SYSTEM AND METHOD FOR VISUAL RECOGNITION
A method for visual recognition of an object in an electronic image includes extracting unique points of an object to be learned and/or a target object. The unique points are obtained by cross-correlating the image with a structure. Generally, the structure and/or the size of the structure may vary to detect extremum information associated with the learned object and/or target object. An icon corresponding to each of the unique points is extracted. The size of the icon corresponds to the scale of the unique point. After extraction of the various icons, an object becomes a collection of icons. Each of these icons is un-rotated and normalized or resized to a constant size so it can be compared with other icons. One of the unique properties of these icons is their stability over scale and angle. Thus, this invention allows the recognition of an image(s) or object(s) from large number of trained images or objects very quickly.
RELATED APPLICATION DATA
This application is a divisional of U.S. application Ser. No. 13/437,370 filed on Apr. 2, 2012, which is a continuation application of U.S. application Ser. No. 12/101,583, filed on Apr. 11, 2008 and issued as U.S. Pat. No. 8,150,165 on Apr. 3, 2012.
TECHNICAL FIELD OF THE INVENTION
The present invention relates to the field of computer vision, and more particularly, to a system and method for visual recognition for use in a wide variety of applications.
DESCRIPTION OF THE RELATED ART
Computer vision generally relates to the theory and technology for building artificial systems that obtain information from images or multi-dimensional data. As used herein “information” means anything that enables a decision to be fully and/or partially based. Exemplary computer vision applications include: visual object recognition and scene interpretation, particularly for image retrieval, video indexing, controlling processes (e.g. an industrial robot or autonomous vehicle such as unmanned aerial/ground/see vehicle), detecting events (e.g. for visual surveillance), organizing information (e.g. for indexing databases of images and image sequences), Image based internet search (e.g., searching for similar image on the Internet), modeling objects or environments (e.g. medical image analysis or topographical modeling), interaction (e.g. as the input to a device for computer-human interaction), etc.
A goal of computer vision is to make a computer truly “see” just like humans do. Understanding the content of everyday images and videos is one of the fundamental challenges of computer vision. In order to make a computer “see” in an unconstrained environment an extraordinary amount of computational power, perhaps on the order of 1015 operations per second likely is needed. Even if such a speed was possible in a commercial computer vision system, it is difficult to perform rapid visual searches in unconstrained, natural environments.
To make search and recognition tasks tractable in commercial computer vision, designers typically limit the task's visual complexity. This may be done in a variety of ways. For example, the vision system may be set up to view and recognize only one or a small class of objects. Second, the presentation (position, orientation, size, view, etc.) of these objects is strictly controlled. Thus, the object variability is limited to the point that the vast majority of variables are eliminated and the search can be implemented with reasonable cost in terms of both computing time and money.
Computer vision systems generally lack the knowledge needed to constrain and interpret a general visual search (e.g., searches performed in an uncontrolled environment). Therefore, practical computer vision search requires the designer to drastically restrict what the vision system sees and to add a priori knowledge about what it will see so that it can interpret the result. Thus, a major drawback to computer vision in real world applications is the time, money and specialized knowledge needed for such applications to be adequately performed.
The evolution of computer vision in the last twenty years was driven by improvements in hardware and algorithms. A variety of computer vision methods have been developed for image detection (also referred to herein as pattern recognition). These techniques include, for example, using binary images to represent gray scale images, normalized grayscale correlation, blob analysis, geometric based search and recognition, contour based search, affine invariant constellation based recognition, corner detection, salient icon detection, scale invariant feature transform, etc.
A strong need exists in the art of computer vision to recognize objects in an image or image sequence similar to vision in human beings. For example, in an airport, an unmanned vehicle needs to recognize other vehicles and obstacles so it can avoid and/or maneuver through the airport. In an unmanned vehicle or other robotic vision application, the robotic application generally needs to “see” the pathway and navigate autonomously or land autonomously.
In visual recognition, achieving invariance to object presentation (position, orientation, distance (scale), and perspective), lighting, occlusion and background is challenging. Aspects of the present invention provide excellent invariance to object presentation, lighting, occlusion and background and generalization for true object recognition.
The human brain processes visual information associated with objects with full independency of the position, orientation, distance (scale), and perspective. For example, if a human being views a “soda pop can”, the human can recognize it regardless of the distance and/or orientation (e.g., distance from can, rotation, tipped, tilted, etc.). The brain essentially “normalizes the view”. Humans are capable of learning a large number of objects and easily retrieve the learned objects. Aspects of the present invention allow learning virtually an unlimited number of objects and recognizing any one of these learned object(s) regardless of object presentation. This is analogous to human visual recognition capability. For example, aspects of the invention enables the recognition of hundreds of trained objects very quickly (e.g., in less than a second) and fundamentally has no limit in learning and recognizing millions of objects. This capability stems from the ability to extract the same icons (image patches) from an image of an object regardless of distance, rotation, presentation that the object is in relation to the viewer and/or the device acquiring the image or series of images.
Aspects of the invention relate to extracting unique points (e.g., x and y coordinate points) in an image. Each one of these unique points has its own unique scale (e.g., size) and orientation that is related directly to the presentation of the object. Having scale and orientation information measured per unique point enables visual recognition that is fully invariant to presentation. In other words, when an object is closer, farther, rotated, tipped, and/or tilted, these unique points have similar relative locations to the object and a unique scale that is related to how close/far the object is and rotation values that are related directly to the object planar rotation. Basically these unique points “normalize the view” of the object.
An icon (image patch) from an image of an object is extracted from each of these unique points. The size of the icon corresponds to the scale of the unique point. And the angle of the icon is the angle of the unique point. After extraction of the various icons, an object becomes a collection of icons. Each of these icons is un-rotated by icon angle and resized to a constant size so it can be compared (distance measure such as absolute difference) one-to-one with other icon (also referred to herein as “normalized”. It has been determined that the icons are virtually identical regardless of object presentation. In other words, the icons (image patches) are the same whether the object is close or far, rotated, tilted, and/or tipped. One of the unique properties of these icons is their stability over scale and angle. Comparing an icon for similarity may also include color information. Generally, when comparing two icons, each icon may also be intensity-normalized.
Searching for an object in database of learned object's images becomes a search of vectors associated with learned object's images. Indexing techniques are one way represent an image for searching.
Computing geometric transformation between a learned object and a found object is done by computing the transformation between the corresponding learned icon's position and found icon's position, as discussed below. The transformation matrix between learned object and found object is computed using a perspective matrix using least square of all corresponding icons positions or by picking two sets of quad icons position from the learned and found objects. Based on rigid body assumptions, every set of four icons can compute a perspective matrix. Many sets of four icons give the same transformation, which provides a robust measure of correct match, also referred to herein as, measure redundancy, as discussed below.
One aspect of the present invention relates to a method for visual recognition of at least one object in an image, the method comprising: providing an image in an electronic format, wherein the image includes at least one object to be learned; generating extremum information associated with the image by cross-correlating at least one structure across at least a portion of the image, wherein the extremum information includes at least one coordinate point associated with cross-correlating the at least one structure across the image; extracting at least one icon from the image, wherein the icon includes the coordinate point associated with the extremum information; determining an angle associated with the at least one icon; normalizing the icon to a fixed size; and storing icon information in a computer readable form, wherein the icon information includes image values associated with at least a portion of the icon; the at least one coordinate point associated with the extremum information; and the angle associated with the at least one icon.
Another aspect of the invention relates to a method for matching a learned object with a target object, the method comprising: providing at least one learned object and at least one target object, wherein the learned object and the target object; extracting unique points from the target object, wherein the unique points are generated from extremum information obtained from the target image by cross-correlating at least one structure across the target image; extracting an icon of the target image corresponding to each of the unique points; determining an angle associated with the at least one icon; normalizing the extracted icon; and determining if the extracted icon from the target images matches a learned object.
Another aspect of the invention relates to a program stored on a machine readable medium, the program being suitable for use in matching a learned object with a target object, wherein when the program is loaded in memory of an associated computer and executed, causes extracting unique points from the target object, wherein the unique points are generated from extremum information obtained from the target image by cross-correlating at least one structure across the target image; extracting an icon of the target image corresponding to each of the unique points; determining an angle associated with the at least one icon; normalizing the extracted icon; and determining if the extracted icon from the target images matches the learned object.
Other systems, devices, methods, features, and advantages of the present invention will be or become apparent to one having ordinary skill in the art upon examination of the following drawings and detailed description. It is intended that all such additional systems, methods, features, and advantages be included within this description, be within the scope of the present invention, and be protected by the accompanying claims.
It should be emphasized that the term “comprise/comprising” when used in this specification is taken to specify the presence of stated features, integers, steps or components but does not preclude the presence or addition of one or more other features, integers, steps, components or groups thereof.”
BRIEF DESCRIPTION OF THE DRAWINGS
The foregoing and other embodiments of the invention are hereinafter discussed with reference to the drawings. The components in the drawings are not necessarily to scale, emphasis instead being placed upon clearly illustrating the principles of the present invention. Likewise, elements and features depicted in one drawing may be combined with elements and features depicted in additional drawings. Moreover, in the drawings, like reference numerals designate corresponding parts throughout the several views.
DETAILED DESCRIPTION OF EMBODIMENTS
The present invention is directed to a system and method for pattern identification of a learned image (or learned pattern) in a target image. Unique and stable points (e.g., x and y coordinate points) are extracted from an image of an object, regardless of object presentation. The uniqueness of the extracted points comes from the fact that these points have the same relative position in the object regardless of distance, orientation (e.g., tip, tilt, rotation, etc.) and illumination of the object from the viewer or viewing device. In other words, the extracted points are invariant to object presentation.
Early computer vision algorithms generally used an image subtraction method (also referred to as golden template matching) as a primitive method of recognition. Image subtraction is a form of distance measure between two images. For the image subtraction method to work, the object in the learned image has to be nearly identical to the object in the target (scene) image. For example, the object generally has to be in the same position, same scale, same planar angle, etc. as the learned object. Any shift in location or other transformations would produce a false result. Such conditions were generally needed because the subtraction method simply subtracted pixel values having coordinates in one image with pixel values located at corresponding coordinates in another image.
If a method can find regions in an object image that are the same regardless of object presentation, it is conceivable that the golden template concept may be used to recognize regions of that object, which could result in recognizing the entire object. Prior methods to find these unique points and associated icons produced either unstable points in position, unstable scale, and/or unstable angle. The prior methods produced few inliers and majority of outliers, which makes pattern recognition generally problematic. Researchers in academia have experimented with several types of these unique points. Example of some of the famous type of these points are, Harris-Corner-Detector, Harris-Laplace, Laplacian-of-Gaussian, SIFT (Difference of Gaussian). Generally these detectors lack stability in position, scale and angle, which produces various problems during the recognition process. For example, these prior methods produce hundreds and/or thousands of unique points and only a handful of such points may survive from one object position to another to aid in pattern recognition.
An invariant point in an image generally needs to have the characteristic of extremum for some metric. For example, in a one dimensional signal, such as a parabola, the peak point of a parabola is an invariant point regardless to the parabola's parameters (e.g., regardless to how wide, narrow, shifted, or rotated a parabola is).
Structure 12 may be circular, conic or Gaussian shape. In one embodiment, the structure 12 is a cone-like structure. Referring to
The cone-like structure 12 has a cone angle θ that generally corresponds to the height (h) of the structure 12. For example, SRI points generally have very interesting properties that aid in recognition. Based on correlation score between cone-like structure (or any other desired structure) and the image: a correlation score close to 1.0 identifies round structures; a correlation score of about 0.5 identifies strip like structures; and a correlation score of about 0.7 identifies an end of strip like structure. One of ordinary skill in the art will readily appreciate that the shape of the structure chosen may correspond to a change in correlation score for the identified structure.
The structure 12 is attempted to be “inserted” into the steel plate 14 in a scanned manner (e.g., one pixel after another pixel). As shown in
Once the image has been cross-correlated, a resulting set of unique points (also referred to herein as x and y coordinate points) and radius values (scale) for each of the points are obtained, as is illustrated by the white circles in each of the images. As stated above, any structure 12 may be used in accordance with aspects of the present invention. For example, a two dimensional Gaussian provided similar results as the cone-like structure. The cone-like structure 12 is utilized to find extremum in the object regardless of scale. This provides robust and reliable relative localization (e.g., x and y coordinate positions in the image), and scale.
Once the correlation image is computed a peak detector is applied to find the coordinate of peaks in sub pixel form, as illustrated in
Once the unique points (e.g., coordinate values) are determined, an icon (image patch) is extracted from the associated at each of these coordinates. The icon size is proportional to the radius (scale) of the cone structure having a high cross-correlation match value. The angle of the icon 50 is computed from a vector between the unique point position 54 and the grayscale centroid position 52 of the icon at scale, as shown in
Once extracted, each one of the icons is normalized to a fixed size, as shown in
Each learned object may be described by a set of icons. Each icon generally includes one or more values, for example: (x,y) coordinate position, a size that correspond to the size of the image structure from which the icon originated from, and an angle. For example,
At block 158, the icon for each for each of the coordinates is computed for the scale. At block 160, given the icon angle and scale for each of the coordinates, extract icons from the image. At block 162, the icons are normalized to a fixed size. At block blocks 154 through 160 are repeated until all icons have been extracted. At block 164, a constant (K) of highly stable and invariant icons that represent the object are obtained and may be stored in a memory or other suitable storage device or pumped into an indexing data base or hash table.
An exemplary method 200 for learning an object is illustrated in
At block 206, SRI points acquired from the image and/or object to be learned are used to extract icons associated with the SRI points. In one embodiment, each icon has its (x, y) coordinate, size (scale) and angle. At block 208, a descriptor for each icon is created. In addition or in the alternative, each icon may also be tagged or otherwise associated with a learned object name.
At block 210, similar icons are found and tagged. Similar icons are generally suitable for recognition, but not unique enough for locating the object unless the spatial relation between the icons is applied such as, for example, nearest neighbor icon; n nearest neighbors; left, top, bottom, left neighbor; etc. Similar icons may have multiple correspondences. Blocks 202-210 are repeated for every object needed to be learned. Once learned, an object becomes a collection of icons (or their descriptors) and the spatial relation that ties the icons together is a set of object contour points. The icon coordinates also can be used for detecting, determining and/or verifying special relationships between the icons.
Using descriptors provides a variety of advantages. Such advantages include, for example, permitting the use of indexing techniques for fast retrieval of similar icons in a database of icons, which hastens retrieval of similar objects. This functionality is highly desirable when recognizing an object(s) from a large database of objects.
During the recognition phase, an object or multiple objects may exist in an image (scene image) and one goal is to recognize the object or multiple objects and provide the x and y coordinates of each object. An exemplary recognition method 250 is illustrated in
At block 262, candidate objects in the image are identified and ranked based on the quality of the match and/or the number of matches. At block 264, the location of the object or objects is determined. At block 266, based on the learned icon or icons of an identified object or objects, corresponding icons in the recognized icons are found. This may be accomplished by a variety of methods.
One exemplary method begins at block 268. At block 268, all corresponding icons are used to compute a perspective transformation. This may be done by generating one or more hypotheses. The hypotheses may be generated in any desired manner. For example, all icons of learned object(s) may be selected, and compared with corresponding matched icons. Using a least square method correlation method or another correlation method, a perspective transform is generated between the learned unique point's coordinates and the corresponding matched points in the scene image. A least squares correlation method is generally preferred because there may be some outliers, at this juncture.
At block 270, the sampled learned contour points of candidate objects are transformed and superimposed on the scene image for verification. A match score between sampled contour and scene contour is computed. The transformation having the best contour match is selected. At block 272, using the transformation of the best match, all unique points of the learned object are transformed with the perspective transform onto the image. In addition, the distance between these transformed points and scene object unique points are computed. At block 274, any outlier corresponding icons are removed and the transform having the best match is saved for later use. Blocks 270 through 274 are repeated N times, where N is the number of corresponding icon points. Once this method has been applied N times, the transform having the best match is saved for later use.
Another method for obtaining precise localization of an icon is illustrated in blocks 280-286. Referring to block 280, for every quad set of corresponding icons points (e.g., coordinates), a perspective transform is computed using least squares or some other correlation method. At block 282, the object contour points with the perspective transform is transformed, in similar manner as block 270. At block 284, transformed contour points are transformed and superimposed onto the image to verify the transform, as discussed above with respect to block 272. At block 286, the transform producing the best match is saved and the process repeats N times (where N is the number of quad sets).
In operation, a method of generating the hypotheses picks a predetermined number of points (e.g., four points) of the leaned unique points and a predetermined number of corresponding points (e.g., four points) in the scene image. The process is repeated for N number of quad points. For every set of four points the perceptive transform is generated and sampled contour points are transformed and superimposed on the scene contour points. The transform of highest match between learned contours and scene contour is kept as the best transformation transform. For a set of ten matching icons between learned object and found object, there are 210 possible combination and, of course, 210 possible hypotheses generation and verifications. The speed of the recognition process (the entire process) for one learned object is about 140 millisecond using standard off the shelf Pentium based processor with 1.6 GHz processor speed. It is approximately 2 milliseconds extra for every learned object using sequential icon matching. This means, for example, that for 430 learned objects, a recognition would take 1 second ((1000−140)/2). A hash table based indexing would have the potential of achieving recognition of a million objects in one second.
As a practical contribution, the aspects of the present invention may be used in a wide variety of application including, for example, Exemplary computer vision applications include: visual object recognition and scene interpretation, particularly for image retrieval, video indexing, controlling processes (e.g. an industrial robot or autonomous vehicle such as unmanned aerial/ground/see vehicle), detecting events (e.g. for visual surveillance), organizing information (e.g. for indexing databases of images and image sequences), Image based internet search (e.g., searching for similar image on the Internet), modeling objects or environments (e.g. medical image analysis or topographical modeling), interaction (e.g. as the input to a device for computer-human interaction), applications wherein a closed-loop guidance and/or control system is utilized that requires a fast searching algorithm, etc.
Computer program elements of the invention may be embodied in hardware and/or in software (including firmware, resident software, micro-code, etc.). The invention may take the form of a computer program product, which can be embodied by a computer-usable or computer-readable storage medium having computer-usable or computer-readable program instructions, “code” or a “computer program” embodied in the medium for use by or in connection with the instruction execution system. In the context of this document, a computer-usable or computer-readable medium may be any medium that can contain, store, communicate, propagate, or transport the program for use by or in connection with the instruction execution system, apparatus, or device. The computer-usable or computer-readable medium may be, for example but not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, device, or propagation medium such as the Internet. Note that the computer-usable or computer-readable medium could even be paper or another suitable medium upon which the program is printed, as the program can be electronically captured, via, for instance, optical scanning of the paper or other medium, then compiled, interpreted, or otherwise processed in a suitable manner. The computer program product and any software and hardware described herein form the various means for carrying out the functions of the invention in the example embodiments.
Specific embodiments of an invention are disclosed herein. One of ordinary skill in the art will readily recognize that the invention may have other applications in other environments. In fact, many embodiments and implementations are possible. The following claims are in no way intended to limit the scope of the present invention to the specific embodiments described above. In addition, any recitation of “means for” is intended to evoke a means-plus-function reading of an element and a claim, whereas, any elements that do not specifically use the recitation “means for”, are not intended to be read as means-plus-function elements, even if the claim otherwise includes the word “means”. It should also be noted that although the specification lists method steps occurring in a particular order, these steps may be executed in any order, or at the same time.
1. A method for matching a learned object with a target object, the method comprising:
- extracting unique points from the target object, wherein the unique points are generated from extremum information obtained from the target image by cross-correlating at least one structure across the target image;
- extracting an icon from the target image corresponding to each of the unique points;
- normalizing the extracted icon; and
- determining if the extracted icon from the target images matches a learned object.
2. The method of claim 1, wherein the step of determining a match includes searching a database of learned objects to determine a match.
3. The method of claim 1 further including ranking objects detected in the target image based on a match score.
4. The method of claim 3, wherein objects having a higher matching score are processed prior to objects having a lower matching score.
5. The method of claim 1 further including localizing the objects detected in the target image to determine the location of the objects in the target image.
6. The method of claim 5 further including finding one or more corresponding icons in the detected icons from one or more learned icons associated with at least one learned object.
7. The method of claim 6 further including computing a perspective transform for all of the unique points.
8. The method of claim 7 further including transforming object contour points with the perspective transform.
9. The method of claim 8 further including verifying the transformed contour points by superimposing the transformed contour points onto the image.
10. The method of claim 8 further including generating a redundancy measure.
11. The method of claim 5 further including selecting a predetermined number of unique points and computing a perspective transform for the predetermined number of unique points.
12. The method of claim 11 further including transforming object contour points with the perspective transform for the predetermined number of unique points.
13. The method of claim 12 further including verifying the transformed contour points by superimposing the transformed contour points onto the image.
14. The method of claim 5 further including computing a perspective transform for every set of 4 unique points.
15. The method of claim 1, wherein at least one of the extracted icons at least partially overlaps with another icon, where each of the extracted icons have a different scale.
16. The method of claim 15, wherein the extracted icons having a larger scale than the another icon.
17. The method of claim 16, wherein the larger icon has a higher ranking than the another icon have a smaller size.