CA1210153A - Scale-space filtering - Google Patents
Scale-space filteringInfo
- Publication number
- CA1210153A CA1210153A CA000454114A CA454114A CA1210153A CA 1210153 A CA1210153 A CA 1210153A CA 000454114 A CA000454114 A CA 000454114A CA 454114 A CA454114 A CA 454114A CA 1210153 A CA1210153 A CA 1210153A
- Authority
- CA
- Canada
- Prior art keywords
- scale
- events
- signal
- space
- phenomena
- Prior art date
- Legal status (The legal status 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 status listed.)
- Expired
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/08—Feature extraction
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Bioinformatics & Computational Biology (AREA)
- Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Cheminformatics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Image Analysis (AREA)
- Processing Or Creating Images (AREA)
- Image Processing (AREA)
- Optical Radar Systems And Details Thereof (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
SCALE-SPACE FILTERING
ABSTRACT OF THE DISCLOSURE
Information indicative discrete events of interest imbedded in raw data are globally classified, or filtered with respect to scale and changes in the scale of observation to effect intelligent perception of phenomena. Large scale components are classified as events while small scale components of an identified event designate points of occurrence of the event. The invention has broad application in artificial intelli-gence and signal processing wherein perceived discrete events, including minima, maxima, inflections, cusps, and discontinuities have a significance other than as noise. Methods and apparatus are described which utilize the subject invention in signal processing applications.
ABSTRACT OF THE DISCLOSURE
Information indicative discrete events of interest imbedded in raw data are globally classified, or filtered with respect to scale and changes in the scale of observation to effect intelligent perception of phenomena. Large scale components are classified as events while small scale components of an identified event designate points of occurrence of the event. The invention has broad application in artificial intelli-gence and signal processing wherein perceived discrete events, including minima, maxima, inflections, cusps, and discontinuities have a significance other than as noise. Methods and apparatus are described which utilize the subject invention in signal processing applications.
Description
I! C~ LO~L~;3 C . .~
S CALE - S PACE F I LTER I NG
ACKGROUND OF THE INVENTION
Field O' Invention This invention relates ~o artificial intelli-gence and to signal processing. In particular, the invention relates to the classification or filtering of inforrnation derived from primitive perception for use in developing a description of observed events i~ the 1~ absence of prior knowledge about the significance of events.
Traditionally, phenomena such as physical - c,bjects, terrain, and the like are considered to have u~ique, 'true" lengths, areas, volumes, tangents, and other metric properties. This "true" value is ihe one o~tained in the limit, as some factor of scale of measurement goes to zero. For example, arc length is defined as the lim-t o~ a polygonal approximation to a cur~e, as the length OI each side of the polygon goes to zero.
The ~raditional notion o~ length doesn't hold for a broad clas~ of irregular 2hysicaL structures anZ
~rocesses. The problem is tllat when measuring the lergtn of, e.g., a coastline by polygonal appro~imation, the measured length increases without apparent ~ound as the length of the poiygonal side decreases (intllitively, because the smaller "yardstick" conforms to ever smaller sub-ba~s and sub-peninsulas.~ Since the limiting value is undefined (or at least uninteresting), there can be no meaninyful evaluation of metric proper~
ties without reference to the scale at which the properties a-e measured (i.e., the length of the "yardstick," or so~e oiher parameter of scale.) ~4 ~2~
It has therefore been assumed that the choice of scale and range of scale is fundamentally a tradeoff between resolution and noise. A small scale observa-tion can resolve closely spaced events but is sensitive to noise. A large scale observa-tion blurs events but is less sensitive to noise. The scale threshold, which :
is often predetermined based on empirical factors for which there is no analytical model, requently deter-mines how events are defined. The inherent ambiguity in scale parameterization is often a serious problem, because there is usually no simple principled basis for setting the scale parameter. Thus, if a ~uery is directed to how long, how wide, or how steeply inclined is an irregular physical structure or the function of a physical measurement process, there is no single "correct" answer, since the answer changes, some-times sisnificantly, as -the scale of measurement is changed.
The ambiguity of scale cannot in general be eliminated by any simple process; rather the problem as herein formulated is to reduce the ambiguity as much as possible and to present the alternatives in an organized, discrete, symbolic fashion.
One technigue for circumventing the problem without really solving it has been simply to select a scale of observation based on a guess at a model for the relationship between the data collected and the conclusions desired. There nevertheless remains a need for a technique for analyzing raw continuous da-ta to develop a discrete syrnbolic description which can be manipulated according to known analytical signal models.
It is to be under~tood that measurement of any sort requires at least some primitive scale parame-teriæa-tion in order to maintain a measurement of properties in a range of scale. The size of any neighborhood determines the scale of description. For example, a linear scale may be able to usefully present events for display with resp~ct to an independent variable between zero and ten with a resolution of one part in one hundred. This represents a uniformly weighted scale range of three orders of magnitude.
Various property measurement techniques may be applied within any scale range as, for example, local fitting of analytical functions to data, spatial averaging and the like. Each property measurement techni~ue has in common the continuous parameterization by scale of some local measure on the signal.
As used herein a scale-dependent measure on a signal might be defined as a function of the signal, a function of a location on the signal, and a function of a parameter of scale, such that values in the signal contribute to the function~s output to a diminishing degree as they become more distant from the original location; and such that the rate at which the values' cont-ibutions diminish with distance decreases as the value of the parameter of scale increases A scale~
dependent measure is for example the result of convolving the signal with a gaussian, using the gaussian's standard deviation ("width"~ as the parameter o~ scale~
Description Of The Prior Art There are no known analytical signal models capable of describing different physical events which occur a-t different physical scales, namely, events which appear like noise ~lithin one scale but correspond to interesting events within another scale. More khan twenty years of research in machine vision has yielded ~0 compelling evidence that even problems once considered to be comparatively simple, such as edge detection, are beyond the capabilikies of known signal processing techniques as measured against performance o the human visual system. In fact, the human visual system is often capable of perceiving and distinguishing meaningful relationships among abstract, non-visual .
(- ~L%~ 53 f ..
,..
variables in graphs, scattergrams, histograms, perspec-tive surface plots, radar returns and like non-visual event description and to organize data in terms of distinguished points, including peaks, steps and inflections, with far greater accuracy than any known signal processing technique.
The scale management probl'em has been addressed extensively in computer vision as for example the works of Rosenfeld, A. and Thurston, M. "Edge and Curve Detec-tion for Visual Scene Analysis," IEEE Transactions on Computers, Vol. C-20, pp.562-56g (May 1971), Marr,-D. and Poggio, T. "A Computational Theory of Human Stereo Vision,i' Proc. R. Soc. Lond., B. 204 (1979) pp. 301-328, and Marr, D. and Hildreth, E.C. "Theory of Edge Detection,"
M.I.T. Artificial Intelligence Memo Number 518, Cambridge, Massachusetts (April 1979). ~owever, the scale manage-ment problem has not been solved satisfactorily. For example, in Marr's works, it is suggeated that multiple descriptions at a fixed series of mask sizes be used for observing events, the mask sizes corresponding to different scales and scale ranges. Marr was unable to integrate these multiple descxiptions effectively. His choice of mask sizes or scale was largely motivated by neural physiological considerations.
SUI~RY OF THE INVENTION
According to the invention, a signàl repre-senting events of interest indicative of phenomena is characterized con-tinuously at all observable scales in a scale-space. Discrete points at which new qualitative features appear in a scale-space image are identified by analysis of directional derivatives of the scale-space image in the direction of lines of constan-t scale. A
discrete tree structure is constructed with reference to the nodes of occurrence of such events in the scale-space planes of the high order derivatives in (~ ~2~
i order to identify ~eatures in successively finer detail without interference from noise. The tree structure may be viewed as a discrete description of a signal at all scales simultaneously or as defining an ordered 1 5 family of descriptions at different scales. Each node may be assigned a value corxesponding to the persistence of events with respect to scale. The value may be employed to identify for filtering purposes a feature of interest. The discrete character of the description permits filtering or categorization of events wherein perceptual scale dependence is an independent variable.
This invention is based on the discovery that there exists a possible correlation between significance of an event and the "persistence or stability" of a lS criterion related to a directional derivative of the ~ scale-space ima~e with respect to variation in scale of i observ2tion. By stability, it is meant the extent of ¦ persistence of an undistinguished interval of data with respect to variation in scale. Stability is not to be confused with the concept of mathematical stability indicative of a converging series.
This discovery suggests that events o significance in differing scales can be observed and identified simultaneously by selection of an appropriate scale of observation for a selected signal in-terval.
The so-called "stability" criterion may be used as a guide to the selection of the scale of observation of particular events in a signal. The invention can be used to develop a discrete approximation in order to identify events from raw data observed in the physical world. Immediate applications include detection of ~ phenomena where prior classical pattern recognition 1 approaches have not provided adequate means to detect edges, line segments~ corners, textures and other visual features in images, such as machine vision, data logging developed from signal returns representative of C` ~ L53 f ` I
underground physical features, radar, ultrasound or any like phenomena.
According to the invention, no distinction is made between signal and noise in any absolute sense.
Features extending over broad ranges of scale may ~e significant for coarse interpretation while features which extend over very narrow and hi~hly resolved scales may be significant for refined interpretation.
Thus, unlike other filtering methods where the object is noise reduction, the object of the present filtering technique is one of separating scale, neighborhoods or descriptions into distinctly significant scale ranges.
According to a preferred embodimen-t of the invention, the plimitive operation providing local parameterization is the convolution of the properties with the normalized gaussian function wherein the standard deviation of this function is the independent variable representing the scale par~meter. The gaussian function convolution provides the advantages of cenier weighting, which tends to preserve local properties. It is also useful in visual edge detection techniques. Other scale parameterization techniques are contemplated should they ultimately prove more efficient or effective, depending upon application.
According to a further aspect of the invention, signals representing phenomena to be analyzed are measured while substantially continuously varying scale range thereby to generate a signal with an extra dimen-sion representing the signal at all scales. In the specific embodiment, measurement is made while continu-ously varying the standard deviation of a gaussian con-volution mask. At the coarser scales, general features are iden-tifiable which can be localized at the finer scales. Moreover, since new qualitative features appear at discrete points in the image of a scale-space plane or its derivatives, ~he continuous surface may be collapsed into a discrete tree structure. The discrete ~2~ 53 tree structure is well adpated to ordered analysis. Thus is a tree structure so generated which is capable of being interpre-ted as a description of a signal viewed simultaneously at all scales or as defining an ordered family of descriptions at different scales.
The technique according to the invention provides a powerful tool for signal analysis since descrete approximation can be developed through this technique without any prior knowledge of the signal or any threshold criteria about raw data.
The invention has immediate application to well logging wherein underground features are reconstructed from raw data and to machine interpretation of co~.plex visual images.
Thus, in accordance with one broad aspect of the invention, there is provided a method for classifying phenomena expressed as detectable signals comprising steps of: observing as a signal of at least a first dimension phenomena within a range of observation resolutîon; and characterized by mapping said signal into scale-space, said phenomena appearing as com-ponents of said signal; classifying said components as events occurring within a range in said scale-space; and tracking said events with respect to said scale-space to a minimun scale value in said scale-space to locate positions of said e~ents with respect to said first dimension.
In accordance with another broad aspect of the inven-tion there is provided an apparatus for filtering a signalrepresented in the form of unscaled data characterized by:
means receptive of said data for adjusting scale o-f observation of said data; means coupled to said scale adjusting means for recognizing occurrence events within said data at a minimum resolution within scale space; means coupled to said events recognizing means for localizing occurence of said events with respect to an independent variable of said signal;
means coupled to said localizing means for classifying sign-ificance of said events; and means coupled to said classifying means for reconstructing a representation of said signal with respect to said independent variable at a plurality of scales based on significance classificati.on of recoynized events.
In accordance with another broad aspect of the invention there is provided a method for classifying phenomena expressed as detectable signals characterized by the steps of:
observing occurrences of events of phenomena within a range of observation resolution; establishing indicia representative of selected ones of said occurrences as a function of scale of said events; and registering said indicia at the point of occurence at a maximum of observation resolution.
In accordance with another broad aspect of the invention there is provided an apparatus for classifying phenomena expressed as detectable signals, said apparatus characterized by: means for observing as a signal of at least a first dimension phenomena within a range of observation res-olution; means for coupled to said observing means for mapping said signal into scale-space, said phenomena appearing as components of said signal; means coupled to said mapping means for classifying said components as events occurring within a range in said scale-space; and means for tracking said events with respect to said scale-space to a minimum scal.e value in said scale-space to locate positions of said events with respect to at least said first dimension.
The invention will be better understood by reference to the following deta:iled description taken in connection with !~ .
`^ ` J.210~53 -7b-~he accompanying drawings.
BRIEF DESCRIPTION OF THE DRAWINGS
Figure 1 is a block diagram of a generalized filtering system according to the invention.
Figure 2 is a waveform diagram of a one dimensional raw data signal with respect to an independent variable such as time whereupon is superimposed a prescaled detected data signal.
Figure 3 is a perspective view of a portion of a convolution surface in scale-space.
Figure 4 is a contour pl~ot in scale-space of the convolution surface of Figure 3.
Figure 5 is a contour plot of the second derivative of the surface of Figure 3 illustrating contours of the loca-tions of the points of inflection in scale-space.
Figure 6 is a scale-space tracking diagram or inter-val tree of the analyzed data of Figure 5 wherein undistinguish-ed intervals with respect to scale nodes are identified as nodes.
s,~
c ~2~ 53 Figure 7 is a tesselated data diagram of the trac~ed data of Figure 6.
Figure 8 is a waveform diagr~n of the raw data of Figure 1 with superimposed thereon a signal approximation based on the maximurn stability criteria according to the invention.
Figure 9 is a family of contours obtained from successively finer gaussian convolutions of a raw data signal.
Figure 10 is a plot of a family of signals with respect to a raw data signal which has been reconstructed based on a stationary variance approxima-tion.
DESCRIPTION OF SPECIFIC EMBODIMENTS
- Figure 1 depicts a generalized filtering system 10 operative according to the invention. As used herein, the term filter refers in general to any mechanism for categorizing data and extracting informa-tion of interest. For simplicity, the system is explained with reference to a one dimensional signal, for example, a signal varying in amplitude only with respect to time. It should be understood, however, that the invention applies to multiple dimensional informaklon as, for example, raw data which can be descrlbed in terms of multiple dependent variables which vary with respect to multiple independent variables, including time, space, amplitude, phase, fre~uency, etc.
As usecl herein signal lines and corresponding signals are assigned the same numerals, and analogous elements are assigned the same numeral throughout.
The filteriny system 10 is responsive to raw data on a raw data signal line 12 to yield processed data on a signal output signal line 22. The processed data may be a reconstructed approximation of the raw data in which information of interest is highlighted.
. __ _ _ . _ _ .___ _ _ ._ _~ . _ _ , . _ . . _ ~.__ .. _ . ,_ ,_ . _ __ . _, _ _ .A .. _ .~_ _ _.. .. .... . .. ..
5~ ( The raw data signal line 12 is coupled to a raw data detector 24. The output of the raw data detector on a signal line 14 herein designated observed data signal line 14 is coupled to a scale-space generator Z6. The scale of observation of the raw data detector 24 is controllable through a scale space parameter control line 27. The output of the scale-space generator 26, which is a signal representative of the observed data projected through scale-space, is provided on a scale-space signal line i6 to a scale-space analyzer 28~ The output of the scale-space analy~er, which may be a derivative of the scale-space signal, such as the second derivative whereby contours of inflation are recorded as zeroes, is provided as analyzed data on analyzed data line 18 to a data tracker 30. The data tracker 30 generates a tracked signal as hereinafter explained on tracked signal line 19 which is coupled to an interval tree constructor 32.
The interval tree constructor 32 generates an interval tree signal, as hereinafter explained, on interval tree signal line 20 which is coupled to a data reconstructor 34. The output of the data reconstructor is the desired processed data on system output signal line 22 and is a selected signal approximating the raw data having been processed by the filtering system 10 according to the invention. The function of a preferred construction of each of these elements will be apparent from the description hereinafter.
By way of further background, perception involves the manipulation of primitive data into meaningful relationships among variables imbedded in the primitive data. According to the present invention, it is an object to organize data in terms of distin-guished points, such as peaks, steps and inflections in a signal and the intervals between such points. It has been found that a description in terms of such local features provides a us~ful alternative to unstructured numerical representation of data and provides appropriate primitives for matching and for more elaborate morpholog-ical descriptions and interpre.ation of data. Further, it has been found that there is significance in both large scale features and small scale or fine-structure features in the same signal interval depending upon the scale of observation. Moreover, the transition from scale to scale is not a continuum, but involves discrete steps among a few scales apparently inherent to the events associated therewith. In the present invention, the understanding of this phenomenon permits the elimination of ambiguity which would otherwise be introduced by limiting the scale range within data as perceived or by arbitrarily establishing a uniform scale.
It is, therefore, a primary objective of the present invention to provide means operative on primitive or raw data to automatically generate an orderly discrete family of meaningful descriptions capturing as nearly as possible discrete events in data as would be perceived and identified by a highly intelligent observer, such as an animal or a human being.
In order for ~aw data to be perceived meaning-fully, it is necessary that the data be detected at some meaningful scale. Since, in the present invention, there i.s no prior knowledge of what is a meaningful scale, the first element of the invention is an adjust-able scale raw data detector 24. .Referring to Figure
S CALE - S PACE F I LTER I NG
ACKGROUND OF THE INVENTION
Field O' Invention This invention relates ~o artificial intelli-gence and to signal processing. In particular, the invention relates to the classification or filtering of inforrnation derived from primitive perception for use in developing a description of observed events i~ the 1~ absence of prior knowledge about the significance of events.
Traditionally, phenomena such as physical - c,bjects, terrain, and the like are considered to have u~ique, 'true" lengths, areas, volumes, tangents, and other metric properties. This "true" value is ihe one o~tained in the limit, as some factor of scale of measurement goes to zero. For example, arc length is defined as the lim-t o~ a polygonal approximation to a cur~e, as the length OI each side of the polygon goes to zero.
The ~raditional notion o~ length doesn't hold for a broad clas~ of irregular 2hysicaL structures anZ
~rocesses. The problem is tllat when measuring the lergtn of, e.g., a coastline by polygonal appro~imation, the measured length increases without apparent ~ound as the length of the poiygonal side decreases (intllitively, because the smaller "yardstick" conforms to ever smaller sub-ba~s and sub-peninsulas.~ Since the limiting value is undefined (or at least uninteresting), there can be no meaninyful evaluation of metric proper~
ties without reference to the scale at which the properties a-e measured (i.e., the length of the "yardstick," or so~e oiher parameter of scale.) ~4 ~2~
It has therefore been assumed that the choice of scale and range of scale is fundamentally a tradeoff between resolution and noise. A small scale observa-tion can resolve closely spaced events but is sensitive to noise. A large scale observa-tion blurs events but is less sensitive to noise. The scale threshold, which :
is often predetermined based on empirical factors for which there is no analytical model, requently deter-mines how events are defined. The inherent ambiguity in scale parameterization is often a serious problem, because there is usually no simple principled basis for setting the scale parameter. Thus, if a ~uery is directed to how long, how wide, or how steeply inclined is an irregular physical structure or the function of a physical measurement process, there is no single "correct" answer, since the answer changes, some-times sisnificantly, as -the scale of measurement is changed.
The ambiguity of scale cannot in general be eliminated by any simple process; rather the problem as herein formulated is to reduce the ambiguity as much as possible and to present the alternatives in an organized, discrete, symbolic fashion.
One technigue for circumventing the problem without really solving it has been simply to select a scale of observation based on a guess at a model for the relationship between the data collected and the conclusions desired. There nevertheless remains a need for a technique for analyzing raw continuous da-ta to develop a discrete syrnbolic description which can be manipulated according to known analytical signal models.
It is to be under~tood that measurement of any sort requires at least some primitive scale parame-teriæa-tion in order to maintain a measurement of properties in a range of scale. The size of any neighborhood determines the scale of description. For example, a linear scale may be able to usefully present events for display with resp~ct to an independent variable between zero and ten with a resolution of one part in one hundred. This represents a uniformly weighted scale range of three orders of magnitude.
Various property measurement techniques may be applied within any scale range as, for example, local fitting of analytical functions to data, spatial averaging and the like. Each property measurement techni~ue has in common the continuous parameterization by scale of some local measure on the signal.
As used herein a scale-dependent measure on a signal might be defined as a function of the signal, a function of a location on the signal, and a function of a parameter of scale, such that values in the signal contribute to the function~s output to a diminishing degree as they become more distant from the original location; and such that the rate at which the values' cont-ibutions diminish with distance decreases as the value of the parameter of scale increases A scale~
dependent measure is for example the result of convolving the signal with a gaussian, using the gaussian's standard deviation ("width"~ as the parameter o~ scale~
Description Of The Prior Art There are no known analytical signal models capable of describing different physical events which occur a-t different physical scales, namely, events which appear like noise ~lithin one scale but correspond to interesting events within another scale. More khan twenty years of research in machine vision has yielded ~0 compelling evidence that even problems once considered to be comparatively simple, such as edge detection, are beyond the capabilikies of known signal processing techniques as measured against performance o the human visual system. In fact, the human visual system is often capable of perceiving and distinguishing meaningful relationships among abstract, non-visual .
(- ~L%~ 53 f ..
,..
variables in graphs, scattergrams, histograms, perspec-tive surface plots, radar returns and like non-visual event description and to organize data in terms of distinguished points, including peaks, steps and inflections, with far greater accuracy than any known signal processing technique.
The scale management probl'em has been addressed extensively in computer vision as for example the works of Rosenfeld, A. and Thurston, M. "Edge and Curve Detec-tion for Visual Scene Analysis," IEEE Transactions on Computers, Vol. C-20, pp.562-56g (May 1971), Marr,-D. and Poggio, T. "A Computational Theory of Human Stereo Vision,i' Proc. R. Soc. Lond., B. 204 (1979) pp. 301-328, and Marr, D. and Hildreth, E.C. "Theory of Edge Detection,"
M.I.T. Artificial Intelligence Memo Number 518, Cambridge, Massachusetts (April 1979). ~owever, the scale manage-ment problem has not been solved satisfactorily. For example, in Marr's works, it is suggeated that multiple descriptions at a fixed series of mask sizes be used for observing events, the mask sizes corresponding to different scales and scale ranges. Marr was unable to integrate these multiple descxiptions effectively. His choice of mask sizes or scale was largely motivated by neural physiological considerations.
SUI~RY OF THE INVENTION
According to the invention, a signàl repre-senting events of interest indicative of phenomena is characterized con-tinuously at all observable scales in a scale-space. Discrete points at which new qualitative features appear in a scale-space image are identified by analysis of directional derivatives of the scale-space image in the direction of lines of constan-t scale. A
discrete tree structure is constructed with reference to the nodes of occurrence of such events in the scale-space planes of the high order derivatives in (~ ~2~
i order to identify ~eatures in successively finer detail without interference from noise. The tree structure may be viewed as a discrete description of a signal at all scales simultaneously or as defining an ordered 1 5 family of descriptions at different scales. Each node may be assigned a value corxesponding to the persistence of events with respect to scale. The value may be employed to identify for filtering purposes a feature of interest. The discrete character of the description permits filtering or categorization of events wherein perceptual scale dependence is an independent variable.
This invention is based on the discovery that there exists a possible correlation between significance of an event and the "persistence or stability" of a lS criterion related to a directional derivative of the ~ scale-space ima~e with respect to variation in scale of i observ2tion. By stability, it is meant the extent of ¦ persistence of an undistinguished interval of data with respect to variation in scale. Stability is not to be confused with the concept of mathematical stability indicative of a converging series.
This discovery suggests that events o significance in differing scales can be observed and identified simultaneously by selection of an appropriate scale of observation for a selected signal in-terval.
The so-called "stability" criterion may be used as a guide to the selection of the scale of observation of particular events in a signal. The invention can be used to develop a discrete approximation in order to identify events from raw data observed in the physical world. Immediate applications include detection of ~ phenomena where prior classical pattern recognition 1 approaches have not provided adequate means to detect edges, line segments~ corners, textures and other visual features in images, such as machine vision, data logging developed from signal returns representative of C` ~ L53 f ` I
underground physical features, radar, ultrasound or any like phenomena.
According to the invention, no distinction is made between signal and noise in any absolute sense.
Features extending over broad ranges of scale may ~e significant for coarse interpretation while features which extend over very narrow and hi~hly resolved scales may be significant for refined interpretation.
Thus, unlike other filtering methods where the object is noise reduction, the object of the present filtering technique is one of separating scale, neighborhoods or descriptions into distinctly significant scale ranges.
According to a preferred embodimen-t of the invention, the plimitive operation providing local parameterization is the convolution of the properties with the normalized gaussian function wherein the standard deviation of this function is the independent variable representing the scale par~meter. The gaussian function convolution provides the advantages of cenier weighting, which tends to preserve local properties. It is also useful in visual edge detection techniques. Other scale parameterization techniques are contemplated should they ultimately prove more efficient or effective, depending upon application.
According to a further aspect of the invention, signals representing phenomena to be analyzed are measured while substantially continuously varying scale range thereby to generate a signal with an extra dimen-sion representing the signal at all scales. In the specific embodiment, measurement is made while continu-ously varying the standard deviation of a gaussian con-volution mask. At the coarser scales, general features are iden-tifiable which can be localized at the finer scales. Moreover, since new qualitative features appear at discrete points in the image of a scale-space plane or its derivatives, ~he continuous surface may be collapsed into a discrete tree structure. The discrete ~2~ 53 tree structure is well adpated to ordered analysis. Thus is a tree structure so generated which is capable of being interpre-ted as a description of a signal viewed simultaneously at all scales or as defining an ordered family of descriptions at different scales.
The technique according to the invention provides a powerful tool for signal analysis since descrete approximation can be developed through this technique without any prior knowledge of the signal or any threshold criteria about raw data.
The invention has immediate application to well logging wherein underground features are reconstructed from raw data and to machine interpretation of co~.plex visual images.
Thus, in accordance with one broad aspect of the invention, there is provided a method for classifying phenomena expressed as detectable signals comprising steps of: observing as a signal of at least a first dimension phenomena within a range of observation resolutîon; and characterized by mapping said signal into scale-space, said phenomena appearing as com-ponents of said signal; classifying said components as events occurring within a range in said scale-space; and tracking said events with respect to said scale-space to a minimun scale value in said scale-space to locate positions of said e~ents with respect to said first dimension.
In accordance with another broad aspect of the inven-tion there is provided an apparatus for filtering a signalrepresented in the form of unscaled data characterized by:
means receptive of said data for adjusting scale o-f observation of said data; means coupled to said scale adjusting means for recognizing occurrence events within said data at a minimum resolution within scale space; means coupled to said events recognizing means for localizing occurence of said events with respect to an independent variable of said signal;
means coupled to said localizing means for classifying sign-ificance of said events; and means coupled to said classifying means for reconstructing a representation of said signal with respect to said independent variable at a plurality of scales based on significance classificati.on of recoynized events.
In accordance with another broad aspect of the invention there is provided a method for classifying phenomena expressed as detectable signals characterized by the steps of:
observing occurrences of events of phenomena within a range of observation resolution; establishing indicia representative of selected ones of said occurrences as a function of scale of said events; and registering said indicia at the point of occurence at a maximum of observation resolution.
In accordance with another broad aspect of the invention there is provided an apparatus for classifying phenomena expressed as detectable signals, said apparatus characterized by: means for observing as a signal of at least a first dimension phenomena within a range of observation res-olution; means for coupled to said observing means for mapping said signal into scale-space, said phenomena appearing as components of said signal; means coupled to said mapping means for classifying said components as events occurring within a range in said scale-space; and means for tracking said events with respect to said scale-space to a minimum scal.e value in said scale-space to locate positions of said events with respect to at least said first dimension.
The invention will be better understood by reference to the following deta:iled description taken in connection with !~ .
`^ ` J.210~53 -7b-~he accompanying drawings.
BRIEF DESCRIPTION OF THE DRAWINGS
Figure 1 is a block diagram of a generalized filtering system according to the invention.
Figure 2 is a waveform diagram of a one dimensional raw data signal with respect to an independent variable such as time whereupon is superimposed a prescaled detected data signal.
Figure 3 is a perspective view of a portion of a convolution surface in scale-space.
Figure 4 is a contour pl~ot in scale-space of the convolution surface of Figure 3.
Figure 5 is a contour plot of the second derivative of the surface of Figure 3 illustrating contours of the loca-tions of the points of inflection in scale-space.
Figure 6 is a scale-space tracking diagram or inter-val tree of the analyzed data of Figure 5 wherein undistinguish-ed intervals with respect to scale nodes are identified as nodes.
s,~
c ~2~ 53 Figure 7 is a tesselated data diagram of the trac~ed data of Figure 6.
Figure 8 is a waveform diagr~n of the raw data of Figure 1 with superimposed thereon a signal approximation based on the maximurn stability criteria according to the invention.
Figure 9 is a family of contours obtained from successively finer gaussian convolutions of a raw data signal.
Figure 10 is a plot of a family of signals with respect to a raw data signal which has been reconstructed based on a stationary variance approxima-tion.
DESCRIPTION OF SPECIFIC EMBODIMENTS
- Figure 1 depicts a generalized filtering system 10 operative according to the invention. As used herein, the term filter refers in general to any mechanism for categorizing data and extracting informa-tion of interest. For simplicity, the system is explained with reference to a one dimensional signal, for example, a signal varying in amplitude only with respect to time. It should be understood, however, that the invention applies to multiple dimensional informaklon as, for example, raw data which can be descrlbed in terms of multiple dependent variables which vary with respect to multiple independent variables, including time, space, amplitude, phase, fre~uency, etc.
As usecl herein signal lines and corresponding signals are assigned the same numerals, and analogous elements are assigned the same numeral throughout.
The filteriny system 10 is responsive to raw data on a raw data signal line 12 to yield processed data on a signal output signal line 22. The processed data may be a reconstructed approximation of the raw data in which information of interest is highlighted.
. __ _ _ . _ _ .___ _ _ ._ _~ . _ _ , . _ . . _ ~.__ .. _ . ,_ ,_ . _ __ . _, _ _ .A .. _ .~_ _ _.. .. .... . .. ..
5~ ( The raw data signal line 12 is coupled to a raw data detector 24. The output of the raw data detector on a signal line 14 herein designated observed data signal line 14 is coupled to a scale-space generator Z6. The scale of observation of the raw data detector 24 is controllable through a scale space parameter control line 27. The output of the scale-space generator 26, which is a signal representative of the observed data projected through scale-space, is provided on a scale-space signal line i6 to a scale-space analyzer 28~ The output of the scale-space analy~er, which may be a derivative of the scale-space signal, such as the second derivative whereby contours of inflation are recorded as zeroes, is provided as analyzed data on analyzed data line 18 to a data tracker 30. The data tracker 30 generates a tracked signal as hereinafter explained on tracked signal line 19 which is coupled to an interval tree constructor 32.
The interval tree constructor 32 generates an interval tree signal, as hereinafter explained, on interval tree signal line 20 which is coupled to a data reconstructor 34. The output of the data reconstructor is the desired processed data on system output signal line 22 and is a selected signal approximating the raw data having been processed by the filtering system 10 according to the invention. The function of a preferred construction of each of these elements will be apparent from the description hereinafter.
By way of further background, perception involves the manipulation of primitive data into meaningful relationships among variables imbedded in the primitive data. According to the present invention, it is an object to organize data in terms of distin-guished points, such as peaks, steps and inflections in a signal and the intervals between such points. It has been found that a description in terms of such local features provides a us~ful alternative to unstructured numerical representation of data and provides appropriate primitives for matching and for more elaborate morpholog-ical descriptions and interpre.ation of data. Further, it has been found that there is significance in both large scale features and small scale or fine-structure features in the same signal interval depending upon the scale of observation. Moreover, the transition from scale to scale is not a continuum, but involves discrete steps among a few scales apparently inherent to the events associated therewith. In the present invention, the understanding of this phenomenon permits the elimination of ambiguity which would otherwise be introduced by limiting the scale range within data as perceived or by arbitrarily establishing a uniform scale.
It is, therefore, a primary objective of the present invention to provide means operative on primitive or raw data to automatically generate an orderly discrete family of meaningful descriptions capturing as nearly as possible discrete events in data as would be perceived and identified by a highly intelligent observer, such as an animal or a human being.
In order for ~aw data to be perceived meaning-fully, it is necessary that the data be detected at some meaningful scale. Since, in the present invention, there i.s no prior knowledge of what is a meaningful scale, the first element of the invention is an adjust-able scale raw data detector 24. .Referring to Figure
2, there is shown a representation of raw data 12 in the form of a continuum in variations in amplitude as a function of an independent variable x. The variable may represent time, distance or the like. The amplitude range shown herein involves a scale, a neighborhood, or a resolution of about one and one-half (l-l/2) orders of magnitude.
Dete~tion can be described as the measurement of neighborhood properties wherein the size of the .
f ~ 53 . .
11 .-.
neighborhood properties determines the scale. The detection scheme, according to a preferred embodiment of the invention, involves the primitive operation of the convolution of the raw data with a function which is dually dependent on the independent variable or variables in the data and on a scale parameter.
Specifically, the normalized center-weighted gaussian function, which is given by the expression: ~
x2 .
g(x ~) = 1 e 2 ~l 2n where x is the independent variable of the data and a, the standard deviation of the gaussian function, is the scale parameter. The convolution of the gaussian function with the data provides several advantages.
First, since the operator is center-weighted, it tends to preserve local properties. At minimum standard deviation, maximum local properti~s are preserved.
Second, the operator is a continuous differentiable function of x and ~. Therefore, scale-par2meterized derivatives can be obtained at each point as a function of a continuous scale parameter. This characteristic is important to the analysis of the characteristics of the raw data.
Referring again to Figure 2, there is shown a waveform 14 corresponding to the siynal on signal line 14 which represents data detected with an arbitrary parameter of scale. ~he signal is a gaussian smoothed representation of the raw data signal 12. It will be immediately apparent that the observed data signal 14 does not fully characterize the raw data signal 12 although it does characterixe more or less significant features of the raw data signal. According to the invention, means are provided for characterizing the raw data signal 12 at all scales in terms of a surface swept out by continuously varying the standard deviation :~Z~L0153 of a gaussian convolution mask. Referring to Figure 1, the scale-space parameter control line 27 may have the function of conveying the signal for varying the mask width, and specifically the standard deviation par~meter of the gaussian convolution mask, in order to generate a family of descriptions of the raw data 12 in scale- :
space.
Figure 3 depicts a three dimensional scale-space signal 16 as would be generated on scale-space signal line 16 (Fig. 1). The observed data signal 14 of Figure 2 is shown as a contour line paralleling the x axis of the scale-space surface 16 of Figure 3. The scàle-space surface 16 has been txuncated at this boundary merely for the ease of illustration. The scale-space 16 may he extended in either direction along the scale parameter axis.
The surface 16 of Figure 3 is defined by the function: _ (x--u)2 ~ F(x,~) = f(x) * g(x,a) = ~ ~ f(u) 1 e du ~2~
where "~" denotes convolution with respect to the independent variable x and u is a dummy variable of integration. The convolution function F (x,a) of Eguation 2 is herein denoted the scale-space representa-tion of the signal f(x) which is the raw data signal 12.
Thus, F(x,a~ is the scale space signal 16. Figure 4 illustrates the projection of the scale-space signal 16 in a contou~ plot in the plane formed by the orthogonal variables x and F. Each of the contours indicates iso-planar magnitudes of the function F(x,a).
An important feature of the invention is that the convolution function F(x,a) is in general continuous and differentiable even when the input signal f(x) is not.
The deriv~tives of the convolu~ion function with respect to the independent variable x are given by the expression:
.53 an an axn F(x u) = f(x) * -g(x,a) where the derivatives of the gaussian function are readily obtained.
To develop a scale-dependent description of a subject function f(x), the extrema in the x direction of the F(x,a~ function and its derivatives with respect to x provide useful benchmarks. The extrema in the raw data signal 12 (Fig. 2) are points in the observed data signal 14, such as first extremum 38 or second extremum 40. Fi~ure 5 illustrates the extrema of Fig. 2 derived from plotting the zero contours in the second derivative, with respect to ~, of the convolution surface illustrated in Fig. 3. The representation of 15 F}g. 5 is the analyzed data 18 and a first ccntour 138 corresponds to the first extremum 38 while a second contour 140 corresponds to the second extremum 40. ~n . accordance. wi.th well understood mathematical rules, zeroes in the second derivative correspond to inflec-tions in the function under analysis. The zero contours, I as shown in Fig. 5 in the plane of analyzed data 18 thus reveal aualitative articulation of the signal ¦ under consideration wherein new features appear and I existing ones sharpen or are localized with increasing ! 25 resolution. More particularly, with each inflection ox ! with each interval between inflections is associated a ~alue of ~ at which a feature iIl the signal un~er con-j sideration first appears. In the particular data under examination, primary interest is in step transitions.
Therefore, attention may be limited for simplicity to extrema in slope.
The analyzed data 18 of Fig. 5, representing the zero contours on the differentiated convolution surface is processed by the data tracker 30 (Fig. 1) tG
Dete~tion can be described as the measurement of neighborhood properties wherein the size of the .
f ~ 53 . .
11 .-.
neighborhood properties determines the scale. The detection scheme, according to a preferred embodiment of the invention, involves the primitive operation of the convolution of the raw data with a function which is dually dependent on the independent variable or variables in the data and on a scale parameter.
Specifically, the normalized center-weighted gaussian function, which is given by the expression: ~
x2 .
g(x ~) = 1 e 2 ~l 2n where x is the independent variable of the data and a, the standard deviation of the gaussian function, is the scale parameter. The convolution of the gaussian function with the data provides several advantages.
First, since the operator is center-weighted, it tends to preserve local properties. At minimum standard deviation, maximum local properti~s are preserved.
Second, the operator is a continuous differentiable function of x and ~. Therefore, scale-par2meterized derivatives can be obtained at each point as a function of a continuous scale parameter. This characteristic is important to the analysis of the characteristics of the raw data.
Referring again to Figure 2, there is shown a waveform 14 corresponding to the siynal on signal line 14 which represents data detected with an arbitrary parameter of scale. ~he signal is a gaussian smoothed representation of the raw data signal 12. It will be immediately apparent that the observed data signal 14 does not fully characterize the raw data signal 12 although it does characterixe more or less significant features of the raw data signal. According to the invention, means are provided for characterizing the raw data signal 12 at all scales in terms of a surface swept out by continuously varying the standard deviation :~Z~L0153 of a gaussian convolution mask. Referring to Figure 1, the scale-space parameter control line 27 may have the function of conveying the signal for varying the mask width, and specifically the standard deviation par~meter of the gaussian convolution mask, in order to generate a family of descriptions of the raw data 12 in scale- :
space.
Figure 3 depicts a three dimensional scale-space signal 16 as would be generated on scale-space signal line 16 (Fig. 1). The observed data signal 14 of Figure 2 is shown as a contour line paralleling the x axis of the scale-space surface 16 of Figure 3. The scàle-space surface 16 has been txuncated at this boundary merely for the ease of illustration. The scale-space 16 may he extended in either direction along the scale parameter axis.
The surface 16 of Figure 3 is defined by the function: _ (x--u)2 ~ F(x,~) = f(x) * g(x,a) = ~ ~ f(u) 1 e du ~2~
where "~" denotes convolution with respect to the independent variable x and u is a dummy variable of integration. The convolution function F (x,a) of Eguation 2 is herein denoted the scale-space representa-tion of the signal f(x) which is the raw data signal 12.
Thus, F(x,a~ is the scale space signal 16. Figure 4 illustrates the projection of the scale-space signal 16 in a contou~ plot in the plane formed by the orthogonal variables x and F. Each of the contours indicates iso-planar magnitudes of the function F(x,a).
An important feature of the invention is that the convolution function F(x,a) is in general continuous and differentiable even when the input signal f(x) is not.
The deriv~tives of the convolu~ion function with respect to the independent variable x are given by the expression:
.53 an an axn F(x u) = f(x) * -g(x,a) where the derivatives of the gaussian function are readily obtained.
To develop a scale-dependent description of a subject function f(x), the extrema in the x direction of the F(x,a~ function and its derivatives with respect to x provide useful benchmarks. The extrema in the raw data signal 12 (Fig. 2) are points in the observed data signal 14, such as first extremum 38 or second extremum 40. Fi~ure 5 illustrates the extrema of Fig. 2 derived from plotting the zero contours in the second derivative, with respect to ~, of the convolution surface illustrated in Fig. 3. The representation of 15 F}g. 5 is the analyzed data 18 and a first ccntour 138 corresponds to the first extremum 38 while a second contour 140 corresponds to the second extremum 40. ~n . accordance. wi.th well understood mathematical rules, zeroes in the second derivative correspond to inflec-tions in the function under analysis. The zero contours, I as shown in Fig. 5 in the plane of analyzed data 18 thus reveal aualitative articulation of the signal ¦ under consideration wherein new features appear and I existing ones sharpen or are localized with increasing ! 25 resolution. More particularly, with each inflection ox ! with each interval between inflections is associated a ~alue of ~ at which a feature iIl the signal un~er con-j sideration first appears. In the particular data under examination, primary interest is in step transitions.
Therefore, attention may be limited for simplicity to extrema in slope.
The analyzed data 18 of Fig. 5, representing the zero contours on the differentiated convolution surface is processed by the data tracker 30 (Fig. 1) tG
3~ extract simple qualit~tive features without undesirable . spatial distortion of the location of those features w1th respect to th~ function x. The procedure, hereinafter called localization by coarse to-fine tracking, comprises the following steps:
After computation of the second-derivative of the scale-space image with respect io x to obtain the analyzed data signal 18, all of the second derivative zeroes passing through a given value of x are found and then each such zero is tracked throu~h successively smaller values of a to the finest scale. ~he resulting sequence of points gives the localized extrema-slope points on the signal for the selected maximum value of a. (In other words, the scale-space analyzer simply extracts the second derivative, and the da-ta tracker maps the values for the second derivative across the differentiated convolution surface.) ~eferring to Figures 9 and 10 together, there is shown an example of coarse-to-fine tracking according to the invention (Fig. 10) wherein data is reconstructed by fitting-least squares-type parabolas between inflec-tion points and standard gaussian smoothing ~Fig. ~) is employed to filter the raw data signal. In Figure 9, gaussian smoothing guickly dampens all high fre~uency components as the value of a is increased. This also is seen in Fig. 3, which is a three dimensional repre-sentation of the data shown in Figure 9. By contrast, however, the filtering procedure illustrated in Figure 10 shows the retention of high frequency components at full amplitude accurately localized with respec-t to x with increasing values of a. Significant features disappear only when an inflection associated with a feature also disappears in a second derivative scale~
space image. One prominent high frequency feature has been labeled extremum 142. Again, significantly, prominent features maintain full amplitude as long as a feature retains its significance in scale-space. This characteristic allows for the assignment of a logical or binary value t:o a feature as a function of x and a, which is extremely useful in decision making applica-53 r tions employing the raw data, such as edge detectionand the like.
The coarse-to-fine tracking procedure is a non-linear filtering procedure which can be refined to further identify and categorize features of prominence within a particular scale range. ~hereas exclusion or inclusion of features in the description of a signal is determined somewhat haphazardly by the values selected for o in the scale-space at which tracking begins, according to a further feature of the present invention, a process is provided for capturing discretely the introduction of new features at successively finer scales. This process is implemented by providing an interval tree structure which provides localized dis-crete control whereby the space of potential descrip-tions of a eature is simultaneously far more tightly constrained yet far more flexible. Moreover, the process, according to the invention, provides a tech-nique for a natural hierarchic ordering of the descrip-tions within the space.
Referring to Figures 5 and 5, features of asignal in the scale-space image, in this case inflec-tions, change continuously with a continuously vaxiable value for a with new features appearing as feature extrema 144, 146 in contours, for example, contour 148 and contour 150. For every distinguished point, such as feature extrema 144 and 146, there is a dual, namely an undistinguished interval 152 between the extrema, with a decrease in a only increasing the precision of locality. An example of an undistinguished interval labeled 152 is the space within the boundary of contour 150 and scale value line 149 between extrema 144 and 146. Interval 152 is mapped into a rectangular region in Figure 7 by tesselation in the x,o plane. It is bounded on x by the final value of the zero contours along the x axis of contour 150 and on the other two parallel sides by the extrema 144 and 146 along the o ~l2~
axis. Interval 152 resolves into three intervals, 154, 156 and 158 at the point of extrema 146 of relatively lower extre~um value. With decreasing scale values, the subintexvals resolve into further sets of ~h~e~ ~
subintervals at each feature introduction. 51 Figure 6 is an illustration of a tree 160 -constructed on the contours of Figure 5. The tree is constructed by assembling the intervals such as 152, 154, 156 an 158 into an order based on source (parent) and branch (offspring).
As a part of the filtering process each node of the tree may be assigned a weight in accordance with the extent of the associated interval with respect to the a axis. Nodes with maximum weight may thereby be associated with features of significance. Significance may ~e based on the criterion that persistence in scale within a range of scale correlates with importance of the event associated with the node corresponding to the interval. The intervals with the selected weights are highlighted in Figure 7.
A description of a signal in terms of a tree can be considered a description of the signal at all scales simultaneously or as a description of the signal as a family of single-scale descriptions, only one scale of which is considered relevant at any point along the x axis. The scale of description may be varied locally, that is, within each interval along the x axis, and the scale may be changed in well-defined steps as between scale values at feature introductions.
Changes in scale may be effected by subdividing a node into its offspring, thereby moving towards smaller values of a, or by merging, that is, combining related offspring into a common parent, thereby moving scale toward greater values of a~
Figure 7 illustrates a tesselation o the zero contours of the scale--space plane in the form of a two dimensional interva1 signal 20. The interval C ~ 53 ~, signal 20 is derived directly from the analyzed data sisnal 18 ~Fig. 5). In Figure 7, a corresponding one dimensional raw data signal 12 is shown along the x axis parallel to the x axis of the two-dimensional interval signal 20. A different interval signal 20 would be generated if a different type of analyzed data were employed, as for example a signal based on contours of zeroes in the first derivative or contours of zeroes in the thixd derivative.
In order to present system output data 22 in the form of a minimum dimensional signal, the concept of undistinguished inter,val selection is introduced into scale-space and particularly selection based on stability of undistinguished intervals as a function of scale from Figure 6. More specifically, there exists for all events in the x axis which define offspring at the minimum value of ~, an ancestor in its interval tree which best represents data of interest. The choice of ancestor may be based on maximum persistence over variation in scale (along the a axis), It has been observed that the stability of an undistinguished interval correlates reasonably well with perceptual salience to the intelligent (human) observer. For example, undistinguished interval 152 is the region having greatest stability as compared to undistinguished intervals 154~ 156, 159 and the offspring of region 158 over the same interval of signal along the x axis.
Therefore the boundaries, in x of range 164 of the undistinguished interval 152, have been selected to indicate significant events, such as an edge of a feature. According to the invention, undistinguished interval 152 and the ~ value corresponding to point 144 (Fig. 5) now tesselated to a line 144' (Fig. 7~ becomes the reference for mapping raw data features along the x axis into a system output sigr.al 22. A reconstructed interval 164 of system output data signal 22 (Fig. 7) is extracted frorn the raw data on the basis of reference ¢ 9~2101S3 to undistinguished interval 152. The shape of a reconstructed interval 164 may be obtained by fitting a least-squares parabola between the boundaries of the region 152 along the x axis, having determined and localized the boundaries by trackin~ the zeroes in the second-derivative to the minimum value of a. Every other interval selected for system output signal 2~ may be constructed in a similar manner based on reference to a selected undistinguislled region with respect to ' 10 the scale-space continuum.
Figure 8 illustrates a plot of the waveform of a one dimensional raw data signal 12 superimposed on a plot of the approximation for system output signal 22 wherein maximum stability criterion has been applied.
Eigh frequency structure is preserved in the output signal 22, as for example interval 171, where the high freguency structure is deemed to be of significance according~to the stability criterion.
The invention has b~en applied to the analysis of gamma ray logs of geological formations with signifi-cant success. Segmentation on the basis of maximum stability by this filtering process has been found to correspond with remarkable accuracy to the tops of major sands as verified by trained geologists who have examined the raw data directly. Success with such geologic data suggests that the same techni~les can be applied wi-th reasonable success to other types o perceptual filtering tasks involving primitive shape description and identification.
9 ~2~ 3 The invention can be applied to virtually any ~ype of filtering system, including time-domain filtering, frequency domain filtering and multi-dimensional domain filtering, particularly where there is no prior knowledge about the signal or the scale of description most relevant to the information contained in the signal.
The invention has been explained with reference to specific embodiments. Other embodiments will be apparent to those of ordinary skill in the art upon reference to the description hereinabove. It is, therefore, not intended that this invention be limited, except as indicated by the appended claims.
After computation of the second-derivative of the scale-space image with respect io x to obtain the analyzed data signal 18, all of the second derivative zeroes passing through a given value of x are found and then each such zero is tracked throu~h successively smaller values of a to the finest scale. ~he resulting sequence of points gives the localized extrema-slope points on the signal for the selected maximum value of a. (In other words, the scale-space analyzer simply extracts the second derivative, and the da-ta tracker maps the values for the second derivative across the differentiated convolution surface.) ~eferring to Figures 9 and 10 together, there is shown an example of coarse-to-fine tracking according to the invention (Fig. 10) wherein data is reconstructed by fitting-least squares-type parabolas between inflec-tion points and standard gaussian smoothing ~Fig. ~) is employed to filter the raw data signal. In Figure 9, gaussian smoothing guickly dampens all high fre~uency components as the value of a is increased. This also is seen in Fig. 3, which is a three dimensional repre-sentation of the data shown in Figure 9. By contrast, however, the filtering procedure illustrated in Figure 10 shows the retention of high frequency components at full amplitude accurately localized with respec-t to x with increasing values of a. Significant features disappear only when an inflection associated with a feature also disappears in a second derivative scale~
space image. One prominent high frequency feature has been labeled extremum 142. Again, significantly, prominent features maintain full amplitude as long as a feature retains its significance in scale-space. This characteristic allows for the assignment of a logical or binary value t:o a feature as a function of x and a, which is extremely useful in decision making applica-53 r tions employing the raw data, such as edge detectionand the like.
The coarse-to-fine tracking procedure is a non-linear filtering procedure which can be refined to further identify and categorize features of prominence within a particular scale range. ~hereas exclusion or inclusion of features in the description of a signal is determined somewhat haphazardly by the values selected for o in the scale-space at which tracking begins, according to a further feature of the present invention, a process is provided for capturing discretely the introduction of new features at successively finer scales. This process is implemented by providing an interval tree structure which provides localized dis-crete control whereby the space of potential descrip-tions of a eature is simultaneously far more tightly constrained yet far more flexible. Moreover, the process, according to the invention, provides a tech-nique for a natural hierarchic ordering of the descrip-tions within the space.
Referring to Figures 5 and 5, features of asignal in the scale-space image, in this case inflec-tions, change continuously with a continuously vaxiable value for a with new features appearing as feature extrema 144, 146 in contours, for example, contour 148 and contour 150. For every distinguished point, such as feature extrema 144 and 146, there is a dual, namely an undistinguished interval 152 between the extrema, with a decrease in a only increasing the precision of locality. An example of an undistinguished interval labeled 152 is the space within the boundary of contour 150 and scale value line 149 between extrema 144 and 146. Interval 152 is mapped into a rectangular region in Figure 7 by tesselation in the x,o plane. It is bounded on x by the final value of the zero contours along the x axis of contour 150 and on the other two parallel sides by the extrema 144 and 146 along the o ~l2~
axis. Interval 152 resolves into three intervals, 154, 156 and 158 at the point of extrema 146 of relatively lower extre~um value. With decreasing scale values, the subintexvals resolve into further sets of ~h~e~ ~
subintervals at each feature introduction. 51 Figure 6 is an illustration of a tree 160 -constructed on the contours of Figure 5. The tree is constructed by assembling the intervals such as 152, 154, 156 an 158 into an order based on source (parent) and branch (offspring).
As a part of the filtering process each node of the tree may be assigned a weight in accordance with the extent of the associated interval with respect to the a axis. Nodes with maximum weight may thereby be associated with features of significance. Significance may ~e based on the criterion that persistence in scale within a range of scale correlates with importance of the event associated with the node corresponding to the interval. The intervals with the selected weights are highlighted in Figure 7.
A description of a signal in terms of a tree can be considered a description of the signal at all scales simultaneously or as a description of the signal as a family of single-scale descriptions, only one scale of which is considered relevant at any point along the x axis. The scale of description may be varied locally, that is, within each interval along the x axis, and the scale may be changed in well-defined steps as between scale values at feature introductions.
Changes in scale may be effected by subdividing a node into its offspring, thereby moving towards smaller values of a, or by merging, that is, combining related offspring into a common parent, thereby moving scale toward greater values of a~
Figure 7 illustrates a tesselation o the zero contours of the scale--space plane in the form of a two dimensional interva1 signal 20. The interval C ~ 53 ~, signal 20 is derived directly from the analyzed data sisnal 18 ~Fig. 5). In Figure 7, a corresponding one dimensional raw data signal 12 is shown along the x axis parallel to the x axis of the two-dimensional interval signal 20. A different interval signal 20 would be generated if a different type of analyzed data were employed, as for example a signal based on contours of zeroes in the first derivative or contours of zeroes in the thixd derivative.
In order to present system output data 22 in the form of a minimum dimensional signal, the concept of undistinguished inter,val selection is introduced into scale-space and particularly selection based on stability of undistinguished intervals as a function of scale from Figure 6. More specifically, there exists for all events in the x axis which define offspring at the minimum value of ~, an ancestor in its interval tree which best represents data of interest. The choice of ancestor may be based on maximum persistence over variation in scale (along the a axis), It has been observed that the stability of an undistinguished interval correlates reasonably well with perceptual salience to the intelligent (human) observer. For example, undistinguished interval 152 is the region having greatest stability as compared to undistinguished intervals 154~ 156, 159 and the offspring of region 158 over the same interval of signal along the x axis.
Therefore the boundaries, in x of range 164 of the undistinguished interval 152, have been selected to indicate significant events, such as an edge of a feature. According to the invention, undistinguished interval 152 and the ~ value corresponding to point 144 (Fig. 5) now tesselated to a line 144' (Fig. 7~ becomes the reference for mapping raw data features along the x axis into a system output sigr.al 22. A reconstructed interval 164 of system output data signal 22 (Fig. 7) is extracted frorn the raw data on the basis of reference ¢ 9~2101S3 to undistinguished interval 152. The shape of a reconstructed interval 164 may be obtained by fitting a least-squares parabola between the boundaries of the region 152 along the x axis, having determined and localized the boundaries by trackin~ the zeroes in the second-derivative to the minimum value of a. Every other interval selected for system output signal 2~ may be constructed in a similar manner based on reference to a selected undistinguislled region with respect to ' 10 the scale-space continuum.
Figure 8 illustrates a plot of the waveform of a one dimensional raw data signal 12 superimposed on a plot of the approximation for system output signal 22 wherein maximum stability criterion has been applied.
Eigh frequency structure is preserved in the output signal 22, as for example interval 171, where the high freguency structure is deemed to be of significance according~to the stability criterion.
The invention has b~en applied to the analysis of gamma ray logs of geological formations with signifi-cant success. Segmentation on the basis of maximum stability by this filtering process has been found to correspond with remarkable accuracy to the tops of major sands as verified by trained geologists who have examined the raw data directly. Success with such geologic data suggests that the same techni~les can be applied wi-th reasonable success to other types o perceptual filtering tasks involving primitive shape description and identification.
9 ~2~ 3 The invention can be applied to virtually any ~ype of filtering system, including time-domain filtering, frequency domain filtering and multi-dimensional domain filtering, particularly where there is no prior knowledge about the signal or the scale of description most relevant to the information contained in the signal.
The invention has been explained with reference to specific embodiments. Other embodiments will be apparent to those of ordinary skill in the art upon reference to the description hereinabove. It is, therefore, not intended that this invention be limited, except as indicated by the appended claims.
Claims (9)
1. A method for classifying phenomena expressed as detectable signals comprising steps of:
observing as a signal of at least a first dimension phenomena within a range of observation resolution; and characterized by mapping said signal into scale-space, said phenomena appearing as components of said signal;
classifying said components as events occurring within a range in said scale-space; and tracking said events with respect to said scale-space to a minimum scale value in said scale-space to locate positions of said events with respect to said first dimension.
observing as a signal of at least a first dimension phenomena within a range of observation resolution; and characterized by mapping said signal into scale-space, said phenomena appearing as components of said signal;
classifying said components as events occurring within a range in said scale-space; and tracking said events with respect to said scale-space to a minimum scale value in said scale-space to locate positions of said events with respect to said first dimension.
2. The method according to claim 1 characterized in that said classifying step comprises mapping points of inflection of said phenomena in said scale-space with respect to said first dimension into regions of stability in said scale-space corresponding to intervals undistinguished by occurrence of observable inflections within variation of range in said scale-space.
3. The method according to claim 2 characterized in that said classifying step further comprises identifying from said regions of stability selected regions of maximal stability as corresponding to events of significance.
4. The method according to claim 3 characterized in that said tracking step further comprises registering said identified events with respect to at least said first dimension for approximating said phenomena.
5. An apparatus for filtering a signal represented in the form of unscaled data characterized by:
means receptive of said data for adjusting scale of observation of said data;
means coupled to said scale adjusting means for recognizing occurrence events within said data at a minimum resolution within scale space;
means coupled to said events recognizing means for localizing occurrence of said events with respect to an independent variable of said signal;
means coupled to said localizing means for classifying significance of said events; and means coupled to said classifying means for reconstructing a representation of said signal with respect to said independent variable at a plurality of scales based on significance classification of recognized events.
means receptive of said data for adjusting scale of observation of said data;
means coupled to said scale adjusting means for recognizing occurrence events within said data at a minimum resolution within scale space;
means coupled to said events recognizing means for localizing occurrence of said events with respect to an independent variable of said signal;
means coupled to said localizing means for classifying significance of said events; and means coupled to said classifying means for reconstructing a representation of said signal with respect to said independent variable at a plurality of scales based on significance classification of recognized events.
6. The apparatus according to claim 5 characterized in that said reconstruction means is operative to approximate transition between event regions defining an interval with respect to said independent variable.
7. The apparatus according to claim 5 or 6 characterized in that said reconstruction means is operative to register events designated by said classifying means as being of most significance within intervals of said independent variable.
8. A method for classifying phenomena expressed as detectable signals characterized by the steps of:
observing occurrences of events of phenomena within a range of observation resolution;
establishing indicia representative of selected ones of said occurrences as a function of scale of said events; and registering said indicia at the point of occurrence at a maximum of observation resolution.
observing occurrences of events of phenomena within a range of observation resolution;
establishing indicia representative of selected ones of said occurrences as a function of scale of said events; and registering said indicia at the point of occurrence at a maximum of observation resolution.
9. An apparatus for classifying phenomena expressed as detectable signals, said apparatus characterized by:
means for observing as a signal of at least a first dimension phenomena within a range of observation resolution;
means for coupled to said observing means for mapping said signal into scale-space, said phenomena appearing as components of said signal;
means coupled to said mapping means for classifying said components as events occurring within a range in said scale-space; and means for tracking said events with respect to said scale-space to a minimum scale value in said scale-space to locate positions of said events with respect to at least said first dimension.
means for observing as a signal of at least a first dimension phenomena within a range of observation resolution;
means for coupled to said observing means for mapping said signal into scale-space, said phenomena appearing as components of said signal;
means coupled to said mapping means for classifying said components as events occurring within a range in said scale-space; and means for tracking said events with respect to said scale-space to a minimum scale value in said scale-space to locate positions of said events with respect to at least said first dimension.
Applications Claiming Priority (2)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| US06/494,194 US4658372A (en) | 1983-05-13 | 1983-05-13 | Scale-space filtering |
| US494,194 | 1983-05-13 |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CA1210153A true CA1210153A (en) | 1986-08-19 |
Family
ID=23963445
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CA000454114A Expired CA1210153A (en) | 1983-05-13 | 1984-05-11 | Scale-space filtering |
Country Status (5)
| Country | Link |
|---|---|
| US (1) | US4658372A (en) |
| EP (1) | EP0129452B1 (en) |
| JP (1) | JPH0721830B2 (en) |
| CA (1) | CA1210153A (en) |
| DE (1) | DE3485960T2 (en) |
Families Citing this family (20)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| USRE38716E1 (en) | 1984-12-20 | 2005-03-22 | Orbotech, Ltd. | Automatic visual inspection system |
| US5774573A (en) * | 1984-12-20 | 1998-06-30 | Orbotech Ltd. | Automatic visual inspection system |
| USRE38559E1 (en) * | 1984-12-20 | 2004-07-27 | Orbotech Ltd | Automatic visual inspection system |
| US5774572A (en) * | 1984-12-20 | 1998-06-30 | Orbotech Ltd. | Automatic visual inspection system |
| US4803736A (en) * | 1985-11-27 | 1989-02-07 | The Trustees Of Boston University | Neural networks for machine vision |
| US4905296A (en) * | 1986-07-22 | 1990-02-27 | Schlumberger Systems & Services, Inc. | System for shape recognition |
| FR2615288B1 (en) * | 1987-05-12 | 1989-09-22 | Commissariat Energie Atomique | AUTOMATIC METHOD OF IDENTIFYING DEFECTS BY ULTRASOUND AND SYSTEM THEREOF |
| EP0603425A1 (en) * | 1992-12-22 | 1994-06-29 | International Business Machines Corporation | Representation of n-ary trees |
| US5815596A (en) * | 1994-04-14 | 1998-09-29 | Narendra Ahuja | Multiscale image edge and region detection method and apparatus |
| US6615680B1 (en) * | 1997-01-08 | 2003-09-09 | Lear Automotive Dearborn, Inc. | Method of testing switch design to quantify feel |
| BR9805971A (en) * | 1997-02-06 | 1999-08-31 | Koninkl Philips Electronics Nv | Process and system for object segmentation and tracking |
| US7318051B2 (en) * | 2001-05-18 | 2008-01-08 | Health Discovery Corporation | Methods for feature selection in a learning machine |
| US7624074B2 (en) * | 2000-08-07 | 2009-11-24 | Health Discovery Corporation | Methods for feature selection in a learning machine |
| US7137062B2 (en) | 2001-12-28 | 2006-11-14 | International Business Machines Corporation | System and method for hierarchical segmentation with latent semantic indexing in scale space |
| FR2977679B1 (en) * | 2011-07-06 | 2013-08-16 | Rockwell Collins France | METHOD AND DEVICE FOR DETECTING A TARGET MASKED BY HIGH ENERGY REFLECTORS |
| KR101528870B1 (en) * | 2013-05-23 | 2015-06-16 | 한림대학교 산학협력단 | Apparatus for detecting copy number variation and method thereof |
| US9460515B2 (en) | 2013-10-25 | 2016-10-04 | Ricoh Co., Ltd. | Processing of light fields by transforming to scale and depth space |
| US9955861B2 (en) | 2015-10-16 | 2018-05-01 | Ricoh Company, Ltd. | Construction of an individual eye model using a plenoptic camera |
| US10136116B2 (en) | 2016-03-07 | 2018-11-20 | Ricoh Company, Ltd. | Object segmentation from light field data |
| CN106886787B (en) * | 2017-03-07 | 2019-05-03 | 华中科技大学 | A kind of extraction of process signal data and storage method based on two-dimensional convolution theory |
Family Cites Families (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US4209853A (en) * | 1974-07-22 | 1980-06-24 | Hyatt Gilbert P | Holographic system for object location and identification |
| US3964064A (en) * | 1972-09-21 | 1976-06-15 | Brandao Ruy L | Data display system having a multilevel video storage device |
| US4062012A (en) * | 1974-07-11 | 1977-12-06 | The United States Of America As Represented By The Secretary Of The Navy | Digital radar signal processor |
| JPS5857758B2 (en) * | 1979-09-28 | 1983-12-21 | 株式会社日立製作所 | Audio pitch period extraction device |
| US4320400A (en) * | 1979-10-12 | 1982-03-16 | Chasek Norman E | Method and apparatus that nearly instantaneously samples, processes and sorts received signals from very large numbers of radar and radio emitters for frequency and angle of arrival analysis |
| US4326259A (en) * | 1980-03-27 | 1982-04-20 | Nestor Associates | Self organizing general pattern class separator and identifier |
| US4453218A (en) * | 1980-11-24 | 1984-06-05 | Oximetrix, Inc. | Signal filter method and apparatus |
| US4396903A (en) * | 1981-05-29 | 1983-08-02 | Westinghouse Electric Corp. | Electro-optical system for correlating and integrating image data from frame-to-frame |
-
1983
- 1983-05-13 US US06/494,194 patent/US4658372A/en not_active Expired - Lifetime
-
1984
- 1984-05-09 EP EP84400945A patent/EP0129452B1/en not_active Expired - Lifetime
- 1984-05-09 DE DE8484400945T patent/DE3485960T2/en not_active Expired - Fee Related
- 1984-05-11 CA CA000454114A patent/CA1210153A/en not_active Expired
- 1984-05-14 JP JP9476684A patent/JPH0721830B2/en not_active Expired - Lifetime
Also Published As
| Publication number | Publication date |
|---|---|
| DE3485960T2 (en) | 1993-05-06 |
| US4658372A (en) | 1987-04-14 |
| EP0129452A3 (en) | 1988-08-31 |
| EP0129452A2 (en) | 1984-12-27 |
| EP0129452B1 (en) | 1992-10-21 |
| DE3485960D1 (en) | 1992-11-26 |
| JPH0721830B2 (en) | 1995-03-08 |
| JPS6037075A (en) | 1985-02-26 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CA1210153A (en) | Scale-space filtering | |
| Kaplan | Extended fractal analysis for texture classification and segmentation | |
| Ma et al. | EdgeFlow: a technique for boundary detection and image segmentation | |
| Förstner | A framework for low level feature extraction | |
| Borah et al. | Wavelet transform based image texture analysis for size estimation applied to the sorting of tea granules | |
| Wiltschi et al. | An automatic assessment scheme for steel quality inspection | |
| Harris et al. | Evolving edge detectors with genetic programming | |
| Al-Sit et al. | Visual texture for automated characterisation of geological features in borehole televiewer imagery | |
| Kasparis et al. | Segmentation of textured images based on fractals and image filtering | |
| CN106815819B (en) | More strategy grain worm visible detection methods | |
| Dickscheid et al. | Coding images with local features | |
| Clark et al. | Texture segmentation using a class of narrowband filters | |
| Rodriguez-Sánchez et al. | The RGFF representational model: a system for the automatically learned partitioning of" visual patterns" in digital images | |
| Zeng et al. | Texture representation based on pattern map | |
| Fritsch | Medial description of gray-scale image structure by gradient-limited diffusion | |
| WO2007004868A1 (en) | Method and apparatus for image characterization | |
| Rosin et al. | Extracting natural scales using Fourier descriptors | |
| Shi et al. | A wavelet-based image edge detection and estimation method with adaptive scale selection | |
| Harris et al. | Evolving edge detectors | |
| Nathalie et al. | Towards the introduction of human perception in a natural scene classification system | |
| Rosenfeld et al. | Picture recognition and scene analysis | |
| Syeda-Mahmood | Detecting perceptually salient texture regions in images | |
| Kakadiya et al. | Shape Extraction methods for fruits: technical review | |
| Choraś | Content-based retrieval using color, texture, and shape information | |
| Boudani et al. | Content-Based Image Retrieval for Surface Defects of Hot Rolled Steel Strip Using Wavelet-Based LBP |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| MKEX | Expiry |