GB1006418A - Adaptive recognition system - Google Patents

Abstract

1,006,418. Automatic character reading. INTERNATIONAL BUSINESS MACHINES CORPORATION. April 16, 1964 [April 29, 1963], No. 15725/64. Heading G4R. In a specimen recognition system of the kind in which a function of the specimen is compared with a plurality of functions corresponding to reference patterns to obtain an indication of the reference function most similar to the specimen, there are means which, during a " training " phase of operation, serve to alter the reference function under the control of the comparing means. In the operating mode the signals derived from the specimen, e.g. by scanning with a flying spot scanner are applied to a circuit 11, Fig. 1, which produces an auto-correlation function. This function is then compared with the reference functions which were previously generated in a " training " mode and are stored in stores 13. Comparison is by multiplying the specimen auto-correlation function, consisting of a series of binary elements, with the corresponding elements of the reference functions. The multipliers 15, one for each reference pattern, send a series of watch signals to accumulators 17 where they are summed. The largest sum is indicative of the identity of the specimen. The accumulator outputs are applied through AND gates 21 enabled by a switch 23 in the " operate " position and a maximum signal indicator produces an output on a corresponding lead. In the "training" mode sample patterns are scanned to provide input signals for the auto-correlation function generator. Switch 23 is set in the " train " position so that AND gates 29 are enabled to pass the outputs of accumulators 17 to test circuits 31. Gates 21 are disabled. The identity of the sample is set on switch 27 and this selects a corresponding one of the test circuits 31. This then compares the input from the associated accumulator and compares it with a first threshold level, producing an output if this level is not exceeded, that is, if the sample pattern has not been properly identified with sufficient positive discrimination. This signal is then used to control the training of the system causing the data in the corresponding store 13 to be altered so as to improve the discrimination when the sample is next applied. The other test circuits compare the corresponding accumulator outputs with a second threshold level, producing a signal if the level is exceeded. This indicates that the system has insufficient negative discrimination of the sample and the corresponding stores 13 are altered so as to improve the rejection of the sample when next applied. The auto-correlation function from generator 11 is stored in store 25. The contents of this store are added to the patterns in the appropriate one of the stores 13 to improve the positive discrimination (recognition) and subtracted to improve the negative discrimination (rejection). Adding is effected by gates 35 and subtracting is effected by complementor 33 and gates 37. Gates 41 controlled by the test circuits apply the additions or subtractions to appropriate stores 13. The sequence of sample patterns is repeated, the system " converging " until substantially no changes in the stored patterns are made by successive applications of the samples. The auto-correlation functions are formed in generator 11 by storing the train of binary bits obtained from the character in two shift registers. The contents of one register are shifted one step, two steps, three steps and so on with respect to the contents of the other and after each shift the contents of both are compared. Matching black points are counted so that a table of values as shown in Fig. 10 is obtained representing the number of black position overlaps in successive shifted positions as illustrated in Figs. 2, 3, 4 respectively for shifts of zero and 1 and 2 positions to the right.