγf(χty) pq X y woo moo
Invariant moments are features derived from the central moments which can be used as image descriptors. Invariant moments can be particularly useful, for instance in describing the shape of an object while ignoring scale, position, and rotation. The manufacturing-specific CBIR system 1 uses the following formulas to derive invariant moments.
φ4=(μ30 +μi2) +2103)2
Figure imgf000014_0001
^^ o-^K^o^^-^i^os^^^ii ^n ^i^os)
Figure imgf000014_0002
+(3μ1230)(μ0321)[3(μ30 + μ12)2-(μ03 +μ21)2]
Since the invariant moments can become very large, the absolute value of the moments can be taken and a logarithmic scaling can be applied thereto if the moment value exceeds 1 0, for instance, φ = 1 0log(φ).
Finally, the following equations describe features which measure quantities about the gray scale levels within the masked area:
1 2 energy = Σ Σ / (x,y) area x y
average intensity = Σ Σ f[x, y) area x y
contrast
Figure imgf000014_0003
fix,y)\ +fix,y)\ Significantly, and unique to the present invention, the system 1 capitalizes on widely-used and automatically generated defect masks which are an inherent component of the defect detection process. A defect mask is a representation of the segmentation of a defect or anomaly residing in a digital image of a manufactured product. The image feature extraction module 2 can develop each feature vector in step 1 6 by isolating the corresponding characteristic with the defect mask.
For example, figure 5 illustrates a typical computer vision process used to detect a defect between a test image 70 and a reference image 72. In the typical process, the reference image 72 is compared to the test image 70, the test image 70 first undergoing spatial and intensity registration 74. The resulting product can be further subjected to threshold and filter processing 76. The result is a binary defect mask 78 that segments the anomaly 80 from the expected image components. Presently, numerous methods are employed for applying a defect detection mask to detect defects in a manufacturing-specific image, each which can be effectively employed by the system 1 . Examples of known defect detection mask techniques follow. For instance, the threshold defect detection masks employs a simple threshold to detect deviations in an expected uniform background. The threshold process usually requires post filtering to remove small nuisance anomalies that arise due to image noise. The simple threshold method has been applied to continuously manufactured web products such as paper and film products, and unpatterned thick and thin-film ceramics.
Alternatively, the golden template defect detection technique utilizes a single reference image of a non-defective product which is maintained as a template for subtractive comparison followed by thresholding and filtering . The golden template method has been applied to the production of printed textiles, printed graphics, semiconductors, printed circuit boards, and patterned technical ceramics. Another defect masking technique, the neighborhood template defect detection mask method, can use a dynamic reference image which is updated and maintained based on a neighboring product. This dynamic reference image is applied in the same manner as a golden template. Thus, as in the golden template method, the neighborhood template defect detecion mask can be applied to the production of printed textiles, printed graphics, semiconductors, printed circuit boards, and patterned technical ceramics.
In contrast, unlike the previous two actual image-based techniques, the CAD specification defect detection mask applies a grey-scale template derived from a CAD design specification. This is used in a similar manner to the golden template. This technique is especially useful for verifying dimensional specifications. The CAD specification technique finds particular application to the production of printed circuit boards, flexible printed circuit product, patterned technical ceramics, and multi-layer packages.
Finally, non-spatial techniques can be used to measure variations in characteristics such as texture or color to locate and segment anomalous conditions. These non-spatial techniques can be applied to products that exhibit non-geometric but statistically expected structure, for instance, non-woven textiles, unpatterned woven and knit textiles, and specialty paper products.
Using defect masks, the image feature extraction module 2 can develop each feature vector in step 1 6 by isolating the corresponding characteristic. Notably, the inventive method can further employ a defect filling technique to facilitate description of the background region of an image by removing the defect region based on an estimate derived from the surrounding regions. This defect filling method allows the background to be mathematically described so that other similar background geometries and structures and be retrieved. Subsequently, in step 1 8, each resulting feature vector can be added to a feature vector list 7. If in decision step 20, additional images remain to be characterized by the image extraction module 2, the process can repeat in step 12, wherein the image extraction module 2 can load another image 8 stored in the image database 5. However, in decision step 20, when no images 8 remain to be processed, passing through jump circle B to figure 3B, the manufacturing- specific CBIR system 1 can form a hierarchical search tree for sorting and storing each feature vector contained in each feature vector list 7.
Having created a series of feature vector lists 7 for images 8 stored in image database 5, the manufacturing-specific CBIR system 1 can initiate an indexing process which organizes the data into a hierarchical tree structure 6 5 allowing for rapid retrieval of imagery during the query process. Specifically, the manufacturing-specific CBIR system 1 employes feature indexing~a process by which the manufacturing-specific CBIR system 1 organizes feature vectors contained in each feature vector list 7 to facilitate rapid access and retrieval of similar imagery. The indexing technique is critical to the efficient retrieval of ιo similar data from the software system. Unlike existing CBIR systems, the manufacturing-specific CBIR system 1 utilizes a sequential, agglomerative, hierarchical, non-overlapping [SAHN] algorithm for sorting the features contained in the feature vector list 7. Historically, the SAHN algorithm has been used as an investigative tool for unsupervised clustering in pattern recognition problems. i s For the CBIR application, SAHN algorithms are used to quickly reduce the number of feature vectors, vιr that must be compared to the query vector, Qv during a retrieval operation.
Figure 6 shows a schematic representation of a hierarchically ordered set of feature vectors 106, v„ i = 1 ,2, ...9. In the illustrated hierarchy, the vectors
20 forming vector pair (v3, v,) have the most similar features to one another. Similarly, the features represented by vector pair (v2, v5) also have the most similar features to one another. The next closest pair of vectors (L v7) maintain a lesser degree of similarity than vector pair (v3, v ), where L 1 02 is the prototype of (l^,^) at level 1 defined by the vector average, < v ,v > . 5 Still, the illustrated hierarchy, when populated with a large number of vectors corresponding to a large image database can exhibit redundancy among similar images. Consequently, the redundancy measure can be used to drive the short-term usage of new image data or to facilitate data archiving for long-term storage, or data purging of data whose redundancy does not add value to the 0 query process. Specifically, the Z., prototype can be used to represent redundant images without loss of database value. Each node in the indexing tree 6 is inserted into the indexing tree 6 according this hybrid tree-filling algorithm, unique to a manufacturing-specific CBIR. Specifically, tree filling is defined as the placement of each descriptive image vector in a location in the indexing tree 6. When creating the indexing tree 6, two basic tree-filling approaches are possible: "bottom up filling" where each vector in the feature vector list 7 is compared to every other vector to determine the closest match, and "top-down filling" where a vector is added to the tree by comparison to the level structure, L„ from the top filling in the tree downward. From a clustering point of view, the bottom up approach is equivalent to a nearest-neighbor algorithm where closest neighbors are joined based on proximity within the n-dimensional feature space of the data. The top down approach can be analogous to a furthest-neighbor algorithm that can result in a similar clustering of the data but also vastly different clustering dependent on the distributions of points in the n-dimensional feature space, for instance elongated groups versus spherical, well separated groups. In essence, in a nearest-neighbor method, the growth of elongated clusters is encouraged while a furthest neighbor approach results in discouraging elongated clusters. Advantageously, in a top-down filling, the tree can be constructed and appended rapidly whereas in bottom-up filling the computational complexity is much greater. In particular, operations on an NxN similarity matrix, can rapidly become enormous for even a moderately sized image database of N images. Still, in top-down filling, appending the indexing tree 6 with new data requires a recalculation of the indexing tree 6.
To facilitate a nearest-neighbor response as well as rapid tree building and appending, the manufacturing-specific CBIR system 1 employs a strategy for tree filling as represented in figure 3B. The initial tree is built around a sub-set of the image feature database to establish the main branch structure in a nearest-neighbor sense. Tree construction is then completed by rapidly filling in the structure from the top-down with the remaining data. Specifically, in step 22, a randomly chosen subset of feature vectors from feature vector list 7 preferably are loaded. Subsequently, in step 24, the main branch structure of the hierarchical search tree 6 is established using leaf nodes to encapsulate each feature vector in the subset of feature vectors. After the initial tree is built, the process continues if, in decision step 26, additional feature vectors in the feature vector list 7 remain to be inserted into the hierarchical search tree 6. In step 28, the next unprocessed feature vector in the feature vector list 7 is loaded, and, in step 30, a node encapsulating the feature vector is created. In step 32, the newly created node can be top-down inserted into the hierarchical search tree 6. The process repeats in step 26 until no unprocessed feature vectors remain in the feature vector list 7, at which point the indexing algorithm exits in step 34. In theory, seeding the tree from the bottom up establishes a minimum structure that models the desired nearest-neighbor method. Top-down filling then fills in the branches in rapid order around this basic structure.
When searching the tree, it is conceivable that a query vector, Qv, could traverse a branch of the tree that does not lead to the most similar match. For example, in figure 6, a query vector Qv could be more similar to Z.4 than L7, but more similar to Z.5 than L4. In this example, a top down search could lead the query down a branch, or partition, that is not the most similar. Thus, to reduce this probability, the manufacturing-specific CBIR system 1 utilizes a multi-level tree searching technique as shown in figure 6. To assist during the searching process, the SAHN algorithm partitions, or clusters, the image data into non- overlapping groups of vectors, for example, each branch of the tree can only refer to a feature vector one time. As a result, the vector average stored in a level node 102 at level L, is representative of the partitioned vectors 1 06 below that level. At each level of the search, the criterion includes at least the number of levels between the top-most branching nodes, for example L4 through Lη, and possibly more nodes as specified by the user, for example the same number of nodes below the bottom-most node. The result is a higher likelihood of finding the most similar imagery associated with a query.
Subsequent to the generation of each feature vector list 7 and corresponding hierarchical search tree 6, both the feature vector list 7 and the hierarchical search tree 6 are maintained for use during the retrieval process. In particular, as new images 8 are added to the image database 5, each feature vector list 7 and the hierarchical search tree 6 are updated accordingly. As in the tree-filling case, appending new vectors to the hierarchical search tree 6 is also completed using a top-down approach. Figure 7 is a flow chart illustrating a method for retrieving manufacturing- specific digital imagery from an indexed image database. Beginning with step 40, once a feature vector list 7 and hierarchical search tree 6 are established for each relevant feature, a user can perform a query by presenting the manufacturing-specific CBIR system 1 with a query image. The image retrieval process is in actuality a feature vector retrieval process. The image can be submitted as a query after, in step 42, conversion to a series of feature vectors analogous to those defined in the feature vector list 7. By passing the query image through the hierarchical search tree 6, a subset of the images can be rapidly obtained, the query image sharing similar feature characteristics. Specifically, in step 44 a first level of the hierarchical search tree 6 can be traversed by comparing the query vector, Qv to the database of feature vectors, vv i = 0, 1 ,... /-1 (where N is the total number of images in the database) is the L- norm distance metric given by | | Qv -v, \ \ . In the present invention, the Euclidean distance definition is applied where | \ QV -v, \ \ - [(Qv -v)2]v2. Still, an exhaustive algorithmic comparison of the query vector, Qv, to all feature vectors, v„ in the set has an algorithmic complexity of O(N), which would be time consuming at best considering the potential size of the image database and the broad range of queries that a user could perform.
Yet, for the CBIR application, SAHN algorithms are used to quickly reduce the number of feature vectors, v„ that must be compared to the query vector, Qv. A query search is performed by comparing the query vector, Qv, to the topmost prototype pair in the tree, and traversing the most similar branch until the number of vectors below the node equals the number of similar matches requested. In the example of figure 6, this equates to a comparison to L7 and LA. If Z.4 is the closest prototype then the next comparison will be to L2 and L3, until the desired number of images is represented below the most similar node. These images can be retrieved for further visualization and analysis during the relevance feedback and analysis stage. Unlike an exhaustive algorithmic comparison of the query vector, Qv, to all feature vectors, v„ the efficiency in searching the SAHN tree is of 0(log2(N)}, which, for a database of 1 00,000 images would be 6,021 times faster than an exhaustive O(N) search. If in step 46, the number of images stored below the present level node 1 02 exceeds a desired number of similar matches requested by the user, in step 48, an additional level of the hierarchical search tree 6 can be traversed. The process can repeat until the number of images stored below the present level node 102 equals the number of similar matches requested.
As a first feedback technique, in step 50, the subset can presented to the user so that other relevant images can be selected. Relevance images are those that are considered by the user to strengthen the query by providing stronger examples of the desired characteristics, for example defect type or substrate structure. The relevance imagery can be used to weight the query process more heavily towards those characteristics that the user is attempting to locate in the image database 5. The relevance process is iterative and the user, in step 52, can select relevant imagery from which a new prototype vector can be calculated. In step 56, a new initial traversal of the indexing tree 6 is performed. As before, in decision step 58, if the number of images stored below the present level exceeds a desired number of similar matches requested by the user, in step 60, an additional level of the indexing tree 6 can be traversed. The process can repeat until the number of images stored below the present level node 1 02 equals the number of similar matches requested. One skilled in the art will recognize that a more accurate search can be performed, at the expense of maximum expediency, by allowing the user to continue to select new relevant imagery after each pass through the hierarchical search tree 6. When the user submits a query based on the image/vector Qv, the manufacturing-specific CBIR system 1 returns all image vectors below some level node L, where the closest match occurs. The images that are returned may contain several other images that have the potential to strengthen, or better describe, the initial query. By modifying the search vector, Qv, based on these similar images, the user is assigning relevance to the new data through manual association. This relevance information is used to modify Qv, to give a new query vector, Qv', based on a prototype composed of the original and relevance data, and it is used to modify the tree search criteria. As a final result, the manufacturing-specific CBIR system 1 in step 62 can present to the user a list of images that approximately match the user's query and relevance requirements. Moreover, the list of images can be displayed with a similarity metric indicating how similar the results are to the original query image. Subsequently, the method can exit in step 64.
A second feedback technique treats the relevance data problem as a binary decision by comparing the relevance data set, {Qv, va, vb, vc, ...} to all other data in the L, group as "correct" versus "incorrect" data. This approach is used to reduce the number of decision tree groups required to describe the relevance data set. Figure 8 shows the decision space associated with the indexing process for defect, background, and global descriptions. Notably, the retrieval process is considered a logical combination of the eight decision trees shown: three for defect (texture, shape, and color), three for background (texture, shape, and color), and two for global (texture and color). The number of hierarchical search trees required to implement a query is made to be dependent on the relevance data. By comparing the relevance set, {Qv, va, vb, vc, ...}, to all other data below L„ it can be determined that only a subset of hierarchical search trees best match the users needs for the particular query. For example, a set of data may show that background color and global texture are not relevant to the current query. The result is a paring down of the hierarchical search tree set prior to a resubmission of Q . The result of applying these relevance feedback techniques in the query process is the ability to achieve a focused search over a broad range of query categories. In consequence, many different users can perform vastly different queries from system while achieving specific target search goals.
As a result, it is an object of the inventive method to provide end-user support for human-level assertions for sourcing manufacturing problems. Specifically, the inventive method can assist with off-line review and analysis of unclassifiable defects, provide assisted defect library generation for supervised automatic defect classification systems, provide unsupervised classification of defects during early yield learning, and, assist in training yield management personnel. The three unique and approximately independent feature vectors allow the user to perform a broad variety of focused, detailed searches based on specific image queries.
It is a further object of the inventive method to capitalize on defect detection masks which are an inherent component of the defect detection process in current commercial systems. Significantly, no other known CBIR systems take advantage of the defect detection mask to assist in the image description and retrieval process. The inventive method can further employ a defect filling technique to facilitate description of the background region of an image by removing the defect region based on an estimate derived from the surrounding regions. This defect filling method allows the background to be mathematically described so that other similar background geometries and structures and be retrieved.
Finally, the inventive method can apply an unsupervised clustering method such as the hierarchical search tree to achieve a first-level reduction of the image data based on the unique image descriptions of modality, background, and defect. The hierarchical search tree allows for rapid and reliable data reduction based on the query image and provides data for logical combination. Moreover, a relevance mechanism can be applied for a second-level data reduction which uses a nearest-prototype method that incorporates the user- selected relevant imagery to further reduce the number of images in the subset. Advantageously, the hierarchical search tree 6 structure permits the reduction of image redundancy by controlling the storage, archiving, and purging of image data in a large image database as a function of the manufacturing-specific CBIR system 1 indexing method.

Claims

1 1 . A method for indexing and retrieving manufacturing-specific digital images
2 based on image content comprising the steps of:
3 extracting at least one feature vector from a manufacturing-specific digital
4 image stored in an image database, said at least one feature vector
5 corresponding to a particular characteristic of said manufacturing-specific image;
6 using an unsupervised clustering method to index said at least one feature
7 vector in a hierarchical search tree, said hierarchical search tree having data
8 encapsulating nodes, said nodes including leaf nodes, each said leaf node
9 encapsulating a feature vector, said nodes further including level nodes, each ιo said level node referencing at least one additional node, each said level node i i encapsulating a vector average defined as an average value of all feature vectors
12 and vector averages encapsulated by nodes referenced by said level node; and,
13 retrieving a manufacturing-specific image corresponding to a feature
14 vector stored in said hierarchicial search tree, said manufacturing-specific image
15 having image content comparably related to image content of a query image.
ι 2. The method according to claim 1 , wherein the extracting step comprises
2 the step of extracting three independent feature vectors, said three independent
3 feature vectors corresponding to an image modality and overall characteristic, a
4 substrate/background characteristic, and an anomaly/defect characteristic,
5 respectively.
ι 3. The method according to claim 2, wherein the extracting step comprises
2 the steps of:
3 generating a defect mask using a detection process selected from the
4 group consisting of thresholding said manufacturing-specific digital image,
5 comparing said manufacturing-specific digital image with a golden template, and
6 comparing said manufacturing-specific digital image with a digital image of a
7 neighboring product; and, extracting a feature vector for said substrate/background characteristic or said anomaly/defect characteristic using said defect mask.
4. The method according to claim 2, wherein the extracting step comprises the steps of: distinguishing a defect-region from a non-defect region in said manufacturing-specific digital image; rendering said defect-region similar to said non-defect region based on an estimate derived from a region surrounding said defect-region, said rendering forming a modified manufacturing-specific digital image; and, extracting a feature vector corresponding to said substrate/background characteristic from said modified manufacturing-specific digital image.
5. The method according to claim 1 , wherein the using step comprises the steps of: forming a subset of feature vectors, said subset including at least one feature vector; establishing a main branch in said hierarchical search tree, said main branch having at least one leaf node encapsulating a feature vector included in said subset of feature vectors; and, adding leaf nodes to said hierarchical search tree using a top-down algorithm, each additional leaf node encapsulating a feature vector exclusive of said subset of feature vectors.
6. The method according to claim 1 , wherein the retrieving step comprises the steps of: converting a query image into at least one query vector corresponding to said particular characteristic of said manufacturing-specific digital image; performing a first-level data reduction of feature vectors stored in said hierarchical data structure, said first-level data reduction based upon said at least one query vector, said first-level data reduction forming a subset of said feature vectors comparable to said query vector; accepting relevance feedback comprising a user-chosen selection of manufacturing-specific digital images corresponding to said subset of feature vectors; calculating at least one prototype vector from said selection, said prototype vector corresponding to said particular characteristic of said manufacturing-specific digital image; and, performing a second-level data reduction of feature vectors stored in said hierarchical data structure, said second-level data reduction based upon said at least one prototype vector, said second-level data reduction forming a subset of said feature vectors comparable to said at least one prototype vector, and further comparable to said at least one query vector.
7. The method according to claim 6, wherein the accepting step comprises the steps of: for each manufacturing-specific digital image in said selection, extracting three independent feature vectors of manufacturing-based digital imagery, said three independent feature vectors corresponding to a digital image modality and overall characteristic, a substrate/background characteristic, and an anomaly/defect characteristic, respectively; and, logically combining each said independent feature vector for each manufacturing-specific digital image in said selection, said logical combination forming a prototype vector for each said independent feature vector.
8. The method according to claim 7, wherein the retrieving step further comprises the step of fetching from said image database a manufacturing- specific digital image defined by an intersection of said three independent feature vectors corresponding to said at least one prototype vector.
9. The method according to claim 1 , further comprising the step of managing said hierarchical search tree.
1 1 0. The method according to claim 9, wherein the managing step comprises
2 the steps of:
3 identifying a level node referencing redundant nodes having redundant
4 feature vectors;
5 equating the vector average encapsulated by said level node with all
6 feature vectors and vector averages encapsulated by nodes referenced by said
7 level node; and,
8 pruning said hierarchical search tree of said redundant nodes referenced
9 by said level node.
1 1 1 . A computer apparatus programmed with a routine set of instructions
2 stored in a fixed medium, said computer apparatus comprising:
3 means for extracting at least one feature vector from a manufacturing-
4 specific digital image stored in an image database, said at least one feature
5 vector corresponding to a particular characteristic of said manufacturing-specific
6 image;
7 means for using an unsupervised clustering method to index said at least
8 one feature vector in a hierarchical search tree, said hierarchical search tree
9 having data encapsulating nodes, said nodes including leaf nodes, each said leaf ιo node encapsulating a feature vector, said nodes further including level nodes, i i each said level node referencing at least one additional node, each said level
12 node encapsulating a vector average defined as an average value of all feature
13 vectors and vector averages encapsulated by nodes referenced by said level
14 node; and,
15 means for retrieving a manufacturing-specific image corresponding to a
16 feature vector stored in said hierarchicial search tree, said manufacturing-specific
17 image having image content comparably related to image content of a query is image.
1 2. The computer apparatus according to claim 1 1 , wherein the extracting means comprises the means for extracting three independent feature vectors, said three independent feature vectors corresponding to an image modality and overall characteristic, a substrate/background characteristic, and an anomaly/defect characteristic, respectively.
1 3. The computer apparatus according to claim 1 2, wherein the extracting means comprises: means for generating a defect mask using a detection process selected from the group consisting of thresholding said manufacturing-specific digital image, comparing said manufacturing-specific digital image with a golden template, and comparing said manufacturing-specific digital image with a digital image of a neighboring product; and, means for extracting a feature vector for said substrate/background characteristic or said anomaly/defect characteristic using said defect mask.
1 4. The computer apparatus according to claim 1 2, wherein the extracting means comprises the steps of: means for distinguishing a defect-region from a non-defect region in said manufacturing-specific digital image; means for rendering said defect-region similar to said non-defect region based on an estimate derived from a region surrounding said defect-region, said rendering means forming a modified manufacturing-specific digital image; and, means for extracting a feature vector corresponding to said substrate/background characteristic from said modified manufacturing-specific digital image.
1 5. The computer apparatus according to claim 1 1 , wherein the using means comprises: means for forming a subset of feature vectors, said subset including at least one feature vector; means for establishing a main branch in said hierarchical search tree, said main branch having at least one leaf node encapsulating a feature vector included in said subset of feature vectors; and, means for adding leaf nodes to said hierarchical search tree using a top- down algorithm, each additional leaf node encapsulating a feature vector exclusive of said subset of feature vectors.
1 6. The computer apparatus according to claim 1 1 , wherein the retrieving means comprises: means for converting a query image into at least one query vector corresponding to said particular characteristic of said manufacturing-specific digital image; means for performing a first-level data reduction of feature vectors stored in said hierarchical data structure, said first-level data reduction based upon said at least one query vector, said first-level data reduction forming a subset of said feature vectors comparable to said query vector; means for accepting relevance feedback comprising a user-chosen selection of manufacturing-specific digital images corresponding to said subset of feature vectors; means for calculating at least one prototype vector from said selection, said prototype vector corresponding to said particular characteristic of said manufacturing-specific digital image; and, means for performing a second-level data reduction of feature vectors stored in said hierarchical data structure, said second-level data reduction based upon said at least one prototype vector, said second-level data reduction forming a subset of said feature vectors comparable to said at least one prototype vector, and further comparable to said at least one query vector. ι 1 7. The computer according to claim 1 6, wherein the accepting means
2 comprises:
3 for each manufacturing-specific digital image in said selection, means for extracting three independent feature vectors of manufacturing-based digital
5 imagery, said three independent feature vectors corresponding to a digital image
6 modality and overall characteristic, a substrate/background characteristic, and an
7 anomaly/defect characteristic, respectively; and,
8 means for logically combining each said independent feature vector for
9 each manufacturing-specific digital image in said selection, said logical o combination forming a prototype vector for each said independent feature 1 vector.
1 1 8. The computer apparatus according to claim 1 7, wherein the retrieving
2 means further comprises means for fetching from said image database a
3 manufacturing-specific digital image defined by an intersection of said three
4 independent feature vectors corresponding to said at least one prototype vector.
1 1 9. The computer apparatus according to claim 1 1 , further comprising means
2 for managing said hierarchical search tree.
ι 20. The computer apparatus according to claim 1 9, wherein the managing
2 means comprises:
3 means for identifying a level node referencing redundant nodes having redundant feature vectors; means for equating the vector average encapsulated by said level node with all feature vectors and vector averages encapsulated by nodes referenced by said level node; and, means for pruning said hierarchical search tree of said redundant nodes referenced by said level node.
PCT/US2000/040959 1999-09-20 2000-09-20 Method for indexing and retrieving manufacturing-specific digital imagery based on image content WO2001022270A2 (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
AU13656/01A AU1365601A (en) 1999-09-20 2000-09-20 Method for indexing and retrieving manufacturing-specific digital imagery based on image content

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
US09/399,394 US6751343B1 (en) 1999-09-20 1999-09-20 Method for indexing and retrieving manufacturing-specific digital imagery based on image content
US09/399,394 1999-09-20

Publications (2)

Publication Number Publication Date
WO2001022270A2 true WO2001022270A2 (en) 2001-03-29
WO2001022270A3 WO2001022270A3 (en) 2003-12-24

Family

ID=23579340

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/US2000/040959 WO2001022270A2 (en) 1999-09-20 2000-09-20 Method for indexing and retrieving manufacturing-specific digital imagery based on image content

Country Status (3)

Country Link
US (1) US6751343B1 (en)
AU (1) AU1365601A (en)
WO (1) WO2001022270A2 (en)

Cited By (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP1429265A2 (en) * 2002-12-12 2004-06-16 Aston University System and method for coding and retrieval of a CAD drawing from a database
EP1280072A3 (en) * 2001-07-25 2006-01-11 Nec Corporation Image retrieval apparatus and image retrieving method
KR100615738B1 (en) * 2002-02-08 2006-08-25 소니 매뉴펙츄어링 시스템즈 코포레이션 Defect classification/inspection system
US7218772B2 (en) 2002-06-10 2007-05-15 Ut-Battelle Llc Method for non-referential defect characterization using fractal encoding and active contours
US7672976B2 (en) * 2006-05-03 2010-03-02 Ut-Battelle, Llc Method for the reduction of image content redundancy in large image databases
US8140594B2 (en) * 2004-09-17 2012-03-20 Sap Ag Advanced message mapping with sub-object key mapping
CN102542058A (en) * 2011-12-29 2012-07-04 天津大学 Hierarchical landmark identification method integrating global visual characteristics and local visual characteristics
CN104281679A (en) * 2014-09-30 2015-01-14 东软集团股份有限公司 Goods classification method and goods classification device both based on image features
EP3391239A4 (en) * 2015-12-14 2020-02-26 Stats Llc System for interactive sports analytics using multi-template alignment and discriminative clustering
CN111159438A (en) * 2018-11-08 2020-05-15 杭州海康威视数字技术股份有限公司 Data storage and retrieval method, electronic device and storage medium

Families Citing this family (137)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US6671407B1 (en) 1999-10-19 2003-12-30 Microsoft Corporation System and method for hashing digital images
US6952798B2 (en) * 2000-03-30 2005-10-04 Barksdale Jr William W Method for determining the performance of a communications system
CA2307404A1 (en) * 2000-05-02 2001-11-02 Provenance Systems Inc. Computer readable electronic records automated classification system
US7386170B2 (en) * 2000-06-30 2008-06-10 Texas Instruments Incorporated Image object ranking
US7215360B2 (en) * 2001-04-06 2007-05-08 Triveni Digital, Inc. Error propagation tree technology
US6975743B2 (en) * 2001-04-24 2005-12-13 Microsoft Corporation Robust and stealthy video watermarking into regions of successive frames
US6996273B2 (en) * 2001-04-24 2006-02-07 Microsoft Corporation Robust recognizer of perceptually similar content
US6973574B2 (en) * 2001-04-24 2005-12-06 Microsoft Corp. Recognizer of audio-content in digital signals
US7020775B2 (en) 2001-04-24 2006-03-28 Microsoft Corporation Derivation and quantization of robust non-local characteristics for blind watermarking
US7356188B2 (en) * 2001-04-24 2008-04-08 Microsoft Corporation Recognizer of text-based work
US6993180B2 (en) * 2001-09-04 2006-01-31 Eastman Kodak Company Method and system for automated grouping of images
US20030072481A1 (en) * 2001-10-11 2003-04-17 Advanced Micro Devices, Inc. Method for evaluating anomalies in a semiconductor manufacturing process
US7353224B2 (en) * 2001-12-04 2008-04-01 Hewlett-Packard Development Company, L.P. System and method for efficiently finding near-similar images in massive databases
US7283659B1 (en) * 2002-01-09 2007-10-16 Kla-Tencor Technologies Corporation Apparatus and methods for searching through and analyzing defect images and wafer maps
US20040006559A1 (en) * 2002-05-29 2004-01-08 Gange David M. System, apparatus, and method for user tunable and selectable searching of a database using a weigthted quantized feature vector
US7095873B2 (en) 2002-06-28 2006-08-22 Microsoft Corporation Watermarking via quantization of statistics of overlapping regions
US7006703B2 (en) * 2002-06-28 2006-02-28 Microsoft Corporation Content recognizer via probabilistic mirror distribution
US7266553B1 (en) * 2002-07-01 2007-09-04 Microsoft Corporation Content data indexing
US20040024755A1 (en) * 2002-08-05 2004-02-05 Rickard John Terrell System and method for indexing non-textual data
US20040034633A1 (en) * 2002-08-05 2004-02-19 Rickard John Terrell Data search system and method using mutual subsethood measures
US20040087042A1 (en) * 2002-08-12 2004-05-06 Bruno Ghyselen Method and apparatus for adjusting the thickness of a layer of semiconductor material
JP2004198265A (en) * 2002-12-18 2004-07-15 Dainippon Printing Co Ltd Visual inspection/selection method of processed product, and visual inspection/selection system of processed product
US6885977B2 (en) * 2002-12-20 2005-04-26 Applied Materials, Inc. System to identify a wafer manufacturing problem and method therefor
US6958103B2 (en) * 2002-12-23 2005-10-25 Kimberly-Clark Worldwide, Inc. Entangled fabrics containing staple fibers
MY138544A (en) * 2003-06-26 2009-06-30 Neuramatix Sdn Bhd Neural networks with learning and expression capability
US7831832B2 (en) * 2004-01-06 2010-11-09 Microsoft Corporation Digital goods representation based upon matrix invariances
US20050165690A1 (en) * 2004-01-23 2005-07-28 Microsoft Corporation Watermarking via quantization of rational statistics of regions
US7770014B2 (en) * 2004-04-30 2010-08-03 Microsoft Corporation Randomized signal transforms and their applications
US20060072778A1 (en) * 2004-09-28 2006-04-06 Xerox Corporation. Encoding invisible electronic information in a printed document
US7397584B2 (en) * 2004-09-28 2008-07-08 Xerox Corporation Encoding invisible electronic information in a printed document
US8838591B2 (en) 2005-08-23 2014-09-16 Ricoh Co., Ltd. Embedding hot spots in electronic documents
US7812986B2 (en) * 2005-08-23 2010-10-12 Ricoh Co. Ltd. System and methods for use of voice mail and email in a mixed media environment
US9384619B2 (en) 2006-07-31 2016-07-05 Ricoh Co., Ltd. Searching media content for objects specified using identifiers
US8086038B2 (en) 2007-07-11 2011-12-27 Ricoh Co., Ltd. Invisible junction features for patch recognition
US7991778B2 (en) 2005-08-23 2011-08-02 Ricoh Co., Ltd. Triggering actions with captured input in a mixed media environment
US8825682B2 (en) * 2006-07-31 2014-09-02 Ricoh Co., Ltd. Architecture for mixed media reality retrieval of locations and registration of images
US8176054B2 (en) 2007-07-12 2012-05-08 Ricoh Co. Ltd Retrieving electronic documents by converting them to synthetic text
US8856108B2 (en) 2006-07-31 2014-10-07 Ricoh Co., Ltd. Combining results of image retrieval processes
US8600989B2 (en) 2004-10-01 2013-12-03 Ricoh Co., Ltd. Method and system for image matching in a mixed media environment
US7970171B2 (en) * 2007-01-18 2011-06-28 Ricoh Co., Ltd. Synthetic image and video generation from ground truth data
US8949287B2 (en) 2005-08-23 2015-02-03 Ricoh Co., Ltd. Embedding hot spots in imaged documents
US8144921B2 (en) * 2007-07-11 2012-03-27 Ricoh Co., Ltd. Information retrieval using invisible junctions and geometric constraints
US8276088B2 (en) * 2007-07-11 2012-09-25 Ricoh Co., Ltd. User interface for three-dimensional navigation
US8156427B2 (en) 2005-08-23 2012-04-10 Ricoh Co. Ltd. User interface for mixed media reality
US8521737B2 (en) 2004-10-01 2013-08-27 Ricoh Co., Ltd. Method and system for multi-tier image matching in a mixed media environment
US7702673B2 (en) 2004-10-01 2010-04-20 Ricoh Co., Ltd. System and methods for creation and use of a mixed media environment
US8156116B2 (en) 2006-07-31 2012-04-10 Ricoh Co., Ltd Dynamic presentation of targeted information in a mixed media reality recognition system
US8335789B2 (en) 2004-10-01 2012-12-18 Ricoh Co., Ltd. Method and system for document fingerprint matching in a mixed media environment
US8005831B2 (en) 2005-08-23 2011-08-23 Ricoh Co., Ltd. System and methods for creation and use of a mixed media environment with geographic location information
US8369655B2 (en) * 2006-07-31 2013-02-05 Ricoh Co., Ltd. Mixed media reality recognition using multiple specialized indexes
US9530050B1 (en) 2007-07-11 2016-12-27 Ricoh Co., Ltd. Document annotation sharing
US8332401B2 (en) 2004-10-01 2012-12-11 Ricoh Co., Ltd Method and system for position-based image matching in a mixed media environment
US8195659B2 (en) 2005-08-23 2012-06-05 Ricoh Co. Ltd. Integration and use of mixed media documents
US9405751B2 (en) 2005-08-23 2016-08-02 Ricoh Co., Ltd. Database for mixed media document system
US9171202B2 (en) 2005-08-23 2015-10-27 Ricoh Co., Ltd. Data organization and access for mixed media document system
US9373029B2 (en) * 2007-07-11 2016-06-21 Ricoh Co., Ltd. Invisible junction feature recognition for document security or annotation
US8184155B2 (en) * 2007-07-11 2012-05-22 Ricoh Co. Ltd. Recognition and tracking using invisible junctions
US8385589B2 (en) 2008-05-15 2013-02-26 Berna Erol Web-based content detection in images, extraction and recognition
US8156115B1 (en) 2007-07-11 2012-04-10 Ricoh Co. Ltd. Document-based networking with mixed media reality
US8868555B2 (en) * 2006-07-31 2014-10-21 Ricoh Co., Ltd. Computation of a recongnizability score (quality predictor) for image retrieval
US8510283B2 (en) 2006-07-31 2013-08-13 Ricoh Co., Ltd. Automatic adaption of an image recognition system to image capture devices
US7282374B1 (en) * 2004-11-03 2007-10-16 Advanced Micro Devices, Inc. Method and apparatus for comparing device and non-device structures
US20060184462A1 (en) * 2004-12-10 2006-08-17 Hawkins Jeffrey C Methods, architecture, and apparatus for implementing machine intelligence and hierarchical memory systems
US7739208B2 (en) * 2005-06-06 2010-06-15 Numenta, Inc. Trainable hierarchical memory system and method
US8219940B2 (en) * 2005-07-06 2012-07-10 Semiconductor Insights Inc. Method and apparatus for removing dummy features from a data structure
US20070070365A1 (en) * 2005-09-26 2007-03-29 Honeywell International Inc. Content-based image retrieval based on color difference and gradient information
US8285744B2 (en) * 2005-09-30 2012-10-09 Rockwell Automation Technologies, Inc. Indexing and searching manufacturing process related information
US20070076869A1 (en) * 2005-10-03 2007-04-05 Microsoft Corporation Digital goods representation based upon matrix invariants using non-negative matrix factorizations
CA2636429C (en) 2006-01-27 2013-09-24 Unz.Org, Llc Presenting digitized content on a network
US8078618B2 (en) 2006-01-30 2011-12-13 Eastman Kodak Company Automatic multimode system for organizing and retrieving content data files
US20070192267A1 (en) * 2006-02-10 2007-08-16 Numenta, Inc. Architecture of a hierarchical temporal memory based system
US20080208966A1 (en) * 2007-02-28 2008-08-28 Numenta, Inc. Hierarchical Temporal Memory (HTM) System Deployed as Web Service
US8732098B2 (en) 2006-02-10 2014-05-20 Numenta, Inc. Hierarchical temporal memory (HTM) system deployed as web service
US7941389B2 (en) * 2006-02-10 2011-05-10 Numenta, Inc. Hierarchical temporal memory based system including nodes with input or output variables of disparate properties
US8243999B2 (en) 2006-05-03 2012-08-14 Ut-Battelle, Llc Method and system for the diagnosis of disease using retinal image content and an archive of diagnosed human patient data
US20070280526A1 (en) * 2006-05-30 2007-12-06 Irfan Malik Determining Information about Defects or Binning Defects Detected on a Wafer after an Immersion Lithography Process is Performed on the Wafer
US9063952B2 (en) 2006-07-31 2015-06-23 Ricoh Co., Ltd. Mixed media reality recognition with image tracking
US8201076B2 (en) * 2006-07-31 2012-06-12 Ricoh Co., Ltd. Capturing symbolic information from documents upon printing
US8489987B2 (en) * 2006-07-31 2013-07-16 Ricoh Co., Ltd. Monitoring and analyzing creation and usage of visual content using image and hotspot interaction
US9020966B2 (en) 2006-07-31 2015-04-28 Ricoh Co., Ltd. Client device for interacting with a mixed media reality recognition system
US8676810B2 (en) * 2006-07-31 2014-03-18 Ricoh Co., Ltd. Multiple index mixed media reality recognition using unequal priority indexes
US9176984B2 (en) * 2006-07-31 2015-11-03 Ricoh Co., Ltd Mixed media reality retrieval of differentially-weighted links
US8073263B2 (en) 2006-07-31 2011-12-06 Ricoh Co., Ltd. Multi-classifier selection and monitoring for MMR-based image recognition
JP4929928B2 (en) * 2006-08-31 2012-05-09 富士通株式会社 Database management program, database management apparatus, and database management method
US20080118160A1 (en) * 2006-11-22 2008-05-22 Nokia Corporation System and method for browsing an image database
EP2087437A4 (en) * 2006-11-28 2010-10-27 Numenta Inc Group-based temporal pooling
US7831108B2 (en) * 2006-12-13 2010-11-09 Adobe Systems Incorporated Universal front end for masks, selections, and paths
US7941392B2 (en) * 2007-02-28 2011-05-10 Numenta, Inc. Scheduling system and method in a hierarchical temporal memory based system
WO2008106615A1 (en) * 2007-02-28 2008-09-04 Numenta, Inc. Spatio-temporal learning algorithms in hierarchical temporal networks
US8112367B2 (en) * 2007-02-28 2012-02-07 Numenta, Inc. Episodic memory with a hierarchical temporal memory based system
JP5139716B2 (en) * 2007-05-16 2013-02-06 キヤノン株式会社 Image search apparatus and image search method
WO2009006231A1 (en) * 2007-06-29 2009-01-08 Numenta, Inc. Hierarchical temporal memory system with enhanced inference capability
US20090138500A1 (en) * 2007-10-12 2009-05-28 Yuan Zhiqiang Method of compact display combined with property-table-view for a complex relational data structure
US20090116413A1 (en) * 2007-10-18 2009-05-07 Dileep George System and method for automatic topology determination in a hierarchical-temporal network
US8175984B2 (en) * 2007-12-05 2012-05-08 Numenta, Inc. Action based learning
GB2457267B (en) * 2008-02-07 2010-04-07 Yves Dassas A method and system of indexing numerical data
US8175985B2 (en) 2008-03-19 2012-05-08 Numenta, Inc. Plugin infrastructure for hierarchical temporal memory (HTM) system
US7983998B2 (en) 2008-03-21 2011-07-19 Numenta, Inc. Feedback in group based hierarchical temporal memory system
US8407166B2 (en) * 2008-06-12 2013-03-26 Numenta, Inc. Hierarchical temporal memory system with higher-order temporal pooling capability
US7962500B2 (en) * 2008-10-24 2011-06-14 Yahoo! Inc. Digital image retrieval by aggregating search results based on visual annotations
US7843237B2 (en) * 2008-11-17 2010-11-30 Infineon Technologies Austria Ag Circuit arrangement for actuating a transistor
US8195582B2 (en) * 2009-01-16 2012-06-05 Numenta, Inc. Supervision based grouping of patterns in hierarchical temporal memory (HTM)
US8385660B2 (en) 2009-06-24 2013-02-26 Ricoh Co., Ltd. Mixed media reality indexing and retrieval for repeated content
US8488873B2 (en) * 2009-10-07 2013-07-16 Apple Inc. Method of computing global-to-local metrics for recognition
CA2774957C (en) * 2009-10-09 2018-06-05 Edgenet, Inc. Automatic method to generate product attributes based solely on product images
US11651277B2 (en) 2010-03-15 2023-05-16 Numenta, Inc. Sparse distributed representation for networked processing in predictive system
WO2011115854A1 (en) 2010-03-15 2011-09-22 Numenta, Inc. Temporal memory using sparse distributed representation
US8861844B2 (en) 2010-03-29 2014-10-14 Ebay Inc. Pre-computing digests for image similarity searching of image-based listings in a network-based publication system
US9792638B2 (en) 2010-03-29 2017-10-17 Ebay Inc. Using silhouette images to reduce product selection error in an e-commerce environment
US8406566B1 (en) 2010-05-27 2013-03-26 Adobe Systems Incorporated Methods and apparatus for soft edge masking
US8412594B2 (en) 2010-08-28 2013-04-02 Ebay Inc. Multilevel silhouettes in an online shopping environment
AU2014206199B2 (en) * 2010-08-28 2016-08-25 Ebay Inc. Search system using multi-level silhouettes
EP2428795A1 (en) * 2010-09-14 2012-03-14 Siemens Aktiengesellschaft Apparatus and method for automatic inspection of through-holes of a component
US8612441B2 (en) * 2011-02-04 2013-12-17 Kodak Alaris Inc. Identifying particular images from a collection
US9058331B2 (en) 2011-07-27 2015-06-16 Ricoh Co., Ltd. Generating a conversation in a social network based on visual search results
EP2737432A4 (en) 2011-07-29 2015-03-25 Hewlett Packard Development Co Incremental image clustering
US8798362B2 (en) 2011-08-15 2014-08-05 Hewlett-Packard Development Company, L.P. Clothing search in images
US8504570B2 (en) 2011-08-25 2013-08-06 Numenta, Inc. Automated search for detecting patterns and sequences in data using a spatial and temporal memory system
US8825565B2 (en) 2011-08-25 2014-09-02 Numenta, Inc. Assessing performance in a spatial and temporal memory system
US8645291B2 (en) 2011-08-25 2014-02-04 Numenta, Inc. Encoding of data for processing in a spatial and temporal memory system
CN103827856A (en) 2011-09-27 2014-05-28 惠普发展公司,有限责任合伙企业 Retrieving visual media
US9159021B2 (en) 2012-10-23 2015-10-13 Numenta, Inc. Performing multistep prediction using spatial and temporal memory system
US9251582B2 (en) 2012-12-31 2016-02-02 General Electric Company Methods and systems for enhanced automated visual inspection of a physical asset
WO2014148176A1 (en) * 2013-03-19 2014-09-25 株式会社smart-FOA Information-gathering system, information-gathering method, and storage medium
JP2016529612A (en) * 2013-08-02 2016-09-23 エモティエント インコーポレイテッド Filters and shutters based on image emotion content
JP6235868B2 (en) * 2013-10-31 2017-11-22 株式会社smart−FOA Information collection system
US10318878B2 (en) 2014-03-19 2019-06-11 Numenta, Inc. Temporal processing scheme and sensorimotor information processing
US20150308920A1 (en) * 2014-04-24 2015-10-29 Honeywell International Inc. Adaptive baseline damage detection system and method
JP2014197412A (en) * 2014-06-12 2014-10-16 トムソン ライセンシングThomson Licensing System and method for similarity search of images
US10769772B2 (en) * 2015-05-21 2020-09-08 Corning Incorporated Methods for inspecting cellular articles
US10564626B2 (en) * 2016-01-29 2020-02-18 Sandvik Intellectual Property Ab Planning of computer numerical control (CNC) machining operations with the aid of a digital computer
US10043259B2 (en) 2016-07-25 2018-08-07 PT Papertech Inc. Facilitating anomaly detection for a product having a pattern
US10558887B2 (en) * 2017-12-04 2020-02-11 Adobe Inc. Digital image search based on arbitrary image features
US20190228306A1 (en) * 2018-01-21 2019-07-25 Stats Llc Methods for Detecting Events in Sports using a Convolutional Neural Network
US11216506B1 (en) * 2018-05-17 2022-01-04 Shutterstock, Inc. Image querying to surface undiscovered images
US10949702B2 (en) * 2019-04-16 2021-03-16 Cognizant Technology Solutions India Pvt. Ltd. System and a method for semantic level image retrieval
US11681922B2 (en) 2019-11-26 2023-06-20 Numenta, Inc. Performing inference and training using sparse neural network

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO1999023578A1 (en) * 1997-10-31 1999-05-14 International Business Machines Corporation Multidimensional data clustering and dimension reduction for indexing and searching
US5915250A (en) * 1996-03-29 1999-06-22 Virage, Inc. Threshold-based comparison

Family Cites Families (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US5579471A (en) 1992-11-09 1996-11-26 International Business Machines Corporation Image query system and method
US5793888A (en) 1994-11-14 1998-08-11 Massachusetts Institute Of Technology Machine learning apparatus and method for image searching
US5963662A (en) * 1996-08-07 1999-10-05 Georgia Tech Research Corporation Inspection system and method for bond detection and validation of surface mount devices
US5815198A (en) * 1996-05-31 1998-09-29 Vachtsevanos; George J. Method and apparatus for analyzing an image to detect and identify defects
US6292582B1 (en) * 1996-05-31 2001-09-18 Lin Youling Method and system for identifying defects in a semiconductor
US5819288A (en) 1996-10-16 1998-10-06 Microsoft Corporation Statistically based image group descriptor particularly suited for use in an image classification and retrieval system
US6108586A (en) * 1997-03-31 2000-08-22 Hitachi, Ltd. Fraction defective estimating method, system for carrying out the same and recording medium
US6330353B1 (en) * 1997-12-18 2001-12-11 Siemens Corporate Research, Inc. Method of localization refinement of pattern images using optical flow constraints
US6546137B1 (en) * 1999-01-25 2003-04-08 Siemens Corporate Research, Inc. Flash system for fast and accurate pattern localization

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US5915250A (en) * 1996-03-29 1999-06-22 Virage, Inc. Threshold-based comparison
WO1999023578A1 (en) * 1997-10-31 1999-05-14 International Business Machines Corporation Multidimensional data clustering and dimension reduction for indexing and searching

Cited By (13)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP1280072A3 (en) * 2001-07-25 2006-01-11 Nec Corporation Image retrieval apparatus and image retrieving method
US7236652B2 (en) 2001-07-25 2007-06-26 Nec Corporation Image retrieval apparatus and image retrieving method
KR100615738B1 (en) * 2002-02-08 2006-08-25 소니 매뉴펙츄어링 시스템즈 코포레이션 Defect classification/inspection system
US7218772B2 (en) 2002-06-10 2007-05-15 Ut-Battelle Llc Method for non-referential defect characterization using fractal encoding and active contours
EP1429265A3 (en) * 2002-12-12 2005-10-19 Aston University System and method for coding and retrieval of a CAD drawing from a database
EP1429265A2 (en) * 2002-12-12 2004-06-16 Aston University System and method for coding and retrieval of a CAD drawing from a database
US8140594B2 (en) * 2004-09-17 2012-03-20 Sap Ag Advanced message mapping with sub-object key mapping
US7672976B2 (en) * 2006-05-03 2010-03-02 Ut-Battelle, Llc Method for the reduction of image content redundancy in large image databases
CN102542058A (en) * 2011-12-29 2012-07-04 天津大学 Hierarchical landmark identification method integrating global visual characteristics and local visual characteristics
CN104281679A (en) * 2014-09-30 2015-01-14 东软集团股份有限公司 Goods classification method and goods classification device both based on image features
CN104281679B (en) * 2014-09-30 2018-04-06 东软集团股份有限公司 Commodity classification method and device based on characteristics of image
EP3391239A4 (en) * 2015-12-14 2020-02-26 Stats Llc System for interactive sports analytics using multi-template alignment and discriminative clustering
CN111159438A (en) * 2018-11-08 2020-05-15 杭州海康威视数字技术股份有限公司 Data storage and retrieval method, electronic device and storage medium

Also Published As

Publication number Publication date
WO2001022270A3 (en) 2003-12-24
AU1365601A (en) 2001-04-24
US6751343B1 (en) 2004-06-15

Similar Documents

Publication Publication Date Title
US6751343B1 (en) Method for indexing and retrieving manufacturing-specific digital imagery based on image content
US7672976B2 (en) Method for the reduction of image content redundancy in large image databases
US9002072B2 (en) System for detection of non-uniformities in web-based materials
US10650508B2 (en) Automatic defect classification without sampling and feature selection
US6606623B1 (en) Method and apparatus for content-based image retrieval with learning function
Mueen et al. Multilevel feature extraction and X-ray image classification
TW202347396A (en) Computer implemented method for the detection and classification of anomalies in an imaging dataset of a wafer, and systems making use of such methods
Wu et al. Content-based image retrieval using growing hierarchical self-organizing quadtree map
Gao et al. An improved XGBoost based on weighted column subsampling for object classification
Chen et al. Semi-supervised framework for wafer defect pattern recognition with enhanced labeling
Tobin et al. Content-based image retrieval for semiconductor process characterization
Gorisse et al. Scalable active learning strategy for object category retrieval
Dijaya et al. Corn Plant Disease Identification Using SURF-based Bag of Visual Words Feature
Grivei SVMIRE-an open source SVM image retrieval with relevance feedback system for earth observation data classification
Chang et al. Image content analysis using modular RBF neural network
Glasgow et al. Case-based reasoning for molecular scene analysis
Tobin Jr et al. Image retrieval in the industrial environment
Kalaiarasan et al. Diagnosis of Skin Cancer with Its Stages Using Multiclass CNN Technique
Tobin et al. The use of historical defect imagery for yield learning
Păvăloi et al. Can keypoint descriptors be used in outdoor image classification?
Bruno et al. Topological multi-contour decomposition for image analysis and image retrieval
Karnowski et al. Content-based image retrieval for semiconductor manufacturing
Rahman SOM Aided Visualization and Retrieval of Flower Database Using Tree-Structured Representation
Vatsal et al. Semantic Scene Classification for Image Annotation
Tambouratzis Image segmentation with the SOLNN unsupervised logic neural network

Legal Events

Date Code Title Description
AK Designated states

Kind code of ref document: A2

Designated state(s): AE AG AL AM AT AU AZ BA BB BG BR BY CA CH CN CU CZ DE DK DM DZ EE ES FI GB GD GE GH GM HR HU ID IL IN IS JP KE KG KP KR KZ LC LK LR LS LT LU LV MA MD MG MK MN MW MX NO NZ PL PT RO RU SD SE SG SI SK SL TJ TM TR TT TZ UA UG UZ VN YU ZA ZW

AL Designated countries for regional patents

Kind code of ref document: A2

Designated state(s): GH GM KE LS MW MZ SD SL SZ TZ UG ZW AM AZ BY KG KZ MD RU TJ TM AT BE CH CY DE DK ES FI FR GB GR IE IT LU MC NL PT SE BF BJ CF CG CI CM GA GN GW ML MR NE SN TD TG

121 Ep: the epo has been informed by wipo that ep was designated in this application
DFPE Request for preliminary examination filed prior to expiration of 19th month from priority date (pct application filed before 20040101)
122 Ep: pct application non-entry in european phase
NENP Non-entry into the national phase

Ref country code: JP