GB1234066A - - Google Patents

Abstract

1,234,066. Pattern recognition. SPETSIALNOE KONSTRUKTORSKOE BJURO VYCHESLITELNYKH MASHIN. 10 July, 1968, No. 32902/68. Heading G4R. An object is recognized from its contour representation by coding and automatically fixing contour representation scanning lines at places where a stable changeover occurs from lines making one number of intersections with the contour to lines making another number of intersections, the sections of the contour between the fixed lines being coded and used to derive closed-open contour characteristics which are used to recognize the object. Character signals at 117 (Fig. 16, bottom) from a vertical raster scan are entered into a shift register 87 which supplies signals N1-N8 to a circuit 32 (Fig. 15, bottom) wherein intersections of the scan lines with the character outline are recognized e.g. by OR 58 and AND 59 and gaps between intersections are recognized e.g. by AND 53 (isolated blacks and whites being ignored), the number of intersections per scan line being counted at 62 to produce signals n1-n8 from a unit 63 indicating the type of the scan line according to the number of intersections, sometimes in accordance with the numbers in previous lines, e.g. n7 means the line is the second successive line with two intersections found after lines with any other number of intersections or on character beginning. The signals n1-n8 control logic and flip-flops at 33 to indicate line-type sequences which are decoded at 34 to obtain commands K1-K8 which are further decoded at 35 to obtain further commands Q1-Q5. During each scan line, clock pulses V1, V2, V3, S1, S2 ... S28, W1, W2, W3, W4, W5 are produced in turn. Returning to Fig. 16, W4 gates the contents of register 87 to a register 89 and W2 ORs the contents of 89 into a shift register 91 used for detecting continuity across scan lines. Register 91 is cleared by command Q2. A shift register 93 used for storing a line code at critical positions across the character, e.g. at a stable changeover from one number of intersections per scan line to another, is cleared by command Q 1 and loaded from register 89 bycommand Q3. Commands Q4, Q5 pass the contents of registers 91, 93 respectively via lines 147, 150, 148, 149 to be gated by K commands (Fig. 17, left) to counters 152, 154, 157 which count whites, blacks and whites respectively. The lower input to each counter is a reset input. Returning to Fig. 16, register 93 also supplies data to a circuit 103 which in the presence of one of the commands K2, K5, K6 determines the number of intervals between intersections using a counter 116. Counter 116 provides signals P1, P2 which together with K commands and the outputs of counter 152, 154, 157 are used (Fig. 17) to detect whether the contour is open between appropriate places (the critical positions mentioned and character begin and end), the information being used by a classifier 31 (Fig. 17, far right) to identify the character by comparison with stored information. Roughly speaking, the upper and lower halves of the detailed circuitry of Fig. 17 relate to characters having a maximum number of intersections per scan line of three and two respectively. The time intervals between successive intersections, and between successive K commands, may also be used for recognition in the classifier 31. The decoder 34 also feeds the classifier to permit a preliminary classification according to the number of intersections. The classifier store may be loaded during a learning process. An uncertain decision adjusts the criteria used for detecting intersections and gaps between intersections in circuit 32 (Fig. 15) and for detecting closed-open contour characteristics at 38 (right four fifths of Fig. 17 excluding classifier 31) by feedback from the classifier 31. Digital processing in Fig. 16 may be replaced by integration followed by A-D conversion if the character input signals are non-binary e.g. analogue. Scanning of the character may be done horizontally as well.