JPS6211388B2 - - Google Patents

Description

[Detailed description of the invention]

The present invention relates to a character feature discrimination circuit. In printed character reading devices, characters that cannot be read the first time are often read again by changing the viewpoint of feature extraction, and the present invention is a feature discrimination circuit used in this case. . Typical feature extraction methods used in printed character reading devices include a stroke analysis method and a pattern matching method. Stroke analysis is a method commonly used in early character reading devices and requires the use of specially designed special characters. An example of characters used in the stroke analysis method is shown in FIG. These characters are determined based on the presence or absence of horizontal lines at the top, middle, and bottom, vertical lines at the top right, bottom right, top left, and bottom left. The pattern matching method is currently the most widely used method in printed character reading devices, and the JISOCR-B font is a character designed to be the easiest to read when reading characters using the pattern matching method. Figure 2 is
JISOCR-B font shape. pattern·
The matching method is a method of superimposing an input character with all standard characters and determining which standard character matches the input character best, and usually involves correcting positional deviations. Misalignment correction involves moving the input character forward, backward, left, and right on top of the standard character to find the position that best matches it. Using this method, especially clearly printed numbers for OCR can be read without any problems. However, as the number of characters to be read increases, including numbers, letters, and symbols, it becomes difficult to distinguish between similar characters using the pattern matching method. An example using JISOCR-B font is:
(0 zero oh), (D-O oh), (1-?),
(2-Z), (8-B), etc. are examples that are difficult to distinguish. The reason for this is that in the pattern matching method, as shown in Figure 3, when creating standard characters, a fairly wide area _R2 , which can be either one, must be provided between the character line area _R1 and the blank area _R3 . This is because small changes in character outlines cannot be detected. Conventionally, the following method has been known as a method for processing unreadable characters used in a printed character reading device using a pattern matching method. The first method is to prepare a modified standard character for a modified character that cannot be accommodated in one standard character, and reread the modified standard character. The second thing is when creating a quantized figure,
This method involves changing the threshold between white and black, re-creating the quantized figure, and re-reading the same standard characters from then on. The third method is to start scanning again and reread. These methods are very effective when the print quality is poor and the characters to be read are only numbers. However, although the print quality is relatively good, it is not very effective when there are many characters to be read. The reason for this is that the causes of unreadability are different: in the case of only numbers, the characters are in bad shape and do not match any standard characters, making them unreadable in most cases, whereas the characters to be read are If there are many characters, the shape of the characters is relatively good and they match well with the standard characters, but the two standard characters cannot be distinguished and are often unreadable. Therefore, an object of the present invention is to identify the most different part between two standard characters for characters that cannot be read because there are many characters to be read and it is impossible to distinguish between the two standard characters. It is an object of the present invention to provide a character feature discriminating circuit that can effectively distinguish between two characters by simply examining the character in detail. Next, the principle of the present invention will be explained. FIG. 4 shows an example of an input figure (pattern) that can be handled by the apparatus according to the present invention, and is represented by a 24×34 mesh. As shown in FIG. 5, such a figure is first stored in a rectangular storage area in a storage device such that the center of the input figure coincides with the center of the rectangle. The center of the input figure is the intersection of the height bisector and the width bisector. Next, the distances between the outer contour of the figure stored in this manner and the four sides of the rectangle are measured at three locations. Now, the input shape is either D or O, but I don't know which one it is.
If so, the distances between the three locations are as shown in Figure 6a,
As shown in b, this is the length from the left side of the rectangle in the L1, L2, and L3 lines to the figure. As shown in Fig. 6 a, b, this length is a, b,
c, the value P for distinguishing between the two is calculated by the following formula. P=(a-b)+(c-b)=a+c-2b After this, check the value of P, and if it is close to zero, D
(Figure 6a), and if it is larger than a certain value, O
(Oh) (Fig. 6b) is determined. Here, as shown in FIG. 7, the value of P hardly changes even if the width of the character line becomes thicker. In other words, P is
This value is not affected much by the thickness of the character line.
Also, as shown in Figure 8, even if the letters are tilted, they will still be read as L1.
If the distance between L2 and the distance between L2 and L3 are almost the same, then
The value of P remains almost unchanged. In other words, P is in 3 places
If you choose L ₁ , L ₂ , and L ₃ carefully, it will not be affected much by the inclination of the letters. These three locations can be determined in advance depending on what kind of standard characters (combinations) the input characters are indistinguishable from each other. Figures 9 to 13 are D, O, and 0, respectively.
(Zero) and O (Oh), 1? , 2 and Z, and 8 and B, the positions of three locations L ₁ , L ₂ , and L ₃ are shown. In this way, by using the value of P, it becomes possible to detect small changes in the contour line quite accurately. When determining the lengths of a, b, and c, consider the dents in one line that often appear in printed characters (a in the same figure), breaks in one line (b in the same figure), etc., as shown in Figure 14. It is necessary to determine by referring to the upper and lower rows (in the case of columns, the left and right columns) so as not to be affected by the protrusion of one row (c in the same figure) or noise in one row (d in the same figure). Therefore, in the present invention, the position of the outer contour line is comprehensively determined based on three consecutive lines.
In other words, the location where the pattern exists in two of the three rows is defined as the position of the outer contour line. Next, the present invention will be explained in detail with reference to the drawings showing one embodiment of the present invention. FIG. 15 is a diagram showing an embodiment of the present invention,
In the figure, a predetermined address is set in address register 1 according to a set of characters to be distinguished, and the contents of memory 2 are read out. Memory 2 has a predetermined number of lines depending on the set of characters to be distinguished (for example, D and O).
In the case of the (O) set, rows 9, 17, and 26) are stored, and this data is output to the address register 3 in order. The memory 2 also stores a threshold value for comparing the value of P used in feature determination, and this data is output to registers 5 and 6. Based on the address stored in the address register 3, specific rows (three consecutive rows in this embodiment) of the pattern stored in the memory 4 are read out in correspondence with the designated number of rows. That is, if 9 lines are specified, a pattern of 8 lines, 9 lines, and 10 lines (this is shown in Figure 4, although the target characters are different) is read out, and the shift registers 9, 10, and 11 are read out one by one. is stored in It is assumed that the pattern stored in the memory 4 has been normalized so that the center of the pattern is located at the center of the rectangular storage area. Three consecutive lines are shift registers 9, 10, 11
, then while the shift signal is present,
It is shifted in synchronization with clock pulse _CL4 , and the information of the edge bits in shift registers 9, 10, and 11 (information as to whether that point in the pattern is white or black) is sent to flip-flops 12, 13, and 14, respectively. I'm going to go. flipflop 1
2, 13, 14 are shift registers 9, 10, 11
This is set when the information from the line is "1", that is, "black", and the position where it is set corresponds to the position of the outer contour of each row. In this embodiment, in preparation for the case shown in FIG. 10, the position of the outer contour is determined when two of the flip-flops 12, 13, and 14 are set. Therefore, the outputs of the flip-flops 12, 13, and 14 are sent to a 2-bit full adder 15, which outputs an end signal through an OR gate 17 when a carry occurs, indicating that the outer contour has been detected. A counter 16 counts shift pulses CL ₄ , e.g.
When it becomes 1 (one line is 34 meshes), the carry output is sent to the OR gate 17. Further, the carry output from the counter 16 is outputted to the outside as a read-unable signal, indicating that this pattern cannot be read. This signal is F・F (not shown)
etc. When the first shift is completed, the address indicating the next row (for example, row 17) from memory 2 is transferred to register 3.
The next three rows (or columns) from memory 4 are set to
is read out and stored in the shift registers 9, 10, and 11, and the outer contour is detected again. and then 3
The outer contour of the second time (for example, the 26th line) is detected. Next, the counting of a, b, c necessary to calculate P and the calculation of P will be explained. Clock pulses CL ₁ and CL ₃ having the same frequency as clock pulse CL ₄ and different phases are separately connected to gates 21 and 22.
are joining. That is, in order to calculate P=a+c-2b, only the clock pulse _CL3 is counted when counting a and c, and both clock pulses _CL1 and _CL3 are counted when counting 2b. Here, the switching of the counts a, c, and 2b is controlled by the counter 19. The values of a, c, and 2b are sent to the counter 20 via the gate 23. Here, due to the structure of gates 21 and 22, gate 22 is supplied with an inverted clock pulse ₃ . The calculation of P is performed by the counter 20,
In order to calculate P, it is necessary to control addition and subtraction, and the counter 19 is also provided for this purpose. The counter 19 counts the end signal from the OR gate 17, and the least significant bit of the output is used as a control signal. That is, when counting a, there is no end signal, so it is "0", when counting b, there is one end signal, so it is "1", and when counting c, it is "2". It is "0" because it contains end signals. This control signal is the other input of the gate 21, and when it is "0", the clock pulse _CL1 is not counted. This control signal also controls up and down of the counter 20, and when it is "0" it is up and when it is "1" it is down. The other input of gate 23 is supplied with a shift signal from flip-flop 18, and counting is performed only when this shift signal is present.
That is, flip-flop 18 is set by ANDing the start signal and clock pulse _CL5 , and is reset by the end signal from OR gate 17. The value of P counted by the counter 20 may be negative, so to prevent complications, a unique value V is given to the counter 20 from the beginning so that the count value PV output from the counter 20 does not become negative. That's what I do. Counting amount PV from counter 20
are supplied to comparators 7 and 8, and are compared with threshold values stored in registers 5 and 6 to obtain feature discrimination signals X and Y.
Output each. Next, an example of discrimination conditions will be shown. In the combination of 0 (zero) and O (oh), P
0 (zero) when ≦5 (PV≦13), P≧7 (PV
≧15), it is O (O). In the combination of D and O, when P≦2 (PV≦10), D, P≧3
When (PV≧11), it is O (O). 1? In combination with 1 when P≦2 (P≦10), P≧4
When (PV≧12)? It is. In the combination of 2 and Z, when P≦6 (PV≦14), Z, P≧8 (PV≧
16), it is 2. In the combination of 8 and B, P
8 when ≦-4 (PV≦4), P≧-2 (PV≧
6), it is B. The combination of characters (categories to be distinguished), the contents of registers 5 and 6, and the
The following table shows the determination relationship when Y=“1” and Y=“1”.

[Table] Clock pulses CL ₁ to _{CL 5} have the same frequency and different phases, and clock pulse CL ₂ is not used. This is to provide sufficient time between clock pulse CL ₁ and clock pulse CL ₃ in order to stably count P. The counter 19 is cleared by the non-initial signal, and the counter 20 is loaded with the value V by the initial signal. The timings of these initial signals, start signals, shift signals, and clock pulses CL ₁ to CL ₅ are shown in FIG. In the figure, the period
T ₁ , T ₂ , T ₃ are pattern shift registers 9, 1
The period is set to 0, 11, and the period t ₁ ,
t ₂ and t ₃ are shift registers 9, 10, 1
This is the period that is shifted within 1. As is clear from the above explanation, according to the present invention, small differences between two characters can be accurately detected using a simple and inexpensive circuit, and characters that cannot be distinguished using conventional pattern matching methods can be effectively detected. It is possible to distinguish between them.

[Brief explanation of the drawing]

Figure 1 shows an example of characters that can be identified using the stroke analysis method, Figure 2 shows an example of characters in JISOCR-B font that can be identified using the pattern matching method, and Figure 3 shows pattern recognition using the pattern matching method. Figure 4 is a diagram showing an example of the input figure; Figure 5 is a diagram explaining how to match the center of the memory with the center of the pattern; A diagram for explaining the principle of the invention. FIG. 7 is a diagram for explaining the principle of the invention, and shows a case where the pattern is thick. 8th
The figure is a diagram for explaining the principle of the present invention and shows a case where the pattern is tilted. Figures 9 to 13 are diagrams showing patterns for various character combinations and outer contour measurement positions, Figure 14 a, b, c, and d are dents a, cuts b, protrusions c, and noise d in the patterns.
FIG. 15 is a diagram showing an embodiment of the present invention, and FIG. 16 is a diagram showing the relationship between various timing signals in FIG. 15.

Claims (1)

[Claims] 1 Make sure the center of the character shape is in the fixed position,
a memory for storing a quantized figure; a measuring means for measuring the positions of at least three outer contour lines predetermined for each type of figure in the figure stored in the memory; Calculating means for calculating the relative positional relationship of the other two positions with respect to the middle position among the three positions, and comparing means for comparing the value obtained by the calculating means with a value given in advance and outputting a discrimination output. A character feature discrimination circuit comprising: