JPS5521384B2 - - Google Patents

Abstract

1454148 Character recognition INTERNATIONAL BUSINESS MACHINES CORP 16 Dec 1974 [10 April 1974] 54200/74 Heading G4R A character recognition machine contains a store holding dictionary words flagged to indicate a propensity to be misread either by HSS where a character in the original is split to form two characters in the recognized word, CS where two characters in the original are joined to form one character in the recognized word or by CRS where two characters in the original are crowded together and recognized as two different characters because of a misread of the segmentation point. The recognized word is fed to a word Separation Detector 4. Alphabetic character fields are fed to an OCR Word Shift Register 14 and the number of characters fed to the register is counted by Counter 18 and the value is fed to Shift Control 20. The end character of the word is stored in cell K1 of register 14. A dictionary store 28 contains words which may be read, e.g. the contents of a dictionary or a directory of street names, &c. depending on the application, and feeds words to a shift Register 26 with the end character of the word in cell L1. The flag bits may be in register 26 or in a separate register 34. A Flag Decode 100 examines the flag of the character stored in cell L1 and has four output lines 102 indicating a probable simple substitution, a probable character splitting (HSS), a probable character pair concatenation (CS) or a probable character pair crowding (CRS). The shift registers 14, 26 feed respective multiplex units 94, 96 which via respective address registers 116, 122 obtain probability values from a Conditional Probability Storage Matrix 124. The matrix contains conditional probabilities P(Kn/Lm) that the character in cell Kn was really the character in cell Lm for values N=1, m=1; n=2, m=2; n=2, m=3; n=3, m=2; P(K1K2/L1) that characters K1K2 were split versions of L1, P(K1/L1L2) that character K1 was a combination of L1L2 and P(K1K2/L1L2) that characters K1K2 were misread by crowding of characters L1L2. Timing of the operations of multiplex units 94, 96 and of a multiplex unit 128 receiving the probability values from the matrix are controlled by timer 108 receiving the Flag Decode signals. The probabilities obtained from the matrix are supplied to registers 130, 132, 134, 136 and pairs of values are multiplied and compared to determine which L character has the most probability of forming the OCR K character and to determine if relative shifting of the characters in registers 14, 26 should occur. Example.-OCR word IWn*C where * indicates rejected or unrecognized character. This could be compared with Break, Wreck or Freak. Break carries a flag bit indicating that r and e can produce crowding errors. The probabili. ties that K was read as C and that A was not recognized gives a running total of 7À3Î10<SP>-5</SP>. For crowding and the best running total is 2 x 10<SP>-9</SP> P(I/B)=2À0Î10<SP>-4</SP> and the running total is 4 Î 10<SP>-13</SP>. Wreck carries three flag bits indicating probability of crowding of r and e, joining of c and k to form one character and splitting of W to form two characters. The probabilities are running total 3 x 10<SP>-7</SP> running total = 1À1 Î 10<SP>-9</SP> for the original word to be Wreck. The running total for Freak is 1À5 x 10<SP>-12</SP>.