Self organizing neural network method and system for general classification of patterns

• US 5,048,100 A
• Filed: 12/15/1988
• Issued: 09/10/1991
• Est. Priority Date: 12/15/1988
• Status: Expired due to Fees

• Current Assignee: EMC Corporation (Dell Technologies Inc.)
• Original Assignee: Michael Kuperstein
• Inventors: Kuperstein, Michael
• Primary Examiner(s): Moore, David K.
• Assistant Examiner(s): Jung, Yon
• Application Number:
  US07/284,975
• Time in Patent Office: 999 Days
• Field of Search: 382/14, 382/15, 382/33, 382/35, 382/36, 382/37, 382/42, 364/715.11, 364/807
• US Class Current: 382/157
• CPC Class Codes: G06N 3/02 Neural networks
• CPC Class Codes: G06V 30/10 Character recognition
• CPC Class Codes: G06V 30/18029 filtering with Haar-like subimages, e.g. computation thereof with the integral image technique
• CPC Class Codes: G06V 30/18057 Integrating the filters into a hierarchical structure, e.g. convolutional neural networks [CNN]

Abstract

A neural network system and method that can adaptively recognize each of many pattern configurations from a set. The system learns and maintains accurate associations between signal pattern configurations and pattern classes with training from a teaching mechanism. The classifying system consists of a distributed input processor and an adaptive association processor. The input processor decomposes an input pattern into modules of localized contextual elements. These elements in turn are mapped onto pattern classes using a self-organizing associative neural scheme. The associative mapping determines which pattern class best represents the input pattern. The computation is done through gating elements that correspond to the contextual elements. Learning is achieved by modifying the gating elements from a true/false response to the computed probabilities for all classes in the set. The system is a parallel and fault tolerant process. It can easily be extended to accommodate an arbitrary number of patterns at an arbitrary degree of precision. The classifier can be applied to automated recognition and inspection of many different types of signals and patterns.

Claims

1. A method of classifying a pattern into one of a plurality of pattern classes, comprising the steps of:

converting the pattern to be classified into an image comprising at least one spatial distribution of intensity amplitudes relative to a spatial window on the pattern;

convolving the image with a plurality of different convolution kernels;

averaging convolutions of the image across spatial sectors within the spatial window;

transforming the averaged convolutions of the image with a distribution of weights for each of the plurality of pattern classes into computed correlations for each of the plurality of pattern classes; and

determining a single pattern class of the plurality of pattern classes from the computed correlations, whereby the pattern is classified in said single pattern class.

2. The method according to claim 1, wherein the spatial sectors are overlapping.

3. The method according to claim 1, further comprising normalizing averaged convolutions of the image before the step of transforming.

4. The method according to claim 1, wherein the step of convolving the image with the plurality of different convolution kernels comprises using kernels of different configurations in different sizes and orientations relative to the image.

5. The method according to claim 1, further comprising defining the distribution of weights for a given pattern class of the plurality of pattern classes by storing a map of initial neural weights, initially converging a given pattern known to be classified in the given pattern class into an image comprising a given spatial distribution of intensity amplitudes, transforming the given spatial distribution with the map of initial neural weights into computed correlations for each of the plurality of pattern classes, determining errors from the computed correlations and expected correlations and updating the stored map of initial neural weights depending upon the determined errors.

6. The method according to claim 4, further comprising defining he distribution of weights for a given pattern class of the plurality of pattern classes and for a given convolution kernel of the plurality of convolution kernels by storing a map of initial neural weights for the given convolution kernel, initially converting a given pattern known to be classified in the given pattern class with the given convolution kernel into an image comprising a given spatial distribution of intensity amplitudes, transforming the given spatial distribution with the map of initial neural weights into computed correlations for each of the plurality of pattern classes, determining errors from the computed correlations and expected correlations and updating the stored map of initial neural weights depending upon the determined errors.

7. The method according to claim 1, wherein the step of converting comprises converting the pattern to be classified into a binary image.

8. The method according to claim 1, wherein the step of converting comprises converting the pattern to be classified into a two dimensional image.

12. The method according to claim 1, wherein the step of determining comprising selecting the single pattern class with a highest probability from the computed correlations.

9. A method of classifying a pattern into one of a plurality of pattern classes, comprising the steps of:

converting the pattern to be classified into an image comprising at least one spatial distribution of intensity amplitudes relative to a spatial window on the pattern by disposing a center of the spatial window at a horizontal center and at a vertical center of the pattern to be classified and totally enclosing the pattern to be classified in the spatial window;

transforming the image with a distribution of weights for each of the plurality of pattern classes into computed correlations for each of the plurality of pattern classes; and

determining a single pattern class of the plurality of pattern classes from the computed correlations, whereby the pattern is classified in said single pattern class.

10. The method according to claim 9, for use in character recognition, wherein the step of converting the pattern to be classified into the image comprises converting a page into a binary image, processing the binary image to determine a number of lines of characters thereon and isolating each character on each line.

11. The method according to claim 10, wherein the step of disposing the spatial window at the horizontal center and vertical center of the pattern comprises disposing the spatial window at the horizontal center of a given character of a given line and at a vertical center of an average height of all characters in the given line.

13. A system for classifying a pattern into one of a plurality of pattern classes, comprising:

means for converting a pattern to be classified into an image comprising at least one spatial distribution of intensity amplitudes relative to a spatial window on the pattern;

means for convolving the image with a plurality of different convolution kernels;

means for averaging convolutions of the image across spatial sectors within the spatial window;

means for transforming the averaged convolutions of the image with a distribution of weights for each of the plurality of pattern classes into computed correlations for each of the plurality of pattern classes; and

means for determining a single pattern class of the plurality of pattern classes from the computed correlations, whereby the pattern is classified in said single pattern class.

14. The system according to claim 13, wherein the spatial sectors are overlapping.

15. The system according to claim 13, further comprising means for normalizing averaged convolutions of the image before transforming by the means for transforming.

16. The system according to claim 13, wherein the plurality of different convolution kernels comprises kernels of different configurations in different sizes and orientations relative to the image.

17. The system according to claim 13, further comprising means for defining the distribution of weights for a given pattern class of the plurality of pattern classes comprising means for storing a map of initial neural weights, means for initially converting a given pattern known to be classified in the given pattern class into a image comprising a given spatial distribution of intensity amplitudes, means for transforming the given spatial distribution with the map of initial neural weights into computed correlations for each of the plurality of pattern classes, means for determining errors from the computed correlations and expected correlations and means for updating the stored map of initial neural weights depending upon the determined errors.

18. The system according to claim 16, further comprising means for defining the distribution of weights for a given pattern class of the plurality of pattern classes and for a given convolution kernel of the plurality of convolution kernels comprising means for storing a map of initial neural weights for the given convolution kernel, means for initially converting a given pattern known to be classified in the given pattern class with the given convolution kernel into an image comprising a given spatial distribution of, intensity amplitudes, means for transforming the given spatial distribution with the map of initial neural weights into computed correlations for each of the plurality of pattern classes, means for determining errors from the computed correlations and expected correlations and means for updating the stored map of initial neural weights.

19. The system according to claim 13, wherein the means for converting comprises means for converting the pattern to be classified into a binary image.

20. The system according to claim 13, wherein the means for converting comprises means for converting the pattern to be classified into a two-dimensional image.

21. The system according to claim 13, wherein the means for determining comprises means for selecting the single pattern class with a highest probability from the computed correlations.

22. A system for classifying a pattern into one of a plurality of pattern classes, comprising:

means for converting a pattern to be classified into an image comprising at least one spatial distribution of intensity amplitudes relative to a spatial window on the pattern, the means for converting comprising means for disposing a center of the spatial window at a horizontal center and at a vertical center of the pattern to be classified and for totally enclosing the pattern to be classified in the spatial window;

means for transforming the image with a distribution of weights for each of the plurality of pattern classes into computed correlations for each of the plurality of pattern classes; and

means for determining a single pattern class of the plurality of pattern classes from the computed correlations, whereby the pattern is classified in said single pattern class.

23. The system according to claim 22, for use in character recognition, wherein the means for converting the pattern to be classified into the image comprises means for converting a page into a binary image, means for processing the binary image to determine a number of lines of characters and means for isolating each character on each line.

24. The system according to claim 23, wherein the means disposing the spatial window at a horizontal and vertical center of the pattern comprises means disposing the spatial window at the horizontal center of a given character of a given line and at a vertical center of an average height of all characters in the given line.