Image analysis – Pattern recognition – Classification
Patent
1996-05-06
1997-12-30
Couso, Jose L.
Image analysis
Pattern recognition
Classification
G06K 962
Patent
active
057039640
DESCRIPTION:
BRIEF SUMMARY
BACKGROUND OF THE INVENTION
Pattern recognition systems automatically identify patterns of input data based on patterns previously received. A system is first trained by inputing multiple training patterns and forming categories. After training, the system receives a pattern for identification. It compares the pattern with the categories and, based on the comparisons, identifies the pattern as belonging to one of the categories.
SUMMARY OF THE INVENTION
The present invention is directed to a pattern recognition system and a method for recognizing input data patterns from a subject and classifying the subject. The system first performs a training operation in which the system generates a set or library of categories. During the training operation, input training patterns are received and grouped into clusters. Each cluster of training patterns is associated with a category having a category definition based on the training patterns in the cluster. As each training pattern is received, a correlation or distance is computed between it and each of the existing categories. Based on the correlations, a best match category is selected. The best match correlation is compared to a preset training correlation threshold. If the correlation is above the threshold, then the training pattern is added to the cluster of the best match category, and the definition of the category is updated in accordance with a learning rule to include the contribution from the new training pattern. If the correlation is below the threshold, a new category defined by the training pattern is formed, the cluster of the new category having only the single training pattern.
Training patterns are usually received from multiple classes of subjects. A class is a particular item or person which the network is trained to identify. A category is defined by particular features or views of the subjects. For example, if the system is used to visually classify automobiles by model, each model of automobile would be a separate class. Specific recognizable features of the automobiles, such as fenders of a particular shape or particular tires, could be categories. Since different models (classes) can have similar appearing fenders or tires (categories), the clusters of each category will generally include training patterns from more than one class. That is, since the fenders of different models may appear similar, the cluster of a fender category will include training patterns from more than one model. When training such a system, multiple photographs of each model (class) taken from different views and/or showing different features could be input to the system to form the categories.
In the case of face recognition, each class would be a separate person. The categories could be defined according to particular view orientations or facial features. To train a face recognition system, photographs showing several views of each person of a group of persons could be input to the system. The views could include front, left side and right side views as well as views of individual persons with and without glasses and/or with and without facial hair. Since more than one person can appear similar from a particular view or with a particular facial feature, each view or feature category can include several persons (classes).
Just as several classes can have similar appearing features and will as a result be grouped in feature clusters, features of different classes can also appear very different. As a result, different categories of corresponding features will be formed for different classes. For example, fenders or tires from different models of automobile may appear very different. So, multiple tire categories and fender categories can be formed, each containing training patterns from different models in its cluster. In the same way, different persons will likely appear different even at the same orientation. So, multiple categories will be formed for a single orientation.
It should also be noted that within a single class, multiple views, although taken from th
REFERENCES:
patent: 4292471 (1981-09-01), Kohn et al.
patent: 4326258 (1982-04-01), Cooper et al.
patent: 4658429 (1987-04-01), Orita et al.
patent: 4979225 (1990-12-01), Tsujiuchi et al.
patent: 5075896 (1991-12-01), Wilcox et al.
patent: 5181259 (1993-01-01), Rorvig et al.
patent: 5187751 (1993-02-01), Tanaka
patent: 5263097 (1993-11-01), Katz et al.
patent: 5283838 (1994-02-01), Togawa et al.
patent: 5295198 (1994-03-01), Maney
patent: 5347595 (1994-09-01), Bokser
patent: 5537488 (1996-07-01), Menon et al.
Kohonen, T., "An Introduction to Neural Computing," Neural Networks, vol. 1, No. 1, 1988, pp. 3-16.
Carpenter, G.A., et al., "ART 2: Self-Organization of Stable Category Recognition Codes for Analog Input Patterns," Proceedings of the First International Conference on Neural Networks, 1987, pp. 2-9.
Kung, S.Y., Digital Neural Networks, PTR Prentice Hall, 1993, p. 149.
Fukunaga, Introduction to Statistical Pattern Recognition, Academic Press, Orlando, Florida, 1972, pp. 213-332.
Duda, R.O. and Hart, P.E., Pattern Classification and Scene Analysis, Wiley, New York, 1973, pp. 213-221.
Keller, J.M., et al., "Membership Function Issues In Fuzzy Pattern Recognition," IEEE 1985 proceedings of the International Conference on Cybernetics and Society, Tucson, Arizona, Nov. 12-15, 1985, pp. 210-214.
Boudreau Eric R.
Menon Murali M.
Couso Jose L.
Massachusetts Institute of Technology
LandOfFree
Pattern recognition system with statistical classification does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Pattern recognition system with statistical classification, we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Pattern recognition system with statistical classification will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-210584