GB1300224A - Improvements in or relating to pattern recognition systems - Google Patents

Abstract

1300224 Pattern recognition NATIONAL RESEARCH DEVELOPMENT CORP 16 March 1970 [18 March 1969] 14240/69 Addition to 1237534 Heading G4R A pattern recognition arrangement using log-autocorrelation or autocorrelation-log-autocorrelation transformations as in the Specification 1,237,534, has comparative recognition means including filtering means to accentuate the regions of difference between otherwise similar patterns or provide each element of the field with a gain proportional to ratio of signal content to square of noise content. The recognition means filters the matrix of signals from the (last) autocorrelation by weighting each element of the matrix with the corresponding element of a weighting matrix. In fact a chain of filters is used, classifying the pattern in progressively smaller groups of possible patterns until the pattern is classified into one of a plurality of one-pattern groups to identify the pattern. A typical weighting matrix, for distinguishing between P and F, is obtained as follows. For each character separately, a series of versions of the character is applied and the output matrix from the (last) autocorrelation for each is subtracted from the mean of these matrices, to form a difference matrix. Each element of each difference matrix is squared, and corressponding elements in the difference matrices are averaged to obtain a mean square noise matrix. A signal matrix is obtained by subtracting the mean values of all the P matrices from the mean values of all the F matrices (or vice versa). The signal matrix elements divided by the corresponding mean square noise matrix elements give the elements of the weighting matrix. In the recognition, the elements resulting from the weighting are added together to give a number which identifies the character as is P or F (the filter for distinguishing P from F is of course at the exit end of the chain of filters). The (last) autocorrelation output is normalized by dividing by the value at the origin of the average value. Further details may be as in the Parent.