JPH0143353B2 - - Google Patents

Description

Detailed Description of the Invention [Technical Field of the Invention] The present invention relates to a method for recognizing patterns such as handwritten characters, symbols, or figures using a pattern matching method. However, the patent application filed by the applicant in 1972-
90272 invention “Character/symbol recognition system”
16271, hereinafter referred to as the "prior invention"), which further improves the calculation method of the distortion distance between the defined pattern and the input pattern using the brain pattern map, and improves the recognition efficiency. It's a diagram.

[Technology background]

The conceptual classification of handwritten character recognition methods is as follows:
There are methods based on the topological structure of characters (for example, how strokes are combined) and methods based on matching with definition patterns (dictionary patterns), but the former method is troublesome to extract strokes, and the latter method results in character deformation. The disadvantage is that it is difficult to process the time. The above-mentioned prior invention, which is the basis of the present invention, is
Although it belongs to the latter category, it 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.

The amount of distortion D of the input pattern is used as a measure of the degree of matching between the definition pattern and the input pattern.
is defined, this is calculated for each definition pattern, and the definition pattern that minimizes the amount of distortion D with respect to the input pattern is set as the recognition pattern. This means that when each definition pattern is distorted (deformed) to match the input pattern, a definition pattern that requires the least amount of distortion, ie, a definition pattern that requires the least amount of deformation, is selected.

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

In Figure 1, 11-1, 11-2,...,1
1-n is a pattern of definition characters/symbols that serve as a reference, and is shown for example, "A", "B", . . . "+", etc. These definition character/symbol patterns are stored in advance in a definition library consisting of, for example, a magnetic disk device. Reference numeral 12 indicates a pattern of input characters/symbols to be recognized, which is detected by an optical reading means such as a scanner.
This character/symbol pattern 12 to be recognized is processed in similarity determination units 13-1, 13-2,..., 13-n.
Corresponding definition patterns 11-1 and 11-, respectively
2, . . . , 11-n, the degree of similarity is calculated, and the similarity amounts R ₁ , R ₂ , . . . Rn are output. The determination unit 14 determines the maximum similarity amount among these similarity amounts, and reads the recognized pattern with a definition pattern corresponding to this maximum similarity amount. This is a typical pattern matching method, but it is difficult to deal with the distortions of handwritten characters and symbols, resulting in the problem that highly accurate character and symbol recognition cannot be achieved.

Therefore, the prior invention effectively solves the problem of the conventional method that accurate reading cannot be performed due to the distortion that appears in handwritten characters and symbols, and replaces the recognition problem with the following.

``When pattern However, the essential difference is that in the latter case,
Based on the fact that any pattern can be created through distortion, we hypothesized that ``observed patterns are distortions of defined patterns.''

Specifically, for multiple defined characters/symbols that serve as a reference, the weighting value, that is, the degree of distortion, increases in proportion to the distance from each pixel of the pattern, so that it is set in advance according to a certain regularity. Generates a distortion pattern/map of defined character/symbol patterns. On the other hand, a distortion pattern map is similarly generated for the input recognition character/symbol pattern in accordance with the certain regularity. Then, the value of the evaluation function is determined using the difference between the respective weighting values of mutually corresponding parts of the distortion pattern maps of the definition pattern and the recognition pattern as variables, that is, matching is performed, and the evaluation function is Recognition is determined by the character/symbol of the defined pattern that has the minimum value.

FIG. 2 is a conceptual diagram of the system according to the prior invention.
In this figure, 21-1, 21-2,...21-n
is a standard definition character/symbol pattern. These patterns are in the form of distortion pattern maps generated from defined characters and symbols,
Patterns 11-1, 11-2,...11- in Figure 1
It is completely different from n. On the other hand, 22 is a pattern of input characters/symbols to be recognized, which is also in the form of a distortion pattern map, and is completely different from pattern 12 in FIG. These patterns 21-1, ..., 21-n, 22 are different from the patterns 11-1, ..., 11-n, 12 in FIG. ” concept has been introduced. This character/symbol pattern 22 to be recognized is processed by the corresponding patterns 21-1, 21-n, respectively, in the distortion amount calculation units 23-1, 23-2, ..., 23-n.
Each distortion amount is calculated between -2,...,21-n,
The distortion amounts D ₁ , D ₂ ,...Dn are output. Then, the determination unit 24 determines the minimum amount of distortion, and reads the pattern to be recognized using the definition pattern for this minimum amount of distortion.

FIG. 3 is a block diagram of an embodiment of the prior invention. In this figure, 31 is a pattern input section;
32 is a pattern standardization section, 33 is a brain pattern map 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 such as a character, symbol, figure, etc. to be recognized, such as "A", with a scanner, converts it into an electrical signal, and outputs it as mesh-like pixel information to the pattern standardization circuit 32.
Output to.

The pattern normalization circuit 32 normalizes the input pattern to a certain size for recognition processing and outputs it to the brain pattern map generation section 33. Note that the line width of the pattern is not changed in this standardization process.

The brain pattern map generation unit 33 generates, for each pixel of an arbitrary input pattern, a distortion pattern map called a brain pattern in which the weight, that is, the degree of distortion increases as the distance from the pixel increases.

Brain pattern maps are created by following certain mathematical rules, for example, as shown in Figure 4 a and b, for each pixel, pixels in the four directions of a and the eight directions of b are sequentially and alternately selected and weighted using ordinal numbers. By the method of performing this, the image is generated as illustrated in c of the same figure. In addition, the above mathematical rule is
It is called “Octagonal Distance”.

FIGS. 5a and 5b, which will be described later, show examples of brain pattern maps generated for the letter "A". a is a definition pattern, and b is an input pattern.

In this brain pattern map, in the case of a pattern with black letters on a white background, a weight of "1" is set for all black pixels at the black-and-white boundary, and the weight is set as 2, 3, 4, etc. in the direction away from the black pixel. An increasing ordinal weight is set. In the brain pattern map example shown in Figure 5, the line width of the letter "A" is a unit width of one pixel, but in the case of a thick line with multiple pixel widths, towards the inside of the line, 0, -1, -
2, . . . are weighted in the negative direction.
By using such a brain pattern map, the distortion distance from the defined character can be easily determined for any distorted handwritten character. For example, if 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 section 34 in FIG. When calculating the distortion distance as described above between the brain pattern map of the input pattern output from (hereinafter referred to as the input brain pattern map), each defined pattern is distorted to match the input pattern. The amount of distortion for each defined pattern is determined.

Next, a method for calculating the amount of distortion will be explained using both brain pattern maps shown in FIGS. 5a and 5b. 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 brain pattern maps, but its calculation method is based on the weight of the defined brain pattern map. This is performed 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 brain pattern map, and from a black pixel with a weight of "1" in the input brain pattern map to a corresponding pixel in the defined brain pattern map. This is to strongly reflect the differences between both patterns in the evaluation function. For example, in the former case, the black pixels of the defined brain pattern map a are compared with those of the input brain pattern map b, as shown by arrow A in FIG. The black pixels in map b are compared with those in map a, and below, in both maps, all black pixels with a weight of "1" are
Each map is compared with the corresponding pixel in the other map.

Here, generally, the weight of a pixel in the defined brain pattern map is d _a , the weight of the corresponding pixel in the input brain pattern map is d _b , and the amount of distortion calculated from the defined pattern to the input pattern is
d ₁ , and the amount of distortion calculated from the input pattern in the opposite direction toward the defined pattern is d ₂ . d ₁
and d ₂ are distortion distances (d _b −
It is given by the following equation as the sum of squares of d _a ) or (d _a −d _b ).

d ₁ =Σ(d _b −d _a )2 d ₂ =Σ(d _a −d _b )2 This total D=d ₁ +d ₂ is the evaluation function, and the distortion amount calculation unit 34 uses this evaluation function D. , the amount of distortion for the input pattern is calculated for each defined pattern.

The minimum value selection unit 36 detects the minimum value among the distortion amounts for each definition pattern calculated by the distortion amount calculation unit 34, and outputs the definition pattern as a recognition result.

The prior invention method described above is extremely effective in recognizing patterns with large distortions such as handwritten characters, and those that have been implemented so far have achieved good results. However, in this method, the amount of distortion is calculated by performing bidirectional matching between the defined pattern and the input pattern, so if there is a large difference between the two patterns, the amount of distortion information increases. Although recognition accuracy can be improved, if there are many matching parts between both patterns, distortion information will be calculated redundantly.
This method has the disadvantage of reducing the difference in the amount of distortion for each definition pattern, increasing ambiguity, and reducing recognition accuracy.

[Object and structure of the invention]

SUMMARY OF THE INVENTION An object of the present invention is to improve the evaluation function that provides the amount of distortion of a pattern, and to improve the precision of pattern recognition and recognition processing efficiency, in order to improve the drawbacks of the prior invention method.

The present invention is a system for recognizing patterns such as characters, symbols, figures, etc. using a pattern matching method, which uses a brain pattern map that represents the distortion distribution for each pixel for each of a plurality of definition patterns and input patterns. means for calculating the distortion distance for each pixel between each definition pattern and the input pattern using the brain pattern map, and determining the amount of distortion of each definition pattern with respect to the input pattern based on the distortion distance for each pixel between each definition pattern and the input pattern; and a means for selecting one definition pattern that gives a minimum amount of distortion to the input pattern and associating it with the input pattern, the means for determining the amount of distortion of the definition pattern with respect to the input pattern for each definition pattern is The first step is to calculate the distance between corresponding pixels of the input pattern as seen from the pattern using both brain pattern maps, and the pixels of the input pattern for which the distance calculation was performed in the first step are eliminated so as not to overlap. The present invention is characterized by having two execution stages: a second stage in which distances between corresponding pixels of the definition pattern viewed from the input pattern are calculated using both brain pattern maps for the remaining pixels of the input pattern.

[Embodiments of the invention]

The present invention will be explained below using examples.

In the distortion amount calculation by double matching between the definition pattern and the input pattern in the prior invention method described above, the present invention calculates the corresponding black pixel of the input pattern that has already been calculated by matching in the direction from the definition pattern to the input pattern. By excluding the input pattern from the calculation target when matching the input pattern to the definition pattern, matching portions between patterns are prevented from being calculated twice. Therefore, as in the case of matching shown by arrow A in FIG. 5, when a black pixel with a weight on the defined brain pattern map a is matched with a pixel with a weight on the input brain pattern map b, It must be possible to identify which black pixel in the input pattern that is the base point of the pixel represents the weight, that is, the distortion. Only when the black pixel serving as the base point can be identified can the black pixel in question be excluded from the calculation when matching is performed in the opposite direction next time.

In the embodiment of the present invention, the black pixel correspondence pattern illustrated in FIG. 8 is used to display the base black pixel for each weight in the brain pattern map as described above. Note that FIG. 8 will be described later.

FIG. 6 is a schematic diagram of a pattern recognition system implementing the method of the present invention. In this figure, 6
1 is a pattern input section, 62 is a pattern standardization section,
63 is a brain pattern map generation unit, 64 is a black pixel corresponding pattern generation unit, 65 is a memory, 66 is a distortion amount calculation unit, 67 is a definition pattern library,
68 is a minimum value selection unit, and its basic configuration is the same as the system according to the embodiment of the prior invention described in FIG. be.

The black pixel corresponding pattern generation unit 64 calculates the weight of each pixel distributed in the brain pattern map as follows:
A black pixel correspondence pattern indicating which black pixel is the base point is generated.

The memory 65 provides a work area for pattern recognition processing, and also stores various processing programs,
That is, it holds a standardization program, a brain program, a black pixel correspondence program, a distortion calculation program, etc. The tables and tables located in memory are working tables used to generate the input brain pattern map, and table P is a numbering table for identifying each black pixel of the input pattern; Q is a table showing black pixel correspondence patterns. Processing using these tables will be described later.

The definition pattern library 67 stores a brain pattern map and a pattern corresponding to black pixels for each definition pattern.

FIG. 7 shows a schematic processing flow of the system of this embodiment. The operation of the embodiment system shown in FIG. 6 will be explained below according to flows a to g in FIG. 7.

(a) The pattern input unit 61 cuts out characters, symbols, and figures to be recognized read by a scanner or the like one unit at a time, and supplies the extracted characters to the pattern standardization unit 62.

(b) The pattern normalization unit 62 normalizes each cut out pattern to a size of N×M dots and supplies it to the brain pattern map generation unit 63.
Note that the standardized patterns here are stored in the memory 65.

(c) The brain pattern map generation unit 63 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, and sets the weight value J to "1". Next, one pixel at a time is extracted from this table, the value of a white pixel (pixel of "0") located at the distance in the four directions is given a weight of J+1 (J=1), and its coordinates are stored in the table. This operation is executed for all pixels in the table. Next, take out each pixel from the table and set the value of only white pixels in 8 directions to J + 1 (J = 2)
and enter the coordinates of that pixel in the table.
Repeat this until there are no more pixels on the table.

The brain pattern map generation unit 63 alternately executes the process using the above table as many times as necessary, weights the ordinal number by J=J+1 with each black pixel as the base point, and generates the input pattern. Generate a brain pattern map for.

(d) The black pixel corresponding pattern generating section 64 performs a process of generating a black pixel corresponding pattern in parallel with the process of (c) above. To do this, first, numbers k (k=1, 2, 3, . . . ) are sequentially assigned to the coordinates of all black pixels on the boundary, that is, pixels with weight value J=1, and stored in table P. Then, in (c), weight values J are sequentially assigned to white pixels (J=J
+1), assign the number k of the black pixel that is the base point of the weight J to the coordinates of the white pixel, generate the black pixel correspondence pattern as shown in Fig. 8a and b, and create a table. Store in Q.

FIG. 8 shows the black pixel correspondence pattern corresponding to the brain pattern map shown in FIG. 5, and (a)
shows one example of the black pixel correspondence pattern of the definition pattern “A” stored in advance in the definition pattern library 67, and (b) shows one example of the black pixel correspondence pattern for the input pattern “A” generated by the black pixel correspondence pattern generation unit 64. Give an example. In the former, black pixels start from 1.
In the latter case, black pixels are numbered from 1 to 27, and for each corresponding pixel on the brain pattern map shown in Figures 5a and b, the number of the black pixel that is the base point is assigned. It shows.

(e) The distortion amount calculation unit 66 first performs matching processing from the definition pattern to the input pattern as shown by arrow A in FIG. At that time, the 8th
The base point black pixel number is determined by referring to the black pixel correspondence pattern for the input pattern as shown in FIG. 9, and a table is created to check whether the calculation as shown in FIG. When the calculation has been performed, "1" is set in the check column for the corresponding black pixel number in the check table.

(f) The distortion amount calculation section 66 then performs reverse matching processing as shown by arrow B in FIG. At this time, as a black pixel of the input pattern,
Only unchecked black pixel numbers for which "1" is not set in the check table of FIG. 9 are selected to calculate the amount of distortion. As a result, between pixels with a high degree of matching between the definition pattern and the input pattern, in other words, between pixels with little distortion,
Duplicate calculations are almost never performed in bidirectional matching, and redundant elements included in the amount of distortion can be reduced.

(g) The minimum value selection unit 68 compares the distortion amount whose redundancy has been improved in the above (f) between each definition pattern, determines the minimum value, and outputs the result 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, and the error rate can be reduced to a fraction of that, resulting in a significant improvement effect. It was done.

[Brief explanation of drawings]

Figure 1 is a conceptual diagram of a conventional general pattern recognition method, Figure 2 is a conceptual diagram of a prior invention method, Figure 3 is a block diagram of an embodiment of the prior invention method, and Figures 4 a to c are brain patterns. An explanatory diagram of the map, Figures 5a and 5b are diagrams showing examples of brain pattern maps,
FIG. 6 is a configuration diagram of an embodiment of the present invention, FIG. 7 is a processing flow diagram thereof, FIG. 8 is a diagram showing an example of a black pixel correspondence pattern, and FIG. 9 is an explanatory diagram of a check table for avoiding duplicate calculations. It is. In the figure, 61 is a pattern input section, 62 is a pattern standardization section, 63 is a brain pattern map generation section, 64 is a black pixel corresponding pattern generation section, 65 is a memory, 66 is a distortion amount calculation section, 67 is a definition pattern library, 68 represents a minimum value selection section.

Claims (1)

[Claims] 1 Characters, symbols,
A system for recognizing patterns such as figures, which includes means for creating a brain pattern map representing the distortion distribution for each pixel for each of a plurality of definition patterns and input patterns, and each definition pattern and input using the brain pattern map. means for calculating the distortion distance for each pixel between the patterns and calculating the distortion amount of each definition pattern with respect to the input pattern based on the distortion distance; and selecting one definition pattern that gives the minimum distortion amount to the input pattern;
The means for determining the amount of distortion of the definition pattern with respect to the input pattern for each definition pattern calculates the distance between the corresponding pixels of the input pattern as seen from the definition pattern using both brains. In the first step, which is performed using a pattern map, the pixels of the input pattern whose distances have been calculated in the first step are removed by eliminating the corresponding pixels that have been calculated once so as not to overlap the pixels of the input pattern. 1. A pattern recognition method comprising two execution steps: a second step in which distances between corresponding pixels of a defined pattern as seen from each other are calculated using both brain pattern maps.