JPS5971585A - Pattern recognizing system - Google Patents

Abstract

PURPOSE:To improve the accuracy of pattern recognition and the recognizing processing efficiency, by combining a phase structure method used for the length of stroke and direction information as recognizing information to the pattern matching method. CONSTITUTION:A length and 8-direction code producing section 85 calculates the length of stroke taking the corresponding picture element as a center at each element of an input pattern so as to generate the length and 8-direction code table T corresponding an 8-direction code. A distortion calculating section 87 executes the calculation of an evaluation function based on tables I , II in a memory 86, black picture element pattern corresponding table S and a table T. The picture element weighting the input pattern is obtained from the black element weighting definition pattern so as to calculate the stroke information distance and the distortion from the evaluating function. A minimum value selecting section 9 detects the definition pattern representing the minimum value from the distortion at each definition pattern and outputs it as the result of recognition.

Description

[Detailed Description of the Invention] [Technical Field of the Invention] The present invention relates to handwritten characters using a pattern matching method,
Regarding the system for recognizing patterns such as symbols or figures, in particular, the present inventor's previous invention, ``Character/Symbol Recognition System'' (Japanese Patent Application Laid-Open No. Regarding the improvement of 1982-16271), information on the direction and length of the stroke in the pattern is combined with the amount of distortion between the definition/ya turn and the input/ya turn calculated using plain pattern massi. This is intended to improve recognition accuracy.

[Technology background]

When handwritten character recognition methods are conceptually classified, there are two methods: one is based on the topological structure of the character (for example, how strokes are combined), and the other is based on matching with a definition pattern (dictionary, pattern). The disadvantage is that extraction is troublesome, and the latter is difficult to process when characters are deformed. The prior invention, which is the basis of the present invention, belongs to the latter category, but provides a unified method in which neither the deformation caused by handwriting nor the thickness and thinness of lines is a constraint. Below, the outline of the prior invention will be explained.

Definition: As a measure of the degree of matching between the input turn and the input cover turn, we define the distortion measure of the input turn.
This is calculated every seven definition turns, and the definition pattern with the minimum distortion amount for the input cover turn is taken as the recognition pattern. This means that when each definition pattern is distorted (deformed) to match the input cover pattern, the definition pattern that requires the least amount of distortion is selected, that is, the definition pattern that requires the least amount of distortion. means.

In general, a recognition problem is defined as ``determining which predetermined pattern the observed pattern is most similar to.'' An example of such a conventional recognition system is shown in FIG.

In FIG. 1, 11 1 , 11-2 , ... 11
−le is the standard definition character/symbol ・gutān,
For example, "A", "B#", ... "+", etc. are shown. These definition character/symbol patterns are stored in advance in a definition library consisting of, for example, a magnetic disk device. 12 is an input This pattern of characters and symbols to be recognized is detected by an optical reading means such as a scanner.This pattern of characters and symbols to be recognized 12
is the similarity determination unit 13-1.

13-2. ..., 13-n, the corresponding definition patterns 1.1-1, 11-2, ...
, 11-ru to calculate the similarity, and the similarity amount R
,,R2,...RrL are output. The determination unit 14
The maximum amount of similarity among these amounts in order of A or higher is determined, and the pattern to be recognized is read using the defined pattern corresponding to this maximum amount of similarity. This is a typical pattern matching method, and it is difficult to deal with the distortion of handwritten characters and symbols, resulting in the problem that highly accurate character and symbol recognition cannot be achieved. This method effectively solves the problem of the conventional method in which accurate reading cannot be performed due to the distortion that appears in handwritten characters and symbols, and the recognition problem is replaced as follows.

``When pattern However, the essential difference between the two is that, in the latter case, although it is possible to create any pattern by distortion, in the latter case, the question is "Is the observed pattern defined or distorted?" This is based on the assumption that each turn is distorted.

Specifically, for multiple defined characters/symbols that serve as standards, they are defined in advance according to a certain regularity so that the weighting value, that is, the degree of distortion, increases in proportion to each pixel of the pattern. Generate a distorted pattern Mazda of character/symbol patterns.

On the other hand, for the 79 turns of the input characters/symbols to be recognized, distortion patterns/symbols are similarly created according to the above-mentioned regularity.
Generate a map. Then, using the difference between the weighting values of mutually corresponding parts of the distortion pattern map of the definition/mother turn and the recognized pattern as a variable, the value of both F functions is calculated, and the value of the F function is calculated. With the minimum value Z) definition/turn letter/symbol,
This is to confirm recognition.

FIG. 2 is a conceptual diagram of the system according to the prior invention. In this figure, 21-1.21-2. ...21-le is a standard definition character/symbol pattern. These patterns are in the form of distortion pattern maps generated from defining characters and symbols, and are similar to No and Turn 11-1 in Figure 1.
, 11-2, ...11-ru.

On the other hand, the pattern of 22 inputted recognized p characters/symbols is also in the distortion pattern Matsuno format,
This is completely different from pattern 12 in FIG. These R turns 21-1. ...21-n and 22 are patterns 11-1 and -11-yL in FIG.

1.2 is that the prior invention method introduces the concept of "distortion amount" as a recognition parameter. This recognized character/symbol pattern 22Vi, distortion amount calculation units 23-1, 23-2, . . . -23-n
, the corresponding nosoturn 21-1.21
-2. . . 21-n are calculated, and distortions 111T D+ , D2, . . . Dn are output. Then, the determination unit 24 determines the minimum distortion,
The recognition pattern or turn is read using the defined pattern corresponding to this minimum distortion l:.

FIG. 3 is a block diagram of an embodiment of the prior invention.

In this figure, 31Viz? 32 is a pattern normalization section, 33 is a plain pattern generation section, 34 is a distortion amount calculation section, 35 is a definition pattern storage section, and 36 is a minimum value selection section.

The pattern input section 31 optically reads a pattern of characters, symbols, figures, etc. to be recognized, for example, °A'' with a scanner, converts it into an electrical signal, and outputs it as mesh-like pixel information to the pattern standardization circuit 32. do.

32 f-turn standardization circuits convert the input pattern into
The map is standardized to a certain size for recognition processing and output to the plane pattern map generation section 33. Note that the line width of the pattern is not changed in this standardization process.

The plane pattern map generation unit 33 generates a plane (Bτai
n) Generate a map of distortion patterns called patterns.

A plain cut turn map is shown in FIG. 4 (α), for example.

As shown in (b) 1. For each pixel, pixels in the four directions of (α) and the eight directions of (A) are sequentially and alternately selected and ordinal numbers are weighted to generate the image as illustrated in (C) of the same figure. . Note that the above mathematical rule is called "-〇ctagonal f) iztance".

5 (α) and (A), which will be described later, show examples of plain pattern maps generated for the character "A'. (α) is the definition pattern, and (b) is the input pattern map.

In this plain pattern map, in the case of a pattern of black characters on a white background, a weight of "1" is set for all black pixels at the black-and-white boundary, and weights of 2, 3, 4, etc. are set in the direction away from the black pixels.
The weight of the ordinal number is set to increase as follows. In addition, in the plain tsukturn pine f example in Figure 5, the letters "
The line width of "A" is one pixel width, but in the case of thick lines with multiple pixel widths, O,
-1゜-2,..., etc. 0 By using such a brush/in-pattern map, it is possible to eliminate the distortion from the defined character for any corrupted handwritten character. Distance can be found in Dutch studies.

For example, when a black pixel with a weight of "1" is distorted by handwriting and written at a pixel position with a weight of "5", the distortion distance between them is calculated as ',5-1=4.

The distortion amount calculation unit 34 in FIG.
A plane pattern map f (hereinafter referred to as a definition plane pattern map) of a plurality of definition patterns created in advance and stored in the plane pattern map generation unit 33
The distortion distance is calculated as described above between the input iR pattern plain pattern pine f (hereinafter referred to as input brain pattern map) output from Then, calculate the amount of distortion for each defined pattern.

Next, a method for calculating the amount of distortion will be explained using both plane pattern maps shown in FIGS. 5(cL) and 5(A). The evaluation function representing the amount of distortion is given by the sum of the squares of the difference (d) in the weight values between corresponding pixels of both blurring and turn maps, but its calculation method is based on the definition play sonogram. This is carried out in two directions: from a black pixel with a weight of "1" in the input brain pattern map to a corresponding pixel in the input plane pattern map, and from a black pixel with a weight of "1" in the input brain seven-turn Mazda to a corresponding pixel in the defined plane pattern map. This is to strongly reflect the differences between both patterns in the evaluation function. For example, in the former case, as shown by arrow A in FIG. In this case, as shown by arrow B, the black pixel ■ of map (6) is
), and hereinafter, all black pixels with a weight of "1" in both maps are mutually compared with the corresponding pixels in the other map.

Here, in general, the weight of the pixel of the definition plane/turn map is set to d, the weight of the corresponding pixel of the input brain pattern map is set to 1, and then the distortion calculated from the definition/mother turn to the input/g turn is set as d. Let the amount be '1, and the amount of distortion calculated from the input pattern in the opposite direction toward the defined pattern be d.

d, and d! are the distortion distances (dj-d,
) or il:(d, -da) is given by the following equation.

d,=Σ(dj-d,)2 The wing section 34 calculates the amount of distortion for each input/turn for each defined pattern using this evaluation function. The minimum value selection section 36 is a distortion amount calculation section. 34 detects the minimum value among the distortion amounts for each definition/gutern, and outputs the definition pattern as a recognition result.

The method of the prior invention described above is extremely effective in recognizing patterns with large distortions such as handwritten characters, and the methods implemented so far have achieved good results. However, this method calculates the amount of distortion by performing bidirectional matching between the definition pattern and the input turn, so if the difference between the two patterns is large, the information amount of distortion is However, since the environmental information (weighting information of stroke direction and length) at each pixel between both patterns is not considered, it is necessary to reduce the difference in distortion amount and reduce ambiguity. This has the disadvantage that it acts to promote the recognition and reduces recognition accuracy.

[Object and structure of the invention]

An object of the present invention is to improve the evaluation function that gives the amount of distortion of each turn, in order to improve the drawbacks of the prior invention method,
The purpose of this invention is to improve the accuracy of turn recognition and the recognition processing efficiency.

Therefore, the present invention achieves the above object by combining the pattern matching method in the prior invention method with a phase structure method that uses stroke length and direction information as recognition information. The present invention weights the distortion distance between the definition pattern obtained by the /'e turn matching method and the input cover turn by the distance in the stroke length and direction information, thereby achieving the invention. The system uses the so-turn matching method to recognize patterns such as characters, symbols, figures, etc., and uses a plain pattern map that represents the distortion distribution for each pixel for each of multiple definition patterns and input turns. means for calculating the distortion distance for each pixel between each definition pattern and the input cover turn using the plane pattern map, and calculating the distortion distance of each definition pattern with respect to the input/turn based on the distortion distance; and a means for selecting one defined nosoturn that gives the minimum amount of distortion for the above-mentioned plurality of definition patterns and the input nosoturn and associating it with the input cover turn, and further, for each of the plurality of definition patterns and input nosoturns, for each pixel. means for creating a stroke information table indicating the length and direction of the stroke including the pixel; and means for determining the stroke information distance for each pixel of 1ljJ between each definition/ya turn and the input cover turn based on the stroke information table. The method is characterized in that the distortion distance for each pixel calculated by the means for determining the distortion amount of each defined pattern with respect to the input pattern is weighted by the stroke correction distance.

[Embodiments of the invention] The present invention will be explained below based on examples.

FIG. 6 shows eight directions and their codes (numbers) for defining the stroke information used in this embodiment. Stroke information is represented by length and these 8-direction codes. FIG. 7 is an explanatory diagram of stroke information using the 8-direction codes shown in FIG. 6.

This figure schematically shows the process of obtaining the stroke information distance between two handwritten character patterns rAJ and rBJ when focusing on the 0 point in matching. The stroke centered at the 0 point shown in (a) is normalized in 8 directions, each length is displayed in vector format in (b), and the difference between the two patterns is calculated by matching processing in (CC) and (cL). 9. This is further subjected to threshold processing and the like to obtain the stroke information distance.

To explain more specifically, the pattern r in the above example
The stroke information distance of the 0 point of AJ and rBJ is determined as shown in the following table.

The processes from (1) to (1) in the above table correspond to those explained with reference to FIG. The threshold processing of (V) is
The threshold value α=4 is set to binarize the stroke difference in each direction. The process (■1) is to calculate the sum of the result values for each direction binarized in (■). The total value C obtained in this way is (vii:
After normalization with + (K is a normalization coefficient), the stroke information distance P is used as a variable for weighting the evaluation function.

The evaluation function D' used in this embodiment is given by the following equation.

D' = d, '+c12' = Σ(1+P1) (tit-d・)2+Σ(1+P
, )(d・-db)2 d, ', and 41 in the above equation are the respective terms dI+'2 of the evaluation function D'''++dt described in the explanation of the prior invention method, and the stroke information distances. It is multiplied by 1 given term (1+PI) and (1+Pt).

Note that P represents the stroke information distance when matching is performed from the defined pattern to the input cover turn, and P2 represents the stroke information distance when matching is performed from the input z4 turn to the defined pattern.

FIG. 8 is a configuration diagram of a system according to an embodiment of the present invention, and FIG. 9 is a processing flow diagram thereof.

In FIG. 8, 81 is a pattern input section, 82 is an R-turn standardization section, 83 is a plain pattern map generation section,
Reference numeral 84 represents a black pixel corresponding turn generation unit, 85 a length and eight direction code generation unit, 86 a memory, 87 a distortion amount calculation unit, 88 a definition pattern library, and 89 a minimum value selection unit. The configuration consisting of each of these elements is based on the system according to the embodiment of the prior invention shown in FIG. 3, with partial additions and modifications according to the present invention.

The black pixel correspondence pattern generation unit 84 generates a black pixel correspondence number or turn indicating which black pixel is the base point of the weight of each pixel distributed in the plain pattern map.

The length and 8-direction code generation unit 85 performs processing for generating the length and 8-direction code of the loaf as described above using FIG. 7 and the table.

The memory 86 provides a work area for pattern recognition processing, and also holds various processing programs, such as a standardization program, a plain program, a black pixel correspondence program, a length and eight-direction code program, and a distortion calculation program. ing. Table I and table ■ installed in the memory are working tables used to generate the input brain pattern Mazda, and the numbering table for identifying each black pixel of the table RH person covert is a table, And table S [table showing black pattern i vs. 9 pattern, chiffle T length and 8 direction code table. Processing using these tables will be described later.

The definition plane pattern library 88 stores a plane/turn map for each definition/mother turn, a black pixel correspondence pattern, a length, and an 8-direction code table.

The operation of the embodiment system shown in FIG. 8 will be explained below according to the flowcharts (a) to ω) shown in FIG.

(α) The pattern input unit 81 cuts out the characters, symbols, figures, etc. to be recognized read by a scanner or the like into −degree units,
It is supplied to the pattern standardization section 82.

Cb) The pattern normalization section 82 normalizes each of the cut out patterns to the size of N.times.IVI dots, and supplies it to the plain pattern map generation section 83. Note that the standardized patterns here are stored in the memory 86.

(C) The plane pattern map generation unit 83u sets the coordinates of all black pixels in the black and white boundary area in the line portion of the standardized input pattern in a table IK, and sets the weight value J to "1". Next, remove one pixel from this table ■, set the value of the white pixel (“O” pixel) located at that distance in four directions to weight J+1 (J=1), and store its coordinates in table H. . This operation is executed for all pixels in table (2). Next, table
Extract each pixel, set the value of only white pixels in eight directions to J+1 (J=2), and enter the coordinates of that pixel in Table I. This is repeated "t" until there are no more pixels in table H.

By alternately performing the processing using the above tables I and I[ in the plane pattern map generation unit 83Vi, as many times as necessary, J=J+
The ordinal numbers are weighted by 1 to generate brain 4-turn mags for the input cover turns.

(d) The black pixel corresponding / turn generation unit 84 is configured as described above (
In parallel with the process of C), a process of generating a pattern corresponding to black pixels is performed. To do this, we first assign numbers A to the coordinates of all black pixels on the boundary, that is, pixels with weight value J=1.
(A=1.2, 3...) and store it in table P. and sequential weight values J for white pixels in C)
(J=J+1), the number k of the black pixel that is the base point of the weight J is assigned to the coordinates of the white pixel, as shown in Figure 10 (α, ψ). A pattern corresponding to black pixels is generated and stored in table S.

FIG. 10 shows a black pixel corresponding pattern corresponding to the plane pattern map shown in FIG. , (b) show an example of the input cover pattern rAJ generated by the black pixel corresponding pattern generation unit 84. In the former, 1 is assigned to the black pixel.
In the latter case, black pixels are numbered from 1 to 27, respectively.
Regarding the corresponding pixels on the plane pattern map of A),
The length and 8-direction code generation unit 85 displays the number of the black pixel that is the base point. - For each pixel of the manual pattern, calculates the length of the stroke centered at the pixel. Then, a length and 8-direction code table T is generated in which the 8-direction code is associated with each other. and correspond to patterns (α) and (b) shown in FIGS. 5 and 10, respectively.

(a) The distortion amount calculation unit 87 calculates the plain pattern pine f (table 1.II) generated for each cover pattern in each table in the memory 86, table R1 with black pixel number, black pixel pattern correspondence table S1 length and 8
Direction code table T and definition pattern standardization unit IJ
Based on similar data regarding each definition pattern in 88, the above-mentioned evaluation function D' is calculated.

At that time, for example, in the matching process from the definition pattern to the input/turn, in the case of the example shown by arrow A in FIG.
0 diagram (α), find the pixel with the weight ■ of the input flat turn (A) corresponding to the pixel number - "1"), and calculate (dA - d
, )2=(3-1)2=4. then,
From the black pixel correspondence/turn in Figure 10 Cb), find the black pixel "1" at the base point, read out the entry with pixel number "1" in both Figures 11 (cL) and (j5), and apply it to the cloth. Therefore, the stroke information distance P is calculated. For example, if P=2, we obtain P,=2. From these values, (1+P+) (” '・)2= (1+2)(3
-1) 2=6 is obtained. Below, calculations will be made for other pixels, and similar calculations will be performed for reverse matching processing.
The amount of distortion is determined from the evaluation function D'.

(,!7) The minimum value selection unit 89 detects the definition/mother turn that exhibits the minimum value from the distortion amount for each definition pattern calculated in (a) above, and outputs it as a recognition result.

〔Effect of the invention〕

As described above, according to the present invention, compared to the prior invention method, the pattern discrimination performance by the evaluation function is improved, the error rate can be reduced to a few fractions of a second, and a significant improvement effect can be obtained. It was done.

[Brief explanation of the drawing]

Figure 1 is a conceptual diagram of a conventional general pattern recognition system, Figure 2 is a conceptual diagram of a prior invention system, Figure 3 is a block diagram of an embodiment of the prior invention system, and Figures 4 (a) to (C ) is an explanatory diagram of the plane pattern map, 5th section 1 (a), (b) is a diagram showing an example of the hub brain/zoo turn mag, FIG. 6 is an explanatory diagram of the 8-direction code, and FIG. 7 (α), (A ), (C), (
d) is an explanatory diagram of stroke information, FIG. 8 is a configuration diagram of the system according to the present invention, FIG. 9 is a processing flow diagram thereof, and
Figures (α) and Cb) are explanatory diagrams of black pixel correspondence patterns, the first
Figure 1 (α), (gradient is the length and is an explanatory diagram of the 8-direction code table. In the figure, 81 is the turn input section, 82 is the main turn standard vs. Ui5 turn generation section, and 85 is the length) and an 8-direction code generator, 86 a memory, 87 a distortion amount calculator, 88
89 represents a definition pattern library, and 89 represents a minimum value selection section. Patent Applicant Fujitsu Law Company Representative Patent Attorney For long diameter Hiroshi Fumi (1 other person) 1-30 -3 -n-10 ◆2 3 years old 5
1 (Q) (b) (c)
(d) 18 figure Nina 10 figure (Q) (b) Sai 11 (C1) 88 button cord Y (b) 8 way copper cord "863

Claims (1)

[Claims] - A system for recognizing patterns such as characters, symbols, figures, etc. by the pattern matching method, which expresses the distortion distribution for each pixel for each of a plurality of definition patterns and input patterns. means for creating a plain pattern map; using the plain pattern map to calculate a pixel-by-pixel distortion distance between each defined pattern and the input cover pattern;
Based on this, the system is equipped with means for determining the amount of distortion of each definition pattern with respect to input turns of 1,000 turns, and means of selecting one definition pattern that gives the minimum amount of distortion for input cuts and turns and associating it with input turns. The invention further includes means for creating a stroke information table indicating the length and direction of a stroke including the pixel for each pixel for each of the plurality of definition patterns and input patterns, and means for creating each stroke information table based on the stroke information table. means for calculating the stroke information distance for each pixel between the definition pattern and the input cover turn, and the means for calculating the distortion amount of each definition/turn with respect to the input cover turn based on the stroke information distance for each pixel. A pattern recognition method characterized by weighting the distortion distance of.