Image analysis – Pattern recognition – Classification
Reexamination Certificate
2002-10-18
2004-08-10
Patel, Jayanti K. (Department: 2625)
Image analysis
Pattern recognition
Classification
C382S190000, C382S276000, C358S001600, C358S450000
Reexamination Certificate
active
06775411
ABSTRACT:
BACKGROUND OF THE INVENTION
1. Field of the Invention
The present invention generally relates to computer systems for image processing. More particularly, the present invention relates to apparatuses and methods to automatically search image data to detect the presence, or likelihood of presence, of specific target images.
2. Description of the Related Art
There are several systems that analyze data to detect the presence of specific images in data. Analog images can occur in a variety of media including photographs, photographic slides, television images in various formats including HDTV, images on computer monitors, holograms, x-rays, sonograms and radar. The analog imagery can be digitized and thus represented by an array of pixels where each pixel may be determined by one or more spectral bands of digital data. As examples, a digital image can be a 1024×1024 array of pixels where each pixel is specified by a 0 or a 1 denoting black or white respectively, or an integer between 0 and 255 denoting 256 shades of gray, or 3 integers between 0 and 255 denoting a red, blue and green component, respectively with 256 levels for each component, or an integer between 0 and 1023 denoting 1024 infra-red levels, or 256 integers each between 0 and 255 and each denoting the output at a different spectral level from a multi-spectral satellite sensor. Digital imagery can also be created directly through computer graphics software and systems as well as through imaging sensors whose output is natively digital.
There are many existing applications that utilize some form of image recognition within a set of data, such as fingerprint identification, quality control for manufacturing, optical character recognition on scanned documents, automatic target recognition for weapons systems, and face recognition. A critical function of these systems is to indicate the likelihood of a match when a target image is indicated by the image data. And in circumstances where the image data is voluminous, such as satellite data or all images available on the Internet, human review for all of the data is impractical.
The prior art image recognition methods generally fall into two categories: constrained and unconstrained. Constrained methods of image recognition work only with images that have a specific structure. For example, one commercial application that performs face recognition based on biometrics requires the calculation of distances between eyes, and between nose and mouth, and may rely on other like relational computations. Such a face recognition calculation normally does not make sense on a general type of image, such as a boat or airplane, since the underlying structure of an image of boat or plane is not the same as that of a face. More generally, constrained image recognition methods are typically not effective on imagery without the specific structure assumed by the method since the sought image may have very different relational aspects than the target image, yet still satisfy the search criteria.
Constrained methods, however, typically require relatively high quality imagery such that the underlying structure of the images that are analyzed by the recognition method can be observed and this would not be suitable for use in poor quality image data. For example, a drug enforcement agent may take a photograph of an individual in a moving car and obtain a blurry image, or may take a telephoto picture of a subject and later want to identify other people in the background of the photo who may be severely out of focus. Another problem occurs in video imagery of a crowd scene, such as at a football stadium, where very tiny images of large numbers of people are present. Once the imagery is too small, or too out of focus, or too blurred, the calculations required in the constrained biometric face recognition system cannot be performed accurately. More generally, once the quality of imagery is too poor, any constrained system cannot take advantage of its underlying structural assumptions and then fails to perform accurately.
On the other hand, unconstrained image recognition systems do not utilize the underlying structure of an image as a basis for comparison, and thus do not suffer the same problems in reviewing low quality image data. Prior art unconstrained image recognition systems however do not work well with complex imagery. For example, given two images A and B, specified by square pixel arrays of data, where each pixel is specified by an integer between 0 and 255, one measure of the difference, ∥A−B∥, between A and B is the square root of the sum of the square of difference between the corresponding pixel values. In such manner, A and B would be said to be similar when ∥A−B∥ is small. However, this measurement may be very large, even when the difference between A and B is barely noticeable For example, create an image B from an image A where all columns of pixels of image A are shifted to the right by one pixel, and a new column of pixels is added at the left which just duplicates the new second column. If A is a 1024×1024 image, A and B will look very similar to a human observer. For example, if A is a picture of a boat, then B will also look like a boat, slightly shifted to the right. However, ∥A−B∥ may be large since every pixel in A will be different from the corresponding pixel in B. In some practical image recognition applications, it is desirable to recognize whether one or more objects in image A are also present, or absent, from image B, not only when they are slightly modified digitally as in the above example, but when they are acquired under very different conditions. Examples of such disparate acquisition conditions include different image acquisition times leading to different conditions of use, with different cameras or other sensors, from different distances from sensor to object(s) or scenes, at different perspectives, under different lighting conditions, different environmental conditions, and in the context of different backgrounds and in the presence or absence of other objects. Moreover, the target object(s) of interest may be partially obscured in a different manner in A and B, and may be rotated, scaled and translated relative to each other or to the background. Additionally, imaging systems and computer software may further distort the imagery, such as through the application of compression to facilitate storage and transmission and such as through insertion of special effects. Prior art unconstrained image recognition methods do not deal effectively with the complexity arising from significantly different conditions of image data acquisition.
Accordingly, it is desirable to have an improved system for image recognition that can adequately search realistic imagery data for the presence of target image(s) and/or image object(s) and successfully indicate the likelihood of target image(s) and/or image object(s) being present, even when the searchable imagery data and target imagery have significantly different conditions of acquisition. Such system should be unconstrained by the target image or searchable image data structure and allow for variation in the appearance of the target image within image data. It is to the provision of such an apparatus and method for recognizing images within digitized data that the present invention is primarily directed.
SUMMARY OF THE INVENTION
The present invention is an apparatus, method, and computer program that can recognize specific images within a collection of digitized image data, or at least indicate the likelihood that a specific image is contained within the image data. In the system, a processor can either receive image data in a digital format or itself digitize data into target images, the collection of which forms the searchable image data. In the system, a processor also can either receive other image data in a digital format, or itself digitize data into query images. The system then generates a set of domain blocks from one or more query images with each domain block rep
Sloan Alan D.
Xie Ruifeng
Arnall Golden & Gregory LLP
Chawan Sheela C
Patel Jayanti K.
Sloan Alan D.
LandOfFree
Apparatus and method for image recognition does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Apparatus and method for image recognition, we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Apparatus and method for image recognition will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-3345659