Receptive field neural network with shift-invariant pattern recognition

• US 5,263,107 A
• Filed: 01/31/1991
• Issued: 11/16/1993
• Est. Priority Date: 01/31/1991
• Status: Expired due to Fees

• Current Assignee: Sharp Kabushiki Kaisha (Hon Hai Precision Industry Co., Ltd.)
• Original Assignee: Sharp Kabushiki Kaisha (Hon Hai Precision Industry Co., Ltd.)
• Inventors: Ueda, Toru, Togawa, Fumio
• Primary Examiner(s): Boudreau, Leo H.
• Assistant Examiner(s): KELLEY, CHRISTOPHER S
• Application Number:
  US07/647,823
• Time in Patent Office: 1,020 Days
• Field of Search: 382/19, 382/15, 382/34, 395/21, 395/22, 395/23
• US Class Current: 382/157
• CPC Class Codes: G06F 18/24137 Distances to cluster centroïds
• CPC Class Codes: G06V 30/10 Character recognition
• CPC Class Codes: G06V 30/18057 Integrating the filters into a hierarchical structure, e.g. convolutional neural networks [CNN]
• CPC Class Codes: G06V 30/287 of Kanji, Hiragana or Katakana characters

Abstract

A neural network system and method of operating same wherein input data are initialized, then mapped onto a predetermined array for learning or recognition. The mapped information is divided into sub-input data or receptive fields, which are used for comparison of the input information with prelearned information having similar features, thereby allowing for correct classification of the input information. The receptive fields are shifted before the classification process, in order to generate a closest match between features which may be shifted at the time of input, and weights of the input information are updated based upon the closest-matching shifted position of the input information.

Claims

1. A method for updating weights relating to previously learned information input into a neural network, the network having a database of learned information mapped onto an array, including the steps of:

(1) mapping the input information onto the array;

(2) isolating a predetermined subset of the array;

(3) shifting the predetermined subset in at least one dimension;

(4) comparing the predetermined subset in both unshifted and shifted positions with a corresponding subset of the learned information; and

(5) based upon the comparison of step (4), determining an optimal shift for the input information; and

(6) updating a weight relating to the learned information based upon the optimal shift.

2. The method of claim 1, wherein step 3 includes shifting the input information in two dimensions.

3. The method of claim 1, wherein step 3 includes rotating the input information.

4. The method of claim 1, wherein the array comprises a matrix of pixels having a first dimension and a second dimension, and wherein step 3 includes the steps of:

(7) shifting the input information a predetermined first number of pixels in each of a negative direction and a positive direction in the first dimension without shifting in the second dimension;

(8) shifting the input information a predetermined second number of pixels in each of a negative direction and a positive direction in the second dimension without shifting in the first dimension; and

(9) shifting the input information the predetermined first number of pixels in each of the negative direction and the positive direction in the first dimension while shifting the input information the predetermined second number of pixels in each of the negative direction and the positive direction in the second dimension.

5. A method for learning information input in a neural network, where the input information corresponds to a feature stored in the neural network, including the steps of:

(1) mapping the information onto at least one receptive field in an unshifted position;

(2) shifting the mapped information into a plurality of shifted positions;

(3) determining which of the plurality of shifted positions matches the stored feature most closely; and

(4) based upon closest matching shifted position, updating a weight relating to the stored feature.

6. An apparatus for updating weights relating to information which has been learned in a neural network, the network including a database of leaned information mapped onto an array, comprising:

means for mapping the input information onto the array;

means coupled to the mapping means for isolating a plurality of predetermined subarrays of the mapped input information;

means in communication with the isolating means for shifting the isolated subarrays of information in at least one dimension;

means for storing the shifted subarrays;

means in communication with the storing means for determining a closest match between each of the respective shifted subarrays and learned information corresponding thereto; and

means in communication with the determining means for updating weights for the learned information based upon the determined closest matches.

7. The apparatus of claim 6, wherein the shifted, isolated subarrays include subarrays which are shifted in each of two dimensions.

8. The apparatus of claim 6, wherein the shifted, isolated subarrays include rotated subarrays which are shifted about an axis.

9. The apparatus of claim 6, wherein the array comprises a matrix of pixels having a first dimension and a second dimension, and wherein the shifted, isolated subarrays include:

first shifted subarrays which are shifted a predetermined first number of pixels in each of a negative direction and a positive direction in the first dimension without shifting in the second dimension;

second shifted subarrays which are shifted a predetermined second number of pixels in each of a negative direction and a positive direction in the second dimension without shifting in the first dimension; and

third shifted subarrays which are shifted the predetermined first number of pixels in each of the negative direction and the positive direction in the first dimension, and are also shifted the predetermined second number of pixels in each of the negative direction and the positive direction in the second dimension.

10. An apparatus for learning input patterns in a neural network, wherein the patterns belong to classes of patterns, comprising:

means for mapping at least one portion of each input pattern onto an input image area which is larger than the mapped portion;

means in communication with the mapping means for shifting the portion of the pattern within the input image area to each one of a plurality of predetermined positions for comparison with prelearned portions of other patterns; and

means in communication with the shifting means for generating weight vector outputs from the neural network based upon a best match between the shifted portion of the pattern at one of the predetermined positions and one of the prelearned portions of the other patterns.

11. The apparatus of claim 10, wherein the weight vector generating means includes means for changing a weight vector with the largest-valued output relating to a class other than a class to which the input pattern belongs farther away from input vector in the best-match position.

12. The apparatus of claim 10, wherein the weight vector generating means includes means for correlating unknown input patterns with prelearned patterns.

13. A method for learning distinguishing characteristics of patterns input into a neural network, including the steps of:

(1) mapping an input pattern onto an array, the array comprising a predetermined number of tiles;

(2) generating values for weights relating to each of the tiles;

(3) repeating steps 1 and 2 for each of a plurality of patterns;

(4) generating contributive values from the weight values based upon a correlation of the weight values for each of the tiles with weight values of corresponding tiles of other patterns;

(5) correlating the contributive values for determining which of the contributive values relate to regions of the array wherein patterns of the class differ most from one another;

(6) generating a set of sums of contributive values from sets of tiles;

(7) determining at least one subset of the set of sums which is higher than other sums;

(8) mapping the determined subset onto the array for generating at least one receptive field for comparison among different ones of said patterns, the receptive field relating to a region representing features of the patterns which are distinct from one another;

(9) defining an input image area which is larger than the generated receptive field;

(10) shifting data of each input pattern within the input image area for determining a best correspondence of the shifted data and a previously learned weight relating to the receptive field; and

(11) updating the weights based upon the shifted data.