WO2008109644A2

WO2008109644A2 - Two stage detection for photographic eye artifacts

Info

Publication number: WO2008109644A2
Application number: PCT/US2008/055864
Authority: WO
Inventors: Yury Prilutsky; Alexei Pososin; Mihai Ciuc; Petronel Bigioi; Eran Steinberg; Peter Corcoran
Original assignee: Fotonation Vision Limited
Priority date: 2007-03-05
Filing date: 2008-03-05
Publication date: 2008-09-12
Also published as: KR20090122275A; CN201788344U; KR101049188B1; WO2008109644A3; JP2010520727A

Abstract

The detection of red-eye defects is enhanced in digital images for embedded image acquisition and processing systems. A two-stage redeye filtering system includes a speed optimized filter that performs initial segmentation of candidate redeye regions and optionally applies a speed-optimized set of falsing/verification filters to determine a first set of confirmed redeye regions for correction. Some of the candidate regions which are rejected during the first stage are recorded and re-analyzed during a second stage by an alternative set of analysis-optimized filters to determine a second set of confirmed redeye regions.

Description

IN THE UNITED STATES PATENT AND TRADEMARK OFFICE APPLICATION FOR PATENT

Title: TWO STAGE DETECTION FOR PHOTOGRAPHIC EYE ARTIFACTS

Inventors: Yury Prilutsky, Alexei Pososin, Mihai Ciuc, Petronel Bigioi, Eran Steinberg and Peter Corcoran

Atty. Docket: FN-121-CIP3-PCT

PRIORITY AND RELATED APPLICATIONS

This application claims the benefit of priority to United States provisional patent application no. 60/892,884, filed March 5, 2007, which is incorporated by reference. This application is also a Continuation-in-Part (CIP) of United States patent applications nos.

11/462,035, filed August 2, 2006 and 11/282,954, filed November 18, 2005 and published as US2007/0116379, which are incorporated by reference.

FIELD OF THE INVENTION The present invention relates to digital image processing, and more particularly to a method and apparatus for detection and correction of red-eye defects and other artifacts in an acquired digital image.

BACKGROUND Redeye is the appearance of an unnatural reddish coloration of the pupils of a person appearing in an image captured by a camera with flash illumination. Redeye is caused by light from the flash reflecting off blood vessels in the person's retina and returning to the camera.

A large number of image processing techniques have been proposed to detect and correct redeye in color images. In general, these techniques typically are semi-automatic or automatic. Semi-automatic redeye detection techniques rely on human input. For example, in some semi-automatic redeye reduction systems, a user must manually identify to the system the areas of an image containing redeye before the defects can be corrected.

Many automatic redeye reduction systems rely on a preliminary face detection step before redeye areas are detected. A common automatic approach involves detecting faces in an image and, subsequently, detecting eyes within each detected face. After the eyes are located, redeye is identified based on shape, coloration, and brightness of image areas corresponding to the detected eye locations. In general, face-detection-based automatic redeye reduction techniques have high computation and memory resource requirements. In addition, most of the face detection algorithms are only able to detect faces that are oriented in an upright frontal view; these approaches cannot detect faces that are rotated in-plane or out-of- plane with respect to the image plane.

A typical prior art redeye filter process is illustrated in Fig l(a). An input image is first analyzed by a speed optimized redeye detection stage 100 at a pixel level 103 and segmented into candidate redeye regions 104. A further series of falsing and verification filters 106 are then applied to the candidate regions and a set of confirmed redeye regions 108 is thus determined. A correction filter (pixel modifier) 102 is next applied to the confirmed regions and a final image 112, corrected for redeye, is generated.

Exemplary prior art includes US patent 6,407,777 to DeLuca which discloses in- camera detection and correction of redeye pixels in an acquired digital image; US patent application 2002/0176623 to Steinberg which discloses automated real-time detection and correction of redeye defects optimized for handheld devices; US patent applications 2005/0047655 and 2005/0047656 to Luo et al which disclose detecting and correcting redeye in a digital image and in embedded systems respectively.

Now it is well known that within an image acquisition subsystem such as is embodied in typical digital cameras, the peak computing load and resource requirements occur around the time of image acquisition. Upon receiving an image acquisition request from the user the main embedded processing system must refine the image focus and exposure to achieve an optimal main acquired image; this image, in turn, must be off-loaded from the main optical sensor of the camera and subjected to further image processing to convert it from its raw format (e.g. Bayer) to a conventional color space such as RGB or YCC. Finally the acquired image must be compressed prior to saving it on a removable storage medium such as a compact flash or multimedia card.

The time taken by the camera to recover from the acquisition of a first image and reinitialize itself to capture a second image is known in the industry as the "click-to-click" time. As this is one of the most important parameters for the comparison and marketing of modern digital cameras it vital for manufacturers to minimize said "click-to-click" time. Thus any additional image processing, such as redeye filtering, which is to be added to the main image acquisition chain should be highly optimized for speed of execution in order to minimize its impact on the click-to-click time of the main system. Evidently such a redeye filter must compromise its overall performance in terms of accuracy of detection of redeye defects and quality of image correction. An alternative would be to wait until after the main image has been acquired and perform the redeye filtering at a later time when the camera may execute the filter as a background process, or to perform the redeye filtering off- camera on a desktop PC or printer.

However there are some drawbacks to this approach. Firstly, images will be displayed on the acquiring device, immediately after acquisition, with uncorrected redeye defects; and, when images are accessed in playback mode, there will be a further delay while images are post-processed before an image can be displayed. Both drawbacks would create a negative impression on end users.

Further, as practically all digital cameras store images using lossy compression techniques there are additional disadvantages with respect to image quality as images must be decompressed and recompressed in order to perform the redeye detection and correction processes in playback or background modes. Such loss of image quality may not become apparent until later when a user wishes to print an image and it is too late to reverse the process.

If redeye processing is delayed until the images are loaded onto another device, such as a desktop PC or printer there are further disadvantages. Firstly, important meta-data relating to the acquiring device and its state at the time the image was acquired may not be available to the redeye filter process. A second disadvantage is that this post-processing device must perform redeye filtering on the entire image; where this is an embedded device such as a printer it may, itself, be relatively constrained in terms of CPU cycles and processing resources for its primary post-processing activity and it may be desirable to optimize the performance of the full redeye filter.

SUMMARY OF THE INVENTION

The detection of red-eye defects is enhanced in digital images for embedded image acquisition and processing systems. A two-stage redeye filtering process is provided whereby a speed optimized filter performs the initial segmentation of candidate redeye regions and optionally applies a speed-optimized set of falsing/verification filters to determine a first set of confirmed redeye regions for correction. Some of the candidate regions which are rejected during the first stage are recorded and re-analyzed during a second stage by an alternative set of analysis-optimized filters to determine a second set of confirmed redeye regions. In another embodiment, the first set of confirmed redeye regions may be passed through the stage-two analysis-optimized filters.

In an another embodiment the second stage filter may incorporate an enhanced correction filter which may be optionally applied to the first set of confirmed redeye regions. A two-stage redeye filter is implemented wherein a first redeye filter process, which is optimized for speed is combined with a second redeye process which is optimized for accurate image analysis. Thus the advantages of a redeye filter implemented within the main image acquisition chain are combined with those of a redeye filter implemented in background/playback mode and the disadvantages of these two approaches are mostly eliminated.

For slow platforms (like Fujitsu for Samsung), in playback mode, a "RE lib" is called after, and preferably immediately after, an image is displayed on the screen (e.g., without waiting for the user's command). The camera is capable of multi-tasking and/or can quickly abandon a background task when there is a need to do something else. This gives a user an instantaneous or real-time result when the RE function is executed. The presence of RE may be provided in a picture, e.g., in a flashing region.

A further technique is provided for processing a digital image. Candidate face regions are obtained of an acquired digital image. The candidate face regions are filtered with a first speed optimized filter to produce a first set of candidate red-eye regions. At least a portion of the acquired digital image is encoded. The encoded portion of the acquired digital image is stored in association with said first set of candidate red-eye regions for later image processing of said encoded image.

A speed-optimized set of one or more falsing/verification filters may also be applied. The first set of candidate redeye regions may be passed through an analysis-optimized filter. Candidate face regions rejected by the filtering may be passed through an analysis-optimized filter sometime after the filtering, encoding and storing, such that one or more may be determined as candidate red-eye regions. The first speed optimized filter may be applied upon image acquisition, and the analysis-optimized filter may be applied in a background or playback mode or both. This is applicable to more complex processes such as Golden Eye removal (e.g., a

"two stage" process may be removing red-eye in real time and Golden Eye in the background. A two-stage process may include a fast filter during an acquisition chain and a slower, analysis optimized filter applied afterwards, when the camera is idling, or even on a secondary device. This helps particularly slow machines, and can also be not only predictive for playback but also a background process when a camera is not taking pictures in general. There can be an icon that shows that Red Eye was performed and/or a flashing region on a display.

BRIEF DESCRIPTION OF THE DRAWINGS Fig. l(a) illustrates a typical prior art redeye process; Fig. 1 (b) illustrates a redeye process according to an embodiment; Fig. l(c) illustrates a redeye process according to an alternative embodiment; Fig. 2(a) illustrates an embodiment of the present invention within a digital image acquisition device;

Fig. 2(b) illustrates an embodiment of the present invention wherein the analysis optimized redeye filter is performed on a separate device to the original acquiring device;

Fig. 3 (a) illustrates a process according to an embodiment of the present invention whereby the speed optimized redeye detector is applied to a partially compressed DCT block image;

Fig. 3(b) is a workflow diagram of an illustrative embodiment of an improved in- camera redeye detection means employing a redeye DCT prefilter;

Fig. 3(c) is a workflow diagram of an illustrative embodiment of the redeye DCT prefilter;

Fig. 3(d) illustrates a segmentation step of the redeye DCT prefilter; Fig. 3(e) shows a 4-DCT block neighborhood; Fig. 4(a) illustrates eye regions mapped onto a rectangular grid;

Fig. 4(b) illustrates the approximate color which will be recorded by the DC coefficient of each DCT block after the image of Fig 4(a) is transformed into the DCT domain;

Figs 4(c), 4(d) and 4(e) illustrate the DCT blocks from Fig 4(a) which can be identified with the colors of a redeye candidate region, an eye-white region and a skin color region, respectively, through the use of an inclusive color determining filter method; Fig. 5 illustrates a functional implementation of modified redeye filtering process according to an embodiment;

Fig 6(a) illustrates an example of the original defect region stored in a header and a corrected defect region applied to a main image body; Fig 6(b) illustrates an example of the corrected defect region stored in the header and the original defect region remaining uncorrected in the main image body;

Fig 6(c) illustrates an example of the original defect region and at least one alternative corrected defect region stored in the header and the optimally determined corrected defect region applied to the main image body; and

Figs 7 and 8 illustrate functional implementations of modified redeye filtering processes according to further embodiments.

DETAILED DESCRIPTION OF THE EMBODIMENTS

The detection of red-eye defects is enhanced in digital images for embedded image acquisition and processing systems. A two-stage redeye filtering process is provided whereby a speed optimized filter performs the initial segmentation of candidate redeye regions and optionally applies a speed-optimized set of falsing/verification filters to determine a first set of confirmed redeye regions for correction. Some of the candidate regions which are rejected during the first stage are recorded and re-analyzed during a second stage by an alternative set of analysis-optimized filters to determine a second set of confirmed redeye regions.

In another embodiment, the first set of confirmed redeye regions may be passed through the stage-two analysis-optimized filters.

In an another embodiment the second stage filter may incorporate an enhanced correction filter which may be optionally applied to the first set of confirmed redeye regions.

A two-stage redeye filter is implemented wherein a first redeye filter process, which is optimized for speed is combined with a second redeye process which is optimized for accurate image analysis. Thus the advantages of a redeye filter implemented within the main image acquisition chain are combined with those of a redeye filter implemented in background/playback mode and the disadvantages of these two approaches are mostly eliminated.

One generalized embodiment is illustrated in Fig l(b). An input image 110 is processed by a pixel analyzer 103, segmented into a set of candidate regions 104 and subsequently passed through a set of falsing & verification filters 106. All of these components form a speed optimized redeye detection filter 100 corresponding generally to the filter 100 of Figure 1 (a), except that in the embodiment the filter 100 is modified so that candidate redeye regions which, in the prior art speed optimised redeye filter, would have been ultimately classified as false positives, based on their size or probability being below a predetermined threshold are, instead saved as candidate regions 109 for a subsequent optimized analysis 101.

Thus, in this embodiment the falsing & verification filters 106 generates a set of secondary candidate regions 109 in addition to the set of confirmed redeye regions 108. The set of secondary candidate regions may include members of the original candidate region set 104, which could be neither confirmed nor eliminated by the speed optimized redeye detection process 100. It may also include combined candidate regions in close proximity to each other. This set of candidate regions 109 is saved either in a RAM buffer, or in non-volatile memory depending on the implementation of the embodiment. Where the data is saved in RAM (or volatile) memory, the image acquisition system must apply the second stage redeye filter to the image prior to powering down. The preferred form of storage is in non-volatile memory, or on a removable media card. In other embodiments this data may be stored in the image header with the part-processed image itself.

In this embodiment, a second stage, analysis optimised redeye filter 101 is next applied to the secondary set of candidate regions 109. During the analysis optimized detection process 101, the saved candidate regions 109 are preferably further analyzed at a higher resolution than during the speed optimized process. Thus, the filter 101 includes an analysis optimized set of falsing and verification filters 116, which differ either in their nature or in their operating parameters from the falsing and verification filters 106 employed in the speed optimized analysis. Nonetheless, it will be appreciated that it may be useful to perform one or more intermediate stages of optimized analysis at increasing image resolutions. This will depend on the hardware capabilities of the imaging appliance and the resources available within the image processing subsystems of the imaging appliance.

Second stage analysis may occur in response to a variety of external events. For example, a user may initiate image playback causing this filter 101 to be applied. Alternatively, a camera may signal that it has been idle for a predetermined interval and thus background redeye processing may be initiated. Where a camera can determine its motion, for example, from auto-focus data, it may be assumed that when a camera is idle, for example, where image focus does not change for a predetermined interval and no user input is received, background image processing, including stage-two redeye filtering, may be initiated.

After a set of confirmed redeye regions 118 is determined by the second stage redeye filter 101 , a correction filter (pixel modifier) 102 is applied and these corrected regions are merged 115 with the initial corrected image 112 to generate a final corrected image 113.

An alternative embodiment is illustrated in Fig l(c) which differs from the embodiment of Fig 1 (b) in that a single correction filter (pixel modifier) 102b is applied after the second stage redeye filter 101, rather than merging the initial corrected image 112 with the corrected regions determined by the stage-two filter 101. The filter 102b corrects both the original confirmed redeye regions 108 and the second stage confirmed redeye regions 118 to produce the final corrected image 113. Fig 2(a) illustrates an embodiment of the present invention within a digital image acquisition device. The speed optimized redeye filter 411 may contains both detection 411-1, 411-2 & 411-4 and, optionally, correction 411-3 processes. Similarly, the analysis optimized redeye filter 412, which may operate as a background process 403, performs additional refinements to the initial determinations and corrections of the speed optimized filter 411. Data related to these initial determinations is provided by the redeye filter metadata 410-5 which is stored with the acquired image 410-2 in an image store 410.

Fig 2(b) illustrates a variation on the embodiment of Fig 2(a) wherein the analysis optimized redeye filter is performed on a separate device 400 to the original acquiring device. This may be, for example, a desktop PC, or a printer. In other embodiments the camera may connect directly to a network or web service. The image data transfer means 404a, 404b may be either a point-to-point communications link between the two devices; a removable storage media which is physically exchanged between the two devices, or alternatively both devices may be connected to a common network such as the internet. In other embodiments the redeye filter metadata 410-5 may be incorporated with the main image data 410-2 by adding the metadata to the JPEG header, see Figure 6. It will also be evident that background redeye filters may operate on both the original acquiring device 400 and the separate device 400'. We remark, however, that supporting multiple redeye filters of increasing sophistication requires very complex and detailed metadata to be exchanged and stored with the image being analyzed and corrected.

For an exemplary, non-exhaustive list of some typical filters that may be found in either the speed or analysis-optimized redeye filters 411, 412, see US Application No. 10/976,336 (Reference FN206US).

In the embodiments above, the speed optimized redeye detection 100 is preferably applied to a sub-sampled input image. The confirmed redeye regions 108 from this speed optimized redeye detection 100 are passed to a redeye correction module 102/ 102a. The corrected redeye image 112 can be displayed on a low-resolution viewing screen of a digital camera immediately after the image acquisition process providing the user with a redeye corrected image almost instantly. However, although this initial corrected image 112 may be adequately corrected, for example, where it is a portrait-style image in which a face occupies most of an image or where large high probability red-eye regions exist, it may not be adequately corrected for images including a large groups of persons, where the candidate redeye regions are smaller or less certain. Accordingly, the second analysis optimized redeye filtering process 101 is implemented after image acquisition but prior to final image 113 display on a larger viewer, or image printing. We remark that one advantage of the present invention is that the analysis optimized redeye detection 101 and correction 102 processes may be delayed until such high resolution viewing or printing is desired by the end user.

It should be noted that in the embodiments of Figures 2(a) and 2(b), the sub-sampled versions of the main image or well as uncorrected full size versions of the main image may be provided directly from main image acquisition device hardware 402 rather than needing to explicitly sub-sample a decoded full size main image.

As in the case, of Figure l(c), it will be seen that image correction need not be performed on images within the acquisition chain and can in fact be performed in the background on acquired images for which speed optimised redeye detection has been performed in the acquisition chain. This is advantageous in many image acquisition appliances where image compression is often implemented in hardware as part of the main image acquisition chain 401. In this embodiment, only the detection process is actually performed in the acquisition chain. A speed optimized correction or a full analysis optimized redeye filter may be subsequently selected in playback mode either based on a predetermined setting within the camera, or on a user selection at the time of image playback/ viewing.

In the embodiment of Fig 3(a), an acquired raw image 402 is partially processed 404 before being provided to DCT compression block 408-1. This block essentially provides a sub-sampled version of the acquired image and, although not shown, this can be provided to the image store 410 as explained above. A speed optimized redeye detector 428 is then applied to the partially compressed DCT block image and DCT red-eye candidate regions both corrected and suspected uncorrected regions are output for storage in the store 410. An advantage of applying speed optimised correction at DCT block level, rather than at the pixel level, is that the need for an explicit image subsampling step is avoided, yet the benefits of applying redeye analysis to a sub-sampled image as detailed in the prior art cited herein are retained.

The regions output by the DCT prefilter 428, incorporated in the main image acquisition chain 401, can advantageously allow much of the DCT block stream to be bypassed without being processed when an image is subsequently corrected by a filter such as a background filter module 426. This allows either much faster or more detailed analysis and filtering of the DCT blocks which are determined to require processing by an analysis optimized redeye filter 406. Those skilled in the art will realize that further embodiments are possible which separate aspects of both the DCT prefilter and the otherwise conventional type redeye filter 406 between the main image acquisition chain, 401 and a background redeye detection and correction process 426.

Fig. 3(b) shows in more detail the operation of the redeye DCT prefilter 428. This particular example illustrates how the DCT prefilter can integrate with the main image acquisition, processing and compression chain, 402, 404 and 408 of Fig 3(a). The DCT image to be filtered is first loaded into memory 902 after which the main DCT prefilter 428 is applied. This has three main steps: firstly the DCT blocks of the image are scanned 904 and the relevant DCT coefficients are extracted. Depending on the sophistication of the filter, only the DC components of each DCT block may be utilized in the subsequent analysis. Alternatively, some of the AC components may be extracted in order to allow some texture or sharpness/blur determination as part of the prefilter operation.

In the second principle stage of the DCT prefilter the DCT blocks are segmented and grouped 906 based on a plurality of criteria determined from the coefficients extracted at step 904. Finally a region based analysis is performed 907 in order to determine the final candidate redeye groupings. Next it is determined if there are any valid candidate grouping 908 and if not the normal JPEG compression process is resumed 408-2. If candidate regions are determined 908 then a bounding region is determined for each 910 which is sufficiently large to include various eye -region features which may be used as part of the main prior-art redeye filter process 406 of Fig 3(a). If the certainty of the region being a flash eye defect is high enough, a bounding box region is decompressed to bitmap format 912 and a speed optimised redeye filter chain 914 is applied to correct that region of the main image 914. The corrected regions in bitmap space are next mapped to an integer number of 8x8 block boundaries and are recompressed 918 and subsequently overwritten 920 onto the DCT domain. Finally, normal JPEG compression is resumed 408-2. As mentioned previously each of the corrected region boundaries and suspected region boundaries are output for use in later analysis optimized detection and correction.

Fig 3(c) shows the region based analysis 907 of Fig 3(b) in more detail. Firstly, the DCT coefficients are read 930 from a DCT image in temporary memory store. These coefficients are then preprocessed into a set of criteria tables 932. Each table is essentially a numeric table of size NxM where there are NxM DCT blocks in the image being analyzed. As examples, one such table will contain the red chrominance component normalized to emphasize a colour range associated with flash eye defects and derived from the DC coefficients for the luminance (Y) and red chrominance (Cr) components of each DCT block. Another table may contain differential values derived from neighbouring DCT blocks and used in edge detection; yet another table may contain variance values calculated across a set of neighbouring DCT blocks. Those skilled in the art will realize that as an implementation of the DCT prefilter becomes increasingly sophisticated that multiple additional criteria may be incorporated into the algorithm. After the calculations required for each criteria table have been completed 932 they are copied into temporary storage 933 and the prefilter algorithm will next perform a filtering and segmentation step 907 for each of the plurality of criteria tables. This particular step is further detailed in Fig 3(d) below. Now the prefilter has determined a plurality of sets of DCT block grouping based on the segmentation analysis of a plurality of criteria tables. It is now necessary to sort and analyze these grouping to determine a final set of flash defect candidate regions.

This region-based analysis 936 is comprised of a number of alternative techniques which will be known to those skilled in the art. In particular, we mention that regions may be combined both in inclusive, exclusive and less frequently in mutually exclusive combinations 936-1; an alternative approach to region-based analysis will employ template matching 936-2, one example of which is disclosed US 5,805,727 to Nakano discloses matching a subregion within a DCT image using both coarse and fine template matching techniques based on the DC coefficients of the DCT blocks within the image.

An important component of the region based analysis is a re-segmentation engine 92-6 which is responsible for analyzing larger regions which may, in fact, be two distinct overlapping regions, or clusters of smaller regions which may, in fact, be a single larger region. Then once the region based analysis 936 is completed a final LUT containing the list of determined flash defect candidate regions is obtained and written to system memory.

Fig 3(d) shows the segmentation step 907 of the redeye DCT prefilter in more detail. The next preprocessed criteria table to be processed by the segmentation process is first loaded 950 and the labeling LUT for the region grouping process is initialized 952. Next the current DCT block and DCT block neighbourhoods are initialized 954.

Fig 3(e) shows a diagrammatic representation of a 4-DCT block neighborhood 992, shaded light grey in the figure and containing the three upper DCT blocks and the DCT block to the left of the current DCT block 994, shaded dark grey in the figure. This 4-block neighborhood is used in the labeling algorithm of this exemplary embodiment. A look-up table, LUT, is defined to hold correspondence labels.

Returning to step 954 we see that after initialization is completed the next step for the workflow of Fig 3(d) is to begin a recursive iteration through all the elements of the current criteria table in a raster-scan from top-left to bottom-right. The workflow next determines if the current criteria table value, associated with the current DCT block satisfies membership criteria for a candidate redeye region 958. Essentially this implies that the current criteria table value has properties which are compatible with a flash eye defect. If the current criteria table value satisfies membership criteria for a segment 958, then the algorithm checks for other member DCT blocks in the 4-block neighborhood 960. If there are no other member blocks, then the current block is assigned membership of the current label 980. The LUT is then updated 982 and the current label value is incremented 984. If there are other member blocks in the 4-block neighborhood 960 then the current block is given membership in the segment with the lowest label value 962 and the LUT is updated accordingly 516. After the current block has been labeled as part of a flash-eye defect segment 962 or 980, or has been categorized as not being a member of a candidate defect region during step 958, a test is then performed to determine if it is the last DCT block in the image 966. If the current block is the last block in the image then a final update of the LUT is performed 970. Otherwise the next criteria table value is obtained by incrementing the current block pointer 968 and returning to step 958 and is processed in the same manner. Once the final DCT block is processed and the final LUT completed 970, all of the blocks with segment membership are sorted into a labeled-segment table of potential eye-defect segments 972. Another test is then made to determine if this is the last criteria table to be processed 966 and if that is the case then control is passed to the region based analysis step of Fig 3(c) 936. Otherwise the block segmentor returns to step 950 and loads the next criteria table for processing.

A number of conventional techniques can advantageously be adapted for use within the redeye DCT prefilter. US 5,949,904 to DeIp discloses querying image colors within a DCT block. In particular it allows the determination of colour within the DCT block from the DC coefficient of the DCT alone. Thus from a knowledge of the DC coefficients alone color matching can be achieved. US 6,621,867 to Sazzad et al discloses determining the presence of edges within DCT blocks based on differences between the DC coefficients in neighbouring DCT blocks.

Now additional image qualities such as texture and image sharpness/blur can be determined through an analysis of additional AC components within a DCT block. Examples of such analysis techniques are described in US patent application No. 2004/0120598 to Feng and US patent application No. 2004/0057623 to Schuhurke et al.

Alternative DCT block segmentation techniques may be employed in other embodiments, specifically techniques described in prior art patents US 6,407,777 to DeLuca, US 6,873,743 to Steinberg, and US patent applications 2005/0047655 and 2005/0047656 to Luo et al may be advantageously employed in certain alternative embodiments.

In Fig 4 we show an example of how an outline colour template can be constructed for redeye regions. Fig 4(a) shows an eye regions mapped onto a rectangular grid. Each block of the grid 201 corresponds to an 8x8 pixel block. The main redeye defect 204 is typically surrounded by an iris region 203 and an additional eye-white region 202 and the boundary of the main redeye region, 206 as determined by a conventional redeye filter.

Next, in Fig 4(b) we show the approximate colour which will be recorded by the DC coefficient of each DCT block after the image in Fig 4(a) is transformed into the DCT domain. The colour combinations shown in Fig 4(b) are as follows: R is a reddish hue indicative of a flash-eye defect phenomenon; S is a hue indicative of a skin colour; W: indicates a whitish colour associated with the eye-white region; I: is the Iris colour of the eye which can vary significantly from person to person; WS: indicates a block with mixed skin and eye-white; RW: is a block with mixed redeye and eye white; and RI: has a hue which is a mix of red and the Iris colour. Now if sufficiently inclusive colour filtering is applied to these image blocks it is possible to determine directly from the DC coefficients of the DCT domain image a colour map for a typical redeye. In Fig 4(c) we illustrate the region which will be determined as red if an inclusive colour filter is used. Figs 4(d) and 4(e) illustrate this for eye white and skin colour regions surrounding the flash eye defect region. This data can, for example, be used to build a set of colour templates for a flash eye defect region. By applying other conventional techniques it is possible to determine DCT blocks which contain sharp edges, or transitions between colour regions. This can provide additional information to improve the DCT prefilter detection process.

A potential disadvantage in the embodiment of Fig 3 (a) is that it requires the entire image to be decompressed in order to perform the second-step redeye filtering process. As most cameras use JPEG compression which is lossy it is desirable for certain embodiments to implement a lossless embodiment which allows a two-stage redeye process to be applied within an image acquisition appliance without loss of image quality.

Accordingly we present in Fig 5 a functional implementation of modified redeye filtering process which allows an analysis optimized redeye detection and correction to occur in playback mode, without loss of image quality. This also allows complex post-processing, to be implemented in incremental steps. Thus, when a camera is idle with respect to user activity, yet is still switched on it may load and commence processing of an image. When user activity recommences the camera can recompress and save the image being processed prior to responding to the user. As the embodiment described below allows lossless saving and restoration of a image within the camera, it thus facilitates incremental process of an image which is not limited to redeye, but may be applied likewise to other in-camera methods such as face detection or recognition. Various means of sensing user activity will be known to those skilled in the art. One exemplary means include detecting camera motion and optionally correlating this with other in-camera functions such as the autofocus subsystem and the user-interface subsystem. Many cameras also incorporate a power-saving mode which determines that a camera has been inactive long enough to disable certain main subsystem. When such a mode is activated by user inactivity then additional background image processing can be initiated without interfering with the use of the appliance by the user.

Returning to Fig 5 we illustrate an embodiment of the present invention which incorporates a speed-optimized redeye filter 411 in the main image acquisition chain 401. In this exemplary embodiment the speed optimization of the filter is achieved by implementing a minimal set of falsing and validation filters and no correction process is applied during the main image acquisition chain. In alternative embodiments the speed optimization techniques described in relation to embodiments above may optionally be incorporated or substituted.

After an image is analyzed by this speed optimized redeye filter 411 it is subsequently compressed 427- 1 and stored 410. In addition data relating to the location of candidate redeye regions and false positives is recorded and associated with the stored image.

Now when the camera can initiate background processing, as described above, or when the user enters playback mode and selects an image for viewing it will be partially decompressed 433 from JPEG to DCT block form. As this decompression step is lossless there is no loss of quality to the main image which is temporarily stored in memory and passed to a DCT region decompressor 430. This DCT region decompressor uses the data stored and associated with the original image to determine the specific DCT blocks which contain candidate redeye regions, and, optionally, false positive regions which may benefit from additional detection processing if sufficient time & system resources are available.

Each decompressed DCT region is then incrementally filtered by one or more redeye filters to determine corrections which should be applied to said DCT image block.

In certain embodiments, DCT blocks may be decompressed to bitmap format and filtered as a pixel block. In other embodiments adjacent, non-candidate DCT blocks may be included in the decompression 430 and filtering 412 processes. Once a decompressed DCT block region, which may consist of multiple DCT blocks, has been corrected by the redeye filter 412 then the corrected DCT image segment is passed onto a DCT block matching module 416 which, in addition to checking the alignment of DCT blocks will also check the integration of the corrected DCT blocks within the partially decompressed and temporarily stored DCT block image. When all candidate DCT blocks and any adjacent DCT blocks included in the redeye filter analysis have been corrected they are overwritten onto the partially decompressed and temporarily stored DCT block image by a DCT region overwriter 418 module. The partially decompressed and temporarily stored DCT block image is next passed to the DCT to JPEG image compression module 427- 1 and is losslessly compressed back to JPEG format. Note that in this way the only regions of the image which are decompressed using lossy techniques are those identified by the speed optimized redeye filter 411 in the image acquisition chain. As these image regions are to be corrected the effect of lossy decompression and recompression on them will thus be negligible.

Several alternative embodiments of the invention can be identified. These include (i) saving a copy of the original defect region prior to overwriting the DCT blocks which contain the image defect in the temporary copy of the DCT image. This alternative embodiment supports lossless restoration of the original image. Those skilled in the art will realize that the saved original DCT block region containing the defect can be stored within the header of the JPEG image. One prior art description of such a technique is provided in US 6,298,166 to Ratnakar et al. where watermark data is incorporated in the image. Thus the corrected image can contain a copy of any original uncorrected regions; alternatively (ii) multiple alternative correction algorithms can be employed and these may be temporarily copied for later storage in the JPEG header for later selection by an end user through a user interface, either on the camera or subsequently in a computer based image processing application. Note that in this the overwriting step is optional; if it is used then certain image analysis criteria can be applied as an additional processing step either immediately prior to overwriting, or as an integral part of detecting or correcting red-eye or combinations thereof.

Further aspects of these embodiments are illustrated in Fig 6(a)-(c) which show respectively: Fig 6(a) an example of the original defect region 506 stored in the header 504 and the corrected defect region 508 applied to the main image body 502; Fig 6(b) an example of the corrected defect region 508 stored in the header 504 and the original defect region 506 remaining uncorrected in the main image body 502; Fig 6(c) an example of the original defect region 506 and at least one alternative corrected defect region 508-2 stored in the header 504 and the optimally determined corrected defect region 508-1 applied to the main image body 502. Note that the graphical representations of "corrected" and "uncorrected" eye regions used in Fig 6 is for illustrative purposes only; those skilled in the art will realize that each graphical eye-region actually represents a transformed block of DCT coefficients.

In other embodiments, the performance of the fast red-eye filter can be further improved by selectively applying it to a limited set of regions within the acquired image. As it is generally impractical to implement extensive image analysis during the main image acquisition chain, these regions are preferably determined prior to the initiation of the main image acquisition.

One convenient approach to pre-determine image regions which have a high probability of containing red-eye candidates is to perform pre-processing on a set of preview images. A digital camera can generally include components for enabled the acquisition of such a stream of images, e.g., captured at video rates of 15-30 frames per second (fps) at a lower resolution than that provided by the main image acquisition. A set of 320x240, or QVGA images is typical of many consumer cameras and the size and frame-rate of this preview images stream can normally be adjusted within certain limits.

In two exemplary embodiments, as illustrated in Figures 7 and 8, the digital camera includes a face detector (600) which operates on the preview image stream (410-3). Figure 7 includes a face detector 600, while Figure 8 illustrates a face detector and tracker 600, which includes tracking of detected faces across multiple frames. Figure 8 also illustrates a preview stream acquisition sub-system 620, including image sensor subsystem 610, and a display 605, as well as the face detector and tracker 600 itself.

Face detecting and tracking typically involve two principle modes: (i) a full image search mode to detect (and confirm) new face-candidate regions (601) and (ii) the main tracking mode which predicts and then confirms the new location of existing face- candidates in subsequent frames of the image stream and compiles statistical information relating to each such confirmed candidate region. Both modes can employ a variety of new and/or conventional methods including face detection, skin region segmentation, feature detection including eye and mouth regions, active contour analysis and even non-image based inputs such as directional voice analysis (e.g. US 2005/0147278 to Rui et al which describes a system for Automatic detection and tracking of multiple individuals using multiple cues). As the first mode, hereafter referred to as the "seeding mode" is applied to the entire image it is computationally more intensive and is only applied occasionally - typically every 30-60 image frames. As such, new faces appearing in the image will still be detected within a couple of seconds which is sufficient for most consumer applications. The second mode is preferably applied to every image frame, although not all of the analysis cues may be applied on every frame.

Thus in normal operation only the output(s) from the second operational mode of a face tracker algorithm will be available after every frame of the preview image stream. There are three principle outputs from this second mode: (i) a list of candidate face regions which are confirmed to still contain faces; and/or (ii) a set of data associated with each such confirmed face region including its location within that frame of the image and various additional data determined from a statistical analysis of the history of said confirmed face region; and/or (iii) a predicted location for each such confirmed face region in the next frame of the preview image stream. Note that if item (ii) is used, item (iii) can be optional as sufficient data may be provided by item (ii) for a determination of predicted location.

Now it can be understood that in our embodiment these outputs from the preview face detector (600) enable the speed optimized red-eye detector 411 to be applied selectively to face regions (601) where it is expected that a red-eye defect will be found. A face detector may advantageously be first applied to an image prior to the application of a red-eye filter (see, e.g., US 20020172419 to Lin et al; US 20020126893 to Held et al; US 20050232490 to Itagaki et al and US 20040037460 to Luo et al., which are incorporated by reference). Under normal circumstances, there generally is not sufficient time available during the main image acquisition chain, which is operable within a digital camera, to allow the application of the face detector prior to the application of a red-eye filter. The present embodiment overcomes this disadvantage of the prior art by employing the predictive output of a face tracker module (600). Although the size of the predicted region will typically be larger than the size of the corresponding face region it is still significantly smaller than the size of the entire image. Thus the advantages of faster and more accurate detection can be achieved within a digital camera or embedded image acquisition system without the need to operate a face detector (600) within the main image acquisition chain.

Note that where multiple face candidate regions (601) are tracked, then in certain embodiments, multiple predicted regions will have the speed-optimized red-eye filter applied.

Referring still to Figures 7 and 8, a main image may be acquired, subsampled and stored before being processed by a face detector/tracker 600 as in Figure 7, or a face detector/tracker 600 may be applied prior to storage as in Figure 8 and perhaps in parallel with main image acquisition. A separate "preview stream" of (uncompressed) images may be piped from the main image sensor, independent of the main acquisition, as illustrated at Figure 8. The preview stream may be sent to the camera display 605. It is this preview stream that the face tracker 600 may be operable on and from whence the candidate face regions may be drawn in the embodiment of Figure 8. In using the face tracker 600, a predicted location of a face is obtained from a preview image, and the speed optimized redeye filter may then be operable on the main acquired image (or a subsampled copy thereof). While an exemplary drawings and specific embodiments of the present invention have been described and illustrated, it is to be understood that that the scope of the present invention is not to be limited to the particular embodiments discussed. Thus, the embodiments shall be regarded as illustrative rather than restrictive, and it should be understood that variations may be made in those embodiments by workers skilled in the arts without departing from the scope of the present invention.

In addition, in methods that may be performed according to preferred embodiments herein and that may have been described above, the operations have been described in selected typographical sequences. However, the sequences have been selected and so ordered for typographical convenience and are not intended to imply any particular order for performing the operations, except for those where a particular order may be expressly set forth or where those of ordinary skill in the art may deem a particular order to be necessary.

In addition, all references cited herein as well as the background, invention summary, abstract and brief description of the drawings, as well as US patent no. 6,407,777, 7,315,631 and 7,336,821, and US published patent applications nos. 2005/0041121, 2005/0031224, 2005/0140801, 2006/0204110, 2006/0093212, 2006/0120599, 2007/0110305 and 2006/0140455, and PCT/EP2006/008358, and US patent applications nos. 60/773,714, 60/804,546, 60/865,375, 60/865,622, 60/829,127, 60/829,127, 60/821,165, 60/892,882, 60/945,558, 60/915,669 10/772,767, 11/554,539, 11/464,083, 11/462,035, 11/282,954, 11/027,001, 10/842,244, 11/024,046, 11/233,513, 11/753,098, 11/753,397, 11/766,674, 11/769,206, 11/772,427 and 11/460,218, are all incorporated by reference into the detailed description of the preferred embodiments as disclosing alternative embodiments.

Claims

What is claimed is

1. A method of detecting a red-eye defect within a digital image

(a) acquiring an image including one or more faces including a red eye defect; (b) performing in a first stage an initial segmentation of candidate redeye regions to determine a first set of one or more confirmed redeye regions designated for correction;

(c) correcting the first set of red eye regions for generating a first red eye corrected image;

(d) recording a subset of rejected candidate red eye regions that were rejected for correction in the first stage without full analysis;

(e) re-analyzing the subset of rejected candidate red eye regions in a later second stage to determine a second set of confirmed redeye regions;

(f) correcting the second set of confirmed red eye regions and generating a second red eye corrected image which has the first and second sets of confirmed red eye regions corrected therein; and

(g) electronically storing, transmitting, further processing or editing, or displaying the second red eye corrected image, or combinations thereof.

2. The method of claim 1, further comprising applying a speed-optimized set of one or more falsing/verification filters in the first stage.

3. The method of claim 1, further comprising re-analyzing the first set of confirmed red eye regions in the second stage.

4. The method of claim 1, wherein the initial segmentation is applied upon acquisition of the image and the second later stage occurs in a background or playback mode or both.

5. The method of claim 1, further comprising displaying in real time the first red eye corrected image on a display screen prior to the later second stage without input from a user.

6. An embedded image acquisition and processing system, comprising.

(a) an image acquisition subsystem; (b) a speed-optimized filter that performs in a first stage an initial segmentation of candidate redeye regions detected within an acquired image to determine a first set of one or more confirmed redeye regions designated for correction;

(c) a storage device to record a subset of rejected candidate red eye regions that were rejected for correction in the first stage without full analysis;

(d) an analysis-optimized filter that determines in a later second stage a second set of one or more confirmed red eye regions designated for correction;

(e) a processor for correcting the red eye defect including the first and second sets of confirmed red eye regions within the acquired image and generating a red eye corrected image; and

(f) wherein the red eye corrected image is electronically stored, transmitted, further processed or edited, or displayed, or combinations thereof.

7. The system of claim 6, wherein the speed-optimized filter applies a speed-optimized set of one or more falsing/verification filters.

8. The system of claim 6, wherein the first set of confirmed redeye regions is also passed through the analysis-optimized filter.

9. The system of claim 8, wherein the analysis-optimized filter includes an enhanced correction filter for application to the first set of one or more confirmed red eye regions.

10. The system of claim 6, wherein the speed-optimized filter is applied upon acquisition of the image and the analysis-optimized filter is applied in a background or playback mode or both.

11. The system of claim 6, further comprising a "RE lib" component for displaying in real time an initial red eye corrected image on a display screen prior to the later second stage without input from a user.

12. A method of detecting eye defects within digital images, comprising:

(a) acquiring an image including one or more faces including red and non-red eye defects; (b) performing in a first stage an initial segmentation of candidate redeye regions to determine a first set of one or more confirmed redeye regions designated for correction;

(c) correcting the first set of red eye regions for generating a first red eye corrected image; (d) performing in a later second stage a segmentation of candidate golden eye or white eye regions, or both, to determine a second set of one or more confirmed non-red eye defect regions also designated for correction;

(e) correcting the second set of non-red eye defect regions within the first red eye corrected image for generating an red and non-red eye defect corrected image; (f) electronically storing, transmitting, further processing or editing, or displaying the second red eye corrected image, or combinations thereof.

13. The method of claim 12, further comprising applying a speed-optimized set of one or more falsing/verification filters in the first stage.

14. The method of claim 12, further comprising re-analyzing the first set of confirmed red eye regions in the second stage.

15. The method of claim 12, wherein the initial segmentation is applied upon acquisition of the image and the second later stage occurs in a background or playback mode or both.

16. The method of claim 12, further comprising displaying in real time the first red eye corrected image on a display screen prior to the later second stage without input from a user.

17. An embedded image acquisition and processing system, comprising.

(a) an image acquisition subsystem;

(b) a speed-optimized filter that performs in a first stage an initial segmentation of candidate redeye regions detected within an acquired image to determine a first set of one or more confirmed redeye regions designated for correction; (c) an analysis-optimized filter that determines in a later second stage a second set of one or more confirmed non-red eye regions designated for correction;

(d) one or more processors for correcting the red and non-red eye defects including the first and second sets of confirmed red and non-red eye regions, respectively, within the acquired image and generating a red and non-red eye corrected image; and (e) wherein the red and non-red eye corrected image is electronically stored, transmitted, further processed or edited, or displayed, or combinations thereof.

18. The system of claim 17, wherein the speed-optimized filter applies a speed-optimized set of one or more falsing/verification filters.

19. The system of claim 17, wherein the first set of confirmed redeye regions is also passed through the analysis-optimized filter.

20. The system of claim 19, wherein the analysis-optimized filter includes an enhanced correction filter for application to the first set of one or more confirmed red eye regions.

21. The system of claim 17, wherein the speed-optimized filter is applied upon acquisition of the image and the analysis-optimized filter is applied in a background or playback mode or both.

22. The system of claim 17, further comprising a "RE lib" component for displaying in real time an initial red eye corrected image on a display screen prior to the later second stage without input from a user.

23. A method of processing a digital image comprising:

(a) obtaining candidate face regions of an acquired digital image;

(b) filtering the candidate face regions with a first speed optimized filter to produce a first set of candidate red-eye regions; (c) encoding at least a portion of the acquired digital image; and

(d) storing the encoded portion of the acquired digital image in association with said first set of candidate red-eye regions for later image processing of said encoded image.

24. The method of claim 23, further comprising applying a speed-optimized set of one or more falsing/verification filters.

25. The method of claim 23, further comprising passing the first set of candidate redeye regions through an analysis-optimized filter.

26. The method of claim 23, further comprising passing candidate face regions rejected by the filtering through an analysis-optimized filter sometime after the filtering, encoding and storing.

27. The method of claim 26, wherein the first speed optimized filter is applied upon image acquisition, and the analysis-optimized filter is applied in a background or playback mode or both.

28. A system for processing a digital image, comprising: (a) an image acquisition subsystem;

(b) a face detector for obtaining candidate face regions of an acquired digital image;

(c) a speed optimized filter for filtering the candidate face regions to produce a first set of candidate red-eye regions;

(d) a processor for encoding at least a portion of the acquired digital image; and (e) storing the encoded portion of the acquired digital image in association with said first set of candidate red-eye regions for later image processing of said encoded image.

29. The system of claim 28, further comprising an analysis-optimized filter for filtering at least initially rejected candidate face regions to produce a second set of candidate red-eye regions.

30. The system of claim 28, further comprising an analysis-optimized filter for confirming one or more of the first set of candidate red eye regions sometime after speed-optimized filtering.

31. The system of claim 30, wherein the first speed optimized filter is applied upon image acquisition, and the analysis-optimized filter is applied in a background or playback mode or both.