Reference data recognition and learning method and pattern recognition system

1. A reference data recognition and learning method of reference data recognition and learning with respect to an associative memory, which compares input search data with several reference data to select and output the most similar reference data, comprising:

giving a rank corresponding to a priority to each of the several reference data stored in the associative memory;

comparing the search data with the several reference data when the search data is input to detect a similarity between the search data and each of the several reference data, and selecting reference data having the highest similarity; and

determining whether or not the search data and the reference data selected by the comparing are identical to each other from their similarity, wherein

the giving the rank includes controlling the rank such that a rank of the reference data selected by the comparing is improved by a predetermined number while a rank of other reference data equal to or lower than the improved rank is reduced when the search data and the reference data selected by the comparing are determined to be identical in the determining, and

when the rank is controlled, a long-term memory hierarchy and a short-term memory hierarchy, each including an arbitrary number of ranks are formed, and when the search data and the reference data selected by the comparing are determined not to be identical in the determining, the search data are registered to an uppermost rank of the short-term memory hierarchy as new reference data, and when the search data are determined to be identical to the reference data in the short-term memory hierarchy, the registration of the search data is shifted to a predetermined rank of the long-term memory hierarchy and thereafter the rank of the reference data in the long-term memory hierarchy determined to be identical to the search data in the determining is increased by the predetermined number.

2. The method according to claim 1 , wherein in the determining, the similarity between the search data and the reference data selected by the comparing is expressed by a distance, and a minimum distance between the search data and the reference data selected by the comparing is calculated, and the determining determines that the search data and the most similar reference data are identical if the minimum distance is within a predetermined threshold and otherwise determines that they are not identical.

3. The method according to claim 1 , wherein in the controlling of the rank, reference data of the lowest rank are abandoned when registering the search data as the new reference data.

4. The method according to claim 1 , wherein in the controlling of the rank, when the search data and the reference data selected by the comparing are determined to be identical, an amount of improvement when improving a rank in a same memory hierarchy is larger in the long-term memory hierarchy than in the short-term memory hierarchy.

5. A pattern recognition system for selecting reference data, which is the most similar to search data of an input pattern, from several reference data to recognize the input pattern, comprising:

an associative memory including

reference data memory means for storing the several reference data for recognizing the search data of the input pattern;

minimum distance search means for expressing a degree of similarity by distance and searching for the reference data having the minimum distance with respect to the search data from said several reference data; and

identity determining means for determining the identity of the search data and the reference data having the minimum distance in accordance with whether or not the minimum distance is more than a threshold value;

rank control means for managing the several reference data stored in the associative memory to give a rank corresponding to the priority to each of the several reference data; and

recognition learning control means for controlling the rank control means so that the rank control means improves a rank of the reference data while reducing a rank of other reference data equal to or lower than the improved rank when the identity determining means determines that the search data and the reference data are identical, wherein

the recognition learning control means forms a long-term memory hierarchy and a short-term memory hierarchy, each having an arbitrary number of ranks,

when the search data and the reference data are determined not to be identical by the identity determining means, the recognition learning control means registers the search data to an uppermost rank of the short-term memory hierarchy as new reference data,

when the search data are determined to be identical to the reference data in the short-term memory hierarchy, the recognition learning control means shifts the registration of the search data to a predetermined rank of the long-term memory hierarchy and improves, thereafter, the rank of the reference data in the long-term memory hierarchy determined to be identical to the search data by the identity determining means by a predetermined number.

6. The system according to claim 5 , wherein the minimum distance search means searches for either Hamming distance, Manhattan distance or Euclidian distance as the distance between the search data and the reference data.

7. The system according to claim 5 , wherein the rank control means manages the rank of the reference data stored in the associative memory based on their address value.

8. The system according to claim 5 , wherein the rank control means carries out rank improvement between the memory hierarchies only when the search data is determined to be identical to the reference data by a predetermined number of times in the identity determining means.

9. The system according to claim 5 , wherein the associative memory, the rank control means and the recognition learning control means are built in an integrated circuit.

10. A pattern recognition system for selecting reference data, which is the most similar to search data of an input pattern, from several reference data to recognize the input pattern, comprising:

an associative memory including

a reference data memory configured to store the several reference data for recognizing the search data of the input pattern;

a minimum distance search circuit configured to express a degree of similarity by distance and to search for the reference data having the minimum distance with respect to the search data from said several reference data; and

an identity determining circuit configured to determine the identity of the search data and the reference data having the minimum distance in accordance with whether or not the minimum distance is more than a threshold value;

a rank control circuit configured to manage the several reference data stored in the associative memory to give a rank corresponding to the priority to each of the several reference data; and

a recognition learning control circuit configured to control the rank control circuit so that the rank control circuit improves a rank of the reference data while reducing a rank of other reference data equal to or lower than the improved rank when the identity determining circuit determines that the search data and the reference data are identical, wherein

the recognition learning control circuit is configured to form a long-term memory hierarchy and a short-term memory hierarchy, each having an arbitrary number of ranks,

when the search data and the reference data are determined not to be identical by the identity determining circuit, the recognition learning control circuit is configured to register the search data to an uppermost rank of the short-term memory hierarchy as new reference data,

when the search data are determined to be identical to the reference data in the short-term memory hierarchy, the recognition learning control circuit is configured to shift the registration of the search data to a predetermined rank of the long-term memory hierarchy and to improve, thereafter, the rank of the reference data in the long-term memory hierarchy determined to be identical to the search data by the identity determining circuit by a predetermined number.