Image analysis – Image transformation or preprocessing – Changing the image coordinates
Reexamination Certificate
2000-11-13
2004-07-13
Couso, Yon J. (Department: 2621)
Image analysis
Image transformation or preprocessing
Changing the image coordinates
C382S295000
Reexamination Certificate
active
06763148
ABSTRACT:
FIELD OF THE INVENTION
This invention relates generally to image processing and, in particular, to methods whereby still or moving images or other objects are transformed into more compact forms for comparison and other purposes.
BACKGROUND OF THE INVENTION
There are many systems in common use today whose function is automatic object identification. Many make use of cameras or scanners to capture images of objects, and employ computers to analyze the images. Examples are bill changing machines, optical character readers, blood cell analyzers, robotic welders, electronic circuit inspectors, to name a few. Each application is highly specialized, and the detailed design and implementation of each system is finely engineered to the specific requirements of the particular application, most notably the visual characteristics of the objects to be recognized. A device that is highly accurate in recognizing a dollar bill would be worthless in recognizing a white blood cell.
The more general problem of identifying an image (or any object through the medium of an image) based solely upon the pictorial content of the image has not been satisfactorily addressed. Considering that the premier model for a generalized identification system is the one which we all carry upon our shoulders, i.e., the human brain, it is not surprising that the general system does not yet exist. Any child can identify a broad range of pictures better than can any machine, but our understanding of the processes involved are so rudimentary as to be of no help in solving the problem.
As a result, the means that have been employed amount to the shrewd applications of heuristic methods. Such methods generally are derived from the requirements of a particular problem. Current technology often uses such an approach to successfully solve specific problems, but the solution to the general image identification problem has remained remote.
The landscape of the patent literature referring to image identification is broad, but very shallow. The following is a summary of two selected patents an three commercial systems which are considered to represent the current state-of-the-art.
U.S. Pat. No. 5,893,095 to Jain et al presents a detailed framework for a pictorial content based image retrieval system and even presents this framework in representative hardware. Flowcharts are given describing the operation of the framework system. The system depends for identification upon the matching of visual features derived from the image pictorial content. Examples of these visual features are hue, saturation and intensity histograms; edge density; randomness; periodicity; algebraic moments of shapes; etc. Some of these features are computed over the entire image and some are computed over a small region of the image. Jain does not reveal the methods through which such visual features are discerned. These visual features are expressed in Jain's system as “primitives”, which appear to be constructed from the visual features at the discretion of a human operator.
A set of primitives and primitive weightings appropriate to each image is selected by the operator and stored in a database. When an unknown image is presented for identification it can either be processed autonomously to create primitives or the user can specify properties and/or areas of interest to be used for identification. A match is determined by comparing the vector of weighted primitive features obtained for the query image against the all the weighted primitive feature vectors for the images in the database.
Given the information provided by Jain, one skilled in the art could not construct a viable image identification system because the performance of the system is dependent upon the skill of the operator at selecting primitives, primitive weightings, and areas of interest. Assuming that Jain ever constructed a functioning system, it is not at all clear that the system described could perform the desired function. Jain does not provide any enlightenment concerning realizable system performance.
U.S. Pat. No. 5,852,823 to De Bonet teach an image recognition system that is essentially autonomous. Image feature information is extracted through application of particular suitable algorithms, independent of human control. The feature information thus derived is stored in a database, which can then be searched by conventional means. De Bonet's invention offers essentially autonomous operation (he suggests that textual information might be associated with collections of images grouped by subject, date, etc. to thereby subdivide the database) and the use of features derived from the whole of the image. Another point of commonality is the so-called “query by example” paradigm, wherein the information upon which a search of the image database is predicated upon information extracted exclusively from the pictorial content of the unknown image.
De Bonet takes some pains to distinguish his technology from that developed by IBM and Illustra Information Technologies, which are described later in this section. He is quite critical of those technologies, declaring that they can address only a small range of image identification and retrieval functions.
De Bonet refers to the features that he extracts from images as the image's signature. The signature for a given image is computed according to the following sequence of operation: (1) The image is split into three images corresponding to the three color bands. Each of these three images is convolved with each of 25 pre-determined and invariant kernels. (2) The 75 resulting images are each summed over the image's range of pixels, and the 75 sums become part of the image's signature. (3) Each of the 75 convolved images is again convolved with the same set of 25 kernels. Each of the resulting 1875 images is summed over its range of pixels, and the 1875 sums become part of the image's signature. (4) Each of the 1875 convolved images it convolved a third time with the same set of 25 kernels. The resulting 46,875 images are each summed over the image's range of pixels, and the 46,875 sums become part of the original image's signature.
In the simplest case, then, the 48,825 sums (46,875+1875+75) serving as the signature are stored in an image database, along with ancillary information concerning the image. It should be noted that this description was obtained from DeBonet's invention summary. Later, he uses just the 46,875 elements obtained from the third convolution. An unknown image is put through the same procedure. The signature of the unknown image is then compared to the signatures stored in the database one at a time, and the best signature matches are reported. The corresponding images are retrieved from an image library for further examination by the system user.
In a somewhat more complex scenario, it is posited that the system user has a group of images that are related in some way (all are images of oak trees; all are images of sailboats; etc.). With the signatures of each member of the group already calculated, the means and variances of each element of their signatures (all 48,825) are computed, thereby creating a composite signature representing all member images of the group, along with a parallel array of variances. When a signature in the database is compared to a given signature, the difference between each corresponding element of the signatures is inversely weighted by the variance associated with that element. The implicit assumption upon which the weighting process is based is that elements exhibiting the least variance would be the best descriptors for that group. In principle, the system would return images representative of the common theme of the group.
Additionally, such composite signatures can be stored in the image database. Then, when a signature matching a composite signature is found, the system returns a group of images which bear a relation to the image upon which the search was based.
The system is obviously very computation-intensive. De Bonet used a 200
Dargel William
Lennington John W.
Sternberg Stanley R.
Voiles Thomas
Couso Yon J.
Gifford Krass Groh Sprinkle Anderson & Citkowski PC
Visual Key, Inc.
LandOfFree
Image recognition methods does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Image recognition methods, we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Image recognition methods will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-3243552