Image analysis – Image segmentation
Reexamination Certificate
1999-04-09
2001-12-18
Boudreau, Leo (Department: 2621)
Image analysis
Image segmentation
C382S305000
Reexamination Certificate
active
06332037
ABSTRACT:
TECHNICAL FIELD
The present invention relates generally to methods of data characterization, storage and retrieval, and more particularly is a method which enables easy characterization, storage and retrieval of multi-dimensional data structures, comprising use of a translation, rotation and scaling invariant index which results from concatenating a series of Eigenvalue calculation mediated index elements, determined at a plurality of hierarchical data depth levels.
BACKGROUND
Data structures, such as two dimensional pixel arrays, are being generated at an ever increasing rate. For instance, algorithm generated and scanned computer screen images, X-ray, CT, MRI and NASA satellite, space telescope and solar explorer systems generate thousands of images every day. To make optimum use of said images, however, convenient methods of data characterization, storage and retrieval are required. For example, a medical doctor might obtain an X-ray image of a patient's chest but has to rely on “diagnostic art” to arrive at a diagnosis. Were it possible to determine an index which characterizes said X-ray image and also enable easy storage and retrieval thereof, it would be possible to compare said index to a catalog of indices of various X-ray images which are known to be associated with various healthy or pathologic conditions. Thus diagnosis could be moved toward the very desirable goal of being objectively definite in a mathematical sense.
Continuing, it must be understood that conventional data bases are stored as text with organization being in terms of fields and values. Examples are business product, customer lists, sales data etc. To retrieve such data a user must issue a query in text format, similar to what is done in natural languages. It is essentially impossible to use such an approach to store and retrieve the contents of most data images, for example, because there is no convenient manageable way to describe such data images in terms of said fields and values. Data Images are instead typically stored in the form of compressed digital files of hundreds of thousands of binary numbers, and said storage technique does not facilitate easy image characterization, storage and retrieval. And, while it is possible to describe a data image with a text Index, to examine the data image data still requires that the data associated with said Index be retrieved. It is also possible to assign an arbitrary serial number to a data image to facilitate data storage and retrieval, but under this approach the serial number provides no insight to the image and again, to examine data image, requires accessing the image data per se.
A preferred approach to the characterization of data images, which provides an index for use in storage and retrieval thereof, is to base the index on features in the data image. To arrive at such an index, however, is typically computationally complex, requiring hundreds of thousands of calculations. That is, determination of said index must typically be extracted from a data image “off-line”. Characteristic indices so determined are called “image indices”, and ideally render a concise description, not only of an image color and intensity content on a row and column basis, but also of the nature and shape of objects therein. A problem arises, however, in that many image features can not be easily described. Geometric shapes in a data image, for example, can require a combination of text annotation and numeric values and often the result is not at all concise.
Relevant considerations in developing an approach to extracting “image indices” from a data image or data set include:
1. Uniqueness—different images/sets should have different associated image indices, (ie. an image index should be non-degenerate);
2. Universality—image/set indices must be extractable from essentially any kind of image to be characterized, stored and retrieved by use thereof;
3. Computation—image/set indices must be easily computed from any data image to be characterized, stored and retrieved by use thereof;
4. Conciseness—image/set indices must concise and easy to store;
5. Invariance—descriptive features in a data image/set must tolerate change of scale, rotation and translation transformations, image object position shifting, calibration of color and pixel intensity and return essentially unchanged image indices;
6. Noise resistant—random noise entry to image/set data should not significantly change the image index extracted therefrom.
Previous attempts at extracting an image index for image/set data have focused on use of:
pixel intensity and color distributions, (see an article titled “Query By Image And Video Content: The QBIC System)”, IEEE Trans. on Computers, (Sep. 1995));
pixel texture patterns (see a book titled “Digital Image Processing”, Gonzales, Addison-Wesley Pub. (1992)); and
edge and boundary-line shapes, (see a book titled “Digital Image Processing And Computer Vision”, Schalkoff, John Wiley & Sons, (1989)),
etc. as the basis of approach. These techniques are mainly based on the calculation of the statistics of a data image in a pixel arrangement. Said techniques often lack Universality in that they work when applied to a certain type of data image, but not when applied to other types of data images. Moreover, many previous approaches are not image transformation invariant and do not tolerate entry of noise.
Continuing, one approach which provides a rotationally invariant result is termed “Equal Angular Sampling”. Said method provides a concatenation of numbers which are distances from a centroid in a data image to an intersection point with an object boundary. Said technique encounters problems, however, where objects with irregular shapes, with concave boundaries and/or wherein holes are encountered.
The use of Moment Invariants to describe the geometrical shape features of data images was proposed more than thirty (30) years ago by Hu in an article titled “Visual Pattern Recognition By Moment Invariants”, IRE Trans. on Information Theory, IT-8, (February 1963). The method is based in modeling an image as a physical object with masses distributed in two dimensional space. It typically treats the pixel intensities as the probability distribution value of the object masses. The central moments in various orders are calculated on distributions. A set of moment invariants is derived from making algebraic combinations of the moments. The most important property of he technique is that the resulting descriptive quantities are transformation invariant, (ie. the moment invariants remain unchanged when the image undergoes scaling, rotation, translation, intensity, or color platter changes). See an article titled “Recognitive Aspects Of Moment Invariants”, by Abu-Mostafa et al., IEEE Trans. on Pattent Analysis and Mach Intell., Vol. PAMI-6, No. 6, (November 1984).
Additional references of interest are:
“Image Analysis Via the General Theory Of Moments”, Teague, J. Opt. Soc. America, Vol. 70, No. 8, (Aug. 1980), which discloses that a 2D shape obtained from moment invariants defined on the second central moments can be viewed as an elliptic approximation of the shape; and
“A Transformation-Invariant Recursive Subdivision Method For Shape Analysis”, Zhu and Poh, IEEE Proc. of the 9th Int. Conf. on Pattern Recog., Rome, Italy, (Nov. 14-17, 1988).
Continuing, it is to be appreciated that Statistical and Moment-based descriptions of data can distinguish data images at only very rough levels. That is, an image index associated with a data image is not unique and could be arrived at by analysis of an alternative data image. In addition, the computations involved in practicing said Statistics and Moment-based approaches can be complicated and time consuming and can require both character and numeric symbols in a resultant image index. And the use of the moment invariant approach can involve the computation of an image index in high orders.
With the present invention in mind a Search of Patents was performed, with the result being that very little was found. A Patent to Windig, U.S. Pat.
Board of Regents of the University of Nebraska
Boudreau Leo
Patel Kanji
Welch James D.
LandOfFree
Invariant, Eigenvalue based, non-degenerate data structure... does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Invariant, Eigenvalue based, non-degenerate data structure..., we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Invariant, Eigenvalue based, non-degenerate data structure... will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-2569675