JPH06348899A - Character recognizing device - Google Patents

Abstract

PURPOSE:To provide a character recognizing device capable of obtaining highly accurate and stable recognition performance even for low quality character patterns with large variation in a local line width. CONSTITUTION:Four kinds of sub patterns for representing the distribution state of the strokes of characters based on the average line width source character patterns are extracted at every scanning direction and thinning pattern composed of the patterns belonging to neither of extracted sub patterns are extracted. The presence/absence of the need of extracting the sub patterns is judged for the remaining patterns for which infinitesimal segments are removed from the thinning patterns and thinning sub patterns based on the average line width of the pertinent thinning pattern are extracted when it is judged that the extraction is needed. Thereafter, either of the sub pattern and the synthesis sub pattern of the sub pattern with the thinning sub pattern is outputted to a feature extraction part 135, feature extraction relating to the pertinent character pattern is performed and the characters are recognized by collating the features with the features of a dictionary.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a high-accuracy character recognizing device which copes with a character pattern having a partially different line width, such as a faint part.

[0002]

2. Description of the Related Art Conventionally, characteristics of input character patterns are extracted,
As a character recognition device that outputs a recognition result by collating with a dictionary prepared in advance, for example, Japanese Patent Publication No. 60-3875.
6 was disclosed. The outline of the processing by the character recognition device will be described below.

First, the brightness of each cell of the input character pattern is converted into a binary image which is a quantized electric signal by photoelectric conversion and stored in a pattern register. Then, the circumscribing frame of the character pattern in the pattern register is detected, and the line width of the character pattern in the circumscribing frame is calculated. Next, the character pattern in the circumscribing frame is scanned horizontally, vertically, diagonally to the right, and diagonally to the left to detect continuous black pixel components with the line width as a threshold value, thereby determining four types of the input character pattern. The sub-pattern of is extracted. Further, the character pattern in the circumscribing frame of the pattern register is non-linearly divided into N × M grid-shaped partial regions so that the number of black pixels in each divided region is the same. Next, for each of the four types of sub patterns,
The number of black pixels of the sub-pattern in the divided partial area is counted and normalized by the size of the character pattern to reflect the distribution state of the character lines in each direction, which is an N × M × 4 dimensional feature. Extract the matrix. The feature matrix is collated with a dictionary, which is a feature matrix of a plurality of standard characters prepared in advance, and the recognition result of the input character pattern is output from the collation result.

[0004]

However, the above character recognition device has the following problems. That is,
In the conventional technique, the binary image in the circumscribed frame of the input character pattern is scanned in each of four directions of horizontal, vertical, right diagonal, and left diagonal, and a threshold value is twice the average line width of the character pattern. As an example, a stroke consisting of continuous black pixels was extracted, and four types of sub-patterns representing their distribution were extracted. However, the conventional sub-pattern extraction method naturally does not extract the stroke component having the number of continuous black pixels smaller than twice the line width. Therefore, a character pattern that is partly faint, that is, a character pattern whose local line width is extremely smaller than other parts, is extracted based on the sub-pattern and sub-pattern. However, there was a problem in that it was not reflected in the feature matrix, and as a result, it contributed to the deterioration of recognition performance.

An example of such a case is shown in FIGS. 4 and 6 (a).
Shown in. FIG. 4 shows a case where the beard portion of the letter “Q” of the alphabet, that is, the segment in the area surrounded by the wavy line 401 is faintly divided into three.
This faint whisker is normally detected as a part of the stroke by scanning in the left diagonal direction, but in this example, as a part of the sub-pattern in scanning in any direction. Not detected. Therefore, the beard portion is not reflected in the feature, and the probability of misreading a similar character without a beard, such as "O", increases.

Further, in FIG. 6 (a), the kanji character "Cause" is used. In this case, of the elements that make up the "cause," the "mouth," which is the outer part, is "large," which is the inner part.
Is written smaller than usual and has a size of not more than twice the average line width of the entire character. Therefore, in scanning using the average line width, FIG. 6 (b), FIG. 6 (c), and FIG. ), FIG. 6 (e)
As shown in, there is a situation in which the character “large” is not reflected in any sub-pattern, which is a big problem.

Further, even when the character pattern is not faint or written small as described above, the value of the average line width becomes very large due to the part being crushed. There is a problem in that the stroke is not detected as a sub-pattern by scanning and is therefore not reflected in the feature matrix, resulting in deterioration of recognition performance.
An example of such a case is shown in FIG. FIG. 5 shows an example in which the lower part of the numeral "5" has been formed into a loop, but at this time, the average line width of the entire character becomes a large value, and as a result, for example, a wavy line 501 is shown. The stroke portion or the like which is normally written as described above has a size of twice the line width or less, and is eventually not extracted as a sub pattern. Therefore, since the feature is extracted as a pattern having no stroke indicated by the wavy line 501, for example, "6".
And so on, and the probability of misreading as "6" increases.

According to the present invention, in the above-mentioned conventional sub-pattern extraction method, when the stroke width of each stroke component constituting a character is distributed near the average stroke width, effective feature extraction can be performed, while local feature extraction can be performed. If the stroke width is significantly different from that of other parts, that is, if the character pattern is partially crushed or faint, proper feature extraction cannot be performed. By eliminating the problem that the recognition performance deteriorates, each strobe with different local line width
Each sub-component is scanned from each direction with an optimal threshold value to extract a sub-pattern, and feature extraction and recognition processing based on the sub-pattern are performed to obtain a local line width with a large variation in quality. It is an object of the present invention to provide a character recognition device that can obtain stable recognition performance with high accuracy even for bad character patterns.

[0009]

In order to solve the above problems, the present invention optically scans a character pattern written on a form or the like and converts it into a binary image which is a quantized electric signal. A photoelectric conversion unit, a pattern register for storing the character pattern converted into the binary image, a circumscribing frame detecting unit for detecting a circumscribing frame of the character pattern in the pattern register, and a circumscribing frame of the pattern register. A character pattern line width calculation unit for calculating the line width of a character pattern, and a black obtained by scanning the character pattern in the circumscribing frame of the pattern register in the horizontal, vertical, right diagonal, and left diagonal directions, respectively. A sub-pattern extraction unit that detects a stroke when the number of consecutive pixels exceeds a threshold value determined based on the line width and extracts four types of sub-patterns representing the distribution of these strokes in each direction. , The pattern From the binary image in the circumscribing frame of the register and the four types of sub-patterns, among the black pixels forming the character pattern, a black pattern that does not belong to any of the four types of sub-patterns is extracted as a fading pattern. Section, a fine segment removal unit that removes a fine segment among the independent segments that form the blur pattern, a blur pattern line width calculation unit that calculates the line width of the blur pattern, and a blur that has removed the fine segment. When it is determined that it is necessary to extract a sub-pattern for a pattern, four types of faint sub-patterns representing the stroke distribution of the faint pattern are extracted for each scanning direction based on the line width of the faint pattern. The blurring sub-pattern extraction unit and the sub-pattern or the sub-pattern and the blurring sub-pattern are combined for each scanning direction. The control unit that outputs any one of the synthesized sub-patterns to the feature extraction unit, the feature extraction unit that extracts the feature matrix of the sub-pattern or the synthesized sub-pattern, and the feature matrix and the dictionary matrix prepared in advance. And an identification unit that outputs a recognition result based on the result.

[0010]

According to the present invention, four types of sub-patterns representing the distribution of the strokes of characters based on the average line width of the original character pattern are extracted for each scanning direction, and do not belong to any of the extracted sub-patterns. A faint pattern consisting of patterns is extracted. It is determined whether or not sub-patterns need to be extracted from the remaining patterns obtained by removing minute segments from this blurring pattern, and if it is determined that extraction is necessary, a blurring sub-pattern based on the average line width of the blurring pattern is extracted. To be done. After that, when the sub pattern or the faint sub pattern is extracted, one of the combined sub patterns of the sub pattern and the faint sub pattern is output to the feature extraction unit under the control of the control unit to extract the feature related to the character pattern. Then, character recognition is performed by collating this feature with the feature of the dictionary. Therefore, even if a part of the stroke component that constitutes the character pattern, which plays an essential role in character recognition, becomes faint or small, it can be relieved and extracted as a faint sub-pattern, so that it can be locally extracted. It is possible to obtain stable recognition performance with high accuracy for other low-quality character patterns having line width variations.

[0011]

[Embodiment] Embodiment 1 of the character recognition apparatus according to the present invention will be described below.
And Example 2 will be described. Here, for example, in FIG.
1, 501 in FIG. 5 and the element “large” and the like that constitute the “factor” in FIG. 6A are collectively referred to as a blur pattern for convenience. In addition, in the first embodiment, an example in which the present embodiment is applied to the binary image of the Chinese character “Cause” in FIG.

FIG. 1 is a block diagram showing a first embodiment of a character recognition device according to the present invention. Here, 101 is an optical signal input obtained by scanning a character pattern with a scanner, 10
2 is a photoelectric conversion unit, 103 is a pattern register, 104 is a circumscribing frame detection unit, 105 is a character pattern line width calculation unit, 106 is a horizontal scanning unit, 107 is a horizontal sub pattern 1 memory, 1
Reference numeral 08 is a vertical scanning unit, 109 is a vertical sub-pattern 1 memory, 110 is a right diagonal scanning unit, 111 is a right diagonal sub pattern 1 memory, 112 is a left diagonal scanning unit, 113 is a left diagonal sub pattern 1 memory, and 114 is a faint pattern. An extraction unit, 115 is a blur pattern memory, 116 is a blur pattern line width calculation unit, 117 is a horizontal scanning unit, 118 is a horizontal blur sub-pattern memory, 119 is a vertical scanning unit, 1
Reference numeral 20 is a vertical blurring sub-pattern memory, 121 is a diagonal right direction scanning unit, 122 is a right diagonal blurring sub pattern memory, 1
23 is a left oblique direction scanning unit, 124 is a left oblique fading sub pattern memory, 125 is a horizontal sub pattern combining unit, 126 is a horizontal sub pattern 2 memory, 127 is a vertical sub pattern combining unit, 128 is a vertical sub pattern 2 memory, and 129 is a right oblique sub pattern. A synthesizing unit, 130 is a right diagonal sub-pattern 2 memory, 131 is a left diagonal sub-pattern 2 synthesizing unit, 132 is a left diagonal sub-pattern 2 memory, 133 is an output control unit, 134 is a line width determining unit, 135 is a feature extracting unit, 136 is an identifying unit. , 13
Reference numeral 7 is a dictionary memory, 138 is a recognition result, and 139 is a minute segment removing unit.

An optical signal 101 obtained by scanning a character pattern handwritten or printed on a form or the like with a scanner is converted into an electric signal in a photoelectric conversion unit 102 and further quantized to be a binary signal. It is converted into a binary image and stored in the pattern register 103.

The circumscribing frame detection unit 104 is used in the pattern register 1
For the binary image stored in 03, the upper and lower ends of the binary image are detected by horizontal scanning, the left and right ends of the binary image are detected by vertical scanning, and as a result, the input character pattern is circumscribed. Get a circumscribed frame that is a rectangle. Then, the coordinate values regarding the circumscribing frame are set to the character pattern line width calculation unit 10
5, the horizontal scanning unit 106, the vertical scanning unit 108, the right oblique direction scanning unit 110, the left oblique direction scanning unit 112, and the blur pattern extracting unit 114, and specifies the character pattern area. In the following processing, the pattern register 103
When the binary image of is used, all the binary images within the circumscribed frame are targeted.

The character pattern line width calculation unit 105 calculates the average line width of the character pattern. Here, as one example of how to obtain the average line width, the following method is adopted in the present embodiment. That is, when the number of black pixels of the binary image of the character pattern in the circumscribed frame of the pattern register 103 is A and the number of black pixels is Q, the average line width Wr is calculated by the following formula. Wr = A / (A−Q) (1) However, 4 black pixels means that all the 2 × 2 windows become black pixels when the binary image is scanned by the 2 × 2 window. The black pixel number Q is a count of such 4 black pixels.

Next, with respect to the character pattern in the circumscribed frame of the pattern register 103, the horizontal scanning unit 106 horizontally, the vertical scanning unit 108 vertically, and the right diagonal scanning unit 110 diagonally right. , The diagonally leftward scanning unit 112 scans diagonally leftward, detects strokes that are continuous black pixels by using the value based on the line width as a threshold, and reflects their distribution state. Generate a sub pattern. At this time, the condition that the continuous black pixels are the stroke components forming the sub-pattern is given by the following equation, where L is the number of continuous black pixels. L> 2 * Wr (2) Here, Wr is the above-mentioned average line width. That is, in the scanning in each direction, a stroke having a length that is more than twice the average line width is extracted as a stroke that constitutes a sub pattern in that direction. The distribution state of strokes as continuous black pixels in the circumscribing frame detected as described above is as follows: horizontal sub-pattern 1, vertical sub-pattern 1, right diagonal sub-pattern 1, left diagonal sub-pattern 1, for each scanning direction. Horizontal sub-pattern 1 memory 107,
The data is stored in the vertical sub-pattern 1 memory 109, the right diagonal sub-pattern 1 memory 111, and the left diagonal sub-pattern 1 memory 113.

Taking FIG. 6 as an example, the original binary image before scanning is shown in FIG. 6A, the horizontal sub-pattern 1 is shown in FIG. 6B, the vertical sub-pattern 1 is shown in FIG. The sub pattern 1 is shown in FIG. 6 (d), and the left diagonal sub pattern 1 is shown in FIG. 6 (e). As described above, it can be seen that the element "large" that constitutes the "factor" is not reflected in each sub-pattern because it is a scale of twice the average line width or less.

The next blur pattern extraction unit 114 extracts a sub-pattern by using the binary image in the circumscribing frame of the pattern register 103 and the horizontal sub-pattern 1, the vertical sub-pattern 1, the right diagonal sub-pattern 1, and the left diagonal sub-pattern 1. The distribution state of the stroke components that did not exist is extracted as a blur pattern. FIG. 3 is a block diagram showing a configuration example of the blur pattern extracting unit, and the inside of the frame indicated by the dotted line represents the inside of the blur pattern extracting unit 114. 301 is an OR circuit unit,
Reference numeral 302 is a memory, 303 is a NOT circuit unit, 304 is a character pattern memory, and 305 is an AND circuit unit.

The faint pattern extraction unit 114 shown in FIG.
The function and processing flow of each block in FIG. First, the sub pattern memory 107 for each direction,
The horizontal sub-pattern 1, the vertical sub-pattern 1, the right diagonal sub-pattern 1 and the left diagonal sub-pattern 1 stored in 109, 111 and 113 are input to the OR circuit unit 301. OR
In the circuit unit 301, assuming that the black pixel of each sub-pattern is 1 and the white pixel is 0, each pixel in the sub-pattern area surrounded by the circumscribing frame has an O value of the pixel value of four sub-patterns 1.
The R logical operation is executed, and the operation result is previously stored in the memory 3
The sub-pattern area prepared in No. 02 is output for each corresponding pixel in the same phrase-shaped area, and finally a pattern that is the union of four sub-patterns 1 is generated in the memory 302. This pattern is a black pixel when the pixel value of at least one sub-pattern among the four sub-patterns in each pixel of the area is 1, that is, a black pixel, and the pixel values of all four sub-patterns 1 are 0, that is, a white pixel is a white pixel. Therefore, the white pixel part of the pattern of the union of this sub-pattern is
Either it was originally a white pixel in the binary image of the character pattern, or it was a black pixel in the binary image but was not extracted as a sub-pattern.

Next, the NOT circuit section 303 performs a NOT operation on the pattern generated on the memory 302. The NOT circuit unit 303 sequentially converts a pixel having a pixel value of 0 into a pixel value of 1 and a pixel having a pixel value of 1 into a pixel value of 0 for each of the pixels forming the pattern on the memory 302, that is, A NOT operation for converting a white pixel into a black pixel and a black pixel into a white pixel is executed, and the operation result is stored in the memory 30.
It outputs on the said pixel in 2. As described above,
On the memory 302, a pattern in which the pattern, which is the union of the sub patterns generated by the OR circuit unit 301, is inverted in black and white is generated.

On the other hand, independently of the above-mentioned processing, only the binary image in the circumscribing frame detected by the circumscribing frame detecting unit 104 among the binary images in the pattern register 103 is transferred to the character pattern memory 304.

Next, for the pattern on the memory 302 and the character pattern on the memory pattern 304, the AND circuit unit 3
An AND operation is executed by 05. In the AND circuit unit 305, for each pixel in the pattern area, an AND operation is performed between the pixel value of the pattern on the memory 302 and the pixel value of the character pattern on the character pattern memory 304 corresponding to the pixel, that is, both pixels. Only when the value is 1, the pixel value 1 is output, and when at least one of them is 0, an operation of outputting the pixel value 0 is executed, and the operation result is used as a blur pattern to make a blur. Pattern memory 115
Output to.

As can be understood from the above description, this blur pattern is obtained by extracting black pixels constituting the character pattern that do not belong to any of the black pixels of the four sub patterns 1. is there. That is, for example, as shown by 401 in FIG. 4, a faint pattern is a stroke formed by a part of the stroke being divided into several segments.
Or a stroke originally isolated as indicated by 501 in FIG. 5 and which does not reach the threshold value of twice the average line width shown in equation (2). When the processing of the blur pattern extracting unit is applied to the original binary image of FIG. 6A, the processing of FIGS. 6A to 6B, 6C, 6D,
All the black pixels of each sub-pattern of (e) are removed, and therefore, as shown in FIG. 6 (f), a blur pattern including only the element “large” that is not extracted as a sub-pattern is obtained.

The faint pattern extracted by the faint pattern extracting unit 114 is stored in the faint pattern memory 115. If necessary, the fine segment removing unit 13 for removing the fine segment in the faint pattern.
It is also possible to install 9. For example, the following can be considered as a removal rule of the minute segment by the minute segment removing unit 139. That is, the number of contour black pixels forming the contour of each segment forming the faint pattern or the total number of black pixels of each segment is a predetermined threshold, for example, β times (β> 0) the line width Wr of the input character pattern. ) It is a rule to consider it as a minute segment when it is below. Here, the segment determined to be minute is erased on the blur pattern memory or excluded from the processing. By deleting the minute segment as described above, it is possible to prevent deterioration of the recognition performance due to the deletion.

The faint pattern line width calculation unit 116 calculates the faint pattern line width. As the method of calculating the line width, for example, the method based on the equation (1) used in the character pattern line width calculation unit 105 is used.

Next, with respect to the blur pattern in the blur pattern memory 115, the horizontal scanning unit 117, the vertical scanning unit 119, the right diagonal scanning unit 121, and the left diagonal scanning unit 123 respectively perform horizontal, vertical, and right scanning. Scanning is performed diagonally and diagonally to the left, and continuous black pixels exceeding a predetermined threshold are detected as strokes. As a result, the sub-patterns of the blurring pattern, that is, the blurring sub-patterns are extracted this time, and the horizontal blurring sub-pattern memory 11 is extracted.
8. The vertical blurring sub pattern memory 120, the right diagonal blurring sub pattern memory 122, and the left diagonal blurring sub pattern memory 124 are stored. Here, the condition for the stroke component forming the sub-pattern is not given by the equation (2), but given by the following equation when the number of consecutive black pixels is L.

[0027]

[Equation 1]

However, Ws is the fading pattern line width calculation unit 1
It is the line width value of the fading pattern calculated by 16.
In addition, the reason why γ is not set to 2 in the equation (4) is because an error caused by calculating the line width in a smaller area than usual is taken into consideration, and it is considered that γ is multiplied by the correction factor. . This correction factor is empirically determined, but, of course, the threshold value may be normally set as γ = 2.

The process of extracting the faint sub-pattern will be described with reference to FIG. First, a partial pattern "large" extracted as a blur pattern from the character "Cause" is shown in FIG. 6F, and scanning is performed horizontally, vertically, diagonally to the right, and diagonally to the left with respect to this blurred pattern. The faint sub-patterns obtained by the above are shown in FIGS. 6 (g), (h), (i), and (j), respectively. As described above, the threshold value in the scan is determined based on the line width of the faint pattern “large” in FIG. 6F, not the line width of the original binary image “factor” in FIG. 6A. . Therefore, when scanning the original binary image of FIG.
It can be seen that the "large" sub-patterns that were not extracted because the line width value was large were properly extracted by scanning with a suitable line width value.

Next, a horizontal sub-pattern 1, a vertical sub-pattern 1, a right diagonal sub-pattern 1, a left diagonal sub-pattern 1 extracted by scanning the original binary image, and a horizontal blur sub-pattern extracted by scanning the blur pattern,
Processing for combining the vertical blurring subpattern, the right diagonal blurring subpattern, and the left diagonal blurring subpattern is performed.
This synthesizing process is executed by the horizontal sub-pattern synthesizing unit 125, the vertical sub-pattern synthesizing unit 127, the right diagonal sub-pattern synthesizing unit 129, and the left diagonal sub-pattern synthesizing unit 131 independently for each sub-pattern in each direction. The combined patterns are a horizontal sub-pattern 2, a vertical sub-pattern 2, a right diagonal sub-pattern 2 and a left diagonal sub-pattern 2, respectively, which are a horizontal sub-pattern 2 memory 126, a vertical sub-pattern 2 memory 128 and a right diagonal sub-pattern 2 memory 13, respectively.
0, stored in the left diagonal sub-pattern memory 132.

Here, the pattern composition in the composition section is performed by, for example, O for each pixel of two sub patterns.
A method of performing R calculation or the like is used. That is, in the case of combining two binary patterns, at least one of the individual pixels forming each outputs the pixel value 1, that is, the pixel value 1 if it is a black pixel, and both output the pixel value 0, That is, a composite pattern is created by a method of outputting a pixel value of 0 when it is a white pixel. In the case of the example of FIG. 6, as the sub-pattern of the original binary image, FIG.
(C), (d), and (e) are given, and subpatterns of the blur pattern are shown in FIGS. 6 (g), 6 (h), and 6 (h), respectively.
When (i) and (j) are given, the sub-pattern 2 synthesized by the synthesizing unit is as shown in FIG.
(L), (m), and (n). Compared with the sub-pattern of the original binary image, these combined sub-patterns 2 accurately reflect the small local scale, so that the feature matrix calculated based on that sub-pattern is also reflected. Naturally, this is reflected, and therefore erroneous reading due to information deterioration due to the conventional sub-pattern extraction can be prevented.

The sub-pattern created by the above method is converted into a further compressed feature in the feature extraction unit 135. In this embodiment, the feature extraction is performed under the control of the output control unit 133. A sub pattern to be input to the unit 135 can be selected. The output control unit 133, when the minute segment removing unit 139 determines that all the segments forming the blurred pattern are all minute, stops the processing for the blurred pattern thereafter and obtains it from the original binary image. The sub-pattern 1 in each direction is stored in each of the memories 107, 109, 111, 1
The data is read from the data 13 and output to the feature extraction unit 135. Also,
The line width of the blur pattern calculated by the blur pattern line width calculation unit 116 is always determined by the line width determination unit 134, and when the line width is determined to be equal to or less than a predetermined threshold value, the determination is made. The result is transmitted to the output control unit 133. At this time, for example, the following equation is given as the threshold for the line width. Ws <δ × Wr (5) 0 <δ << 1 (6) where Ws is the average line width of the blur pattern, and Wr is the average line width of the original binary image, and the equations (5) and (6) When the condition of is satisfied, it means that Ws is extremely smaller than Wr.

When Ws is extremely smaller than Wr from the line width determination unit 134, the output control unit 133 stops the scan for the blur pattern and the sub pattern 1 in each direction obtained by the scan for the original binary image. Is read from each of the memories 107, 109, 111 and 113 and is output to the feature extraction unit 135. The above processing of the output control unit 133 is performed in view of the problems described below.

Combining the faint sub-pattern with the sub-pattern 1 obtained by scanning the original binary image may restore the important information that has been removed while adding a noise-like component of the stroke. is there. Therefore, in this embodiment, in order to remove the extrinsic stroke component, first, as described above, the minute segment removing unit 13 is first performed.
At 9, the fine segment in the blur pattern is removed,
Further, as a matter of course, when it is determined that all the segments are minute, only the sub pattern 1 obtained by scanning the original binary image is output to the feature extraction unit 135. Further, a line width determination unit 134 is provided, and even if the line width of the fading pattern does not reach a predetermined threshold, the fading pattern is determined to be extrinsic for recognition, and in such a case, the fading pattern is determined. Without performing a scan,
Only the sub pattern 1 is output to the feature extraction unit 135. By doing so, the extrinsic stroke component is removed from the sub-pattern, and it is possible to prevent erroneous reading and the like due to it.

The feature extraction unit 135 performs feature extraction based on the sub-pattern of the input original binary image or the four types of synthesized sub-patterns. Before performing this feature extraction,
For the character pattern in the circumscribing frame of the pattern register 103, for example, nonlinearity is made in N × M partial regions in a grid pattern in the vertical and horizontal directions so that the number of black pixels in each divided region is the same. Split it. Next, for each of the four types of sub-patterns, the number of black pixels of the sub-pattern in the divided partial area is counted, and the number of black pixels is normalized by the size of the character pattern to thereby distribute the character lines in each direction. An N × M × 4 dimensional feature matrix that reflects the state is extracted and output to the identification unit 136.

The identification unit 136 collates the inputted characteristic matrix with the characteristic matrix of a plurality of standard characters stored in the dictionary memory 137 in advance, judges from the collation result, and finally makes one The candidate category is output as the recognition result 138 of the input character pattern. The above is the first embodiment of the character recognition device according to the present invention.

Here, the first embodiment was an extremely effective method when the local line widths of strokes forming characters and figures can be classified into two. But the usual simple letter is 2
It can be considered that the scanning can be almost reflected in the sub-pattern even by scanning with different line widths, but in the case of complicated figures and Chinese characters composed of strokes of three or more types of stroke, strokes that cannot be captured by two-step scanning There may be ingredients. The second embodiment is capable of coping with such a problem, and the second embodiment generalizes this, and the second embodiment is the N-stage ( N
Scanning of ≧ 2) is possible. The second embodiment will be described below.

FIG. 2 is a block diagram showing a second embodiment according to the present invention. Here, 201 is an optical signal input, 202 is a photoelectric conversion unit, 203 is a pattern register, 204 is a circumscribing frame detection unit, 205 is a register, 206 is a line width calculation unit, and 207.
Is a horizontal scanning unit, 208 is a horizontal pattern memory, 209
Is a vertical scanning unit, 210 is a vertical pattern memory, 211
Is a right diagonal scanning unit, 212 is a right diagonal pattern memory, 2
13 is a left oblique direction scanning unit, 214 is a left oblique pattern memory, 215 is a blurred pattern extraction unit, 216 is a minute segment removal unit, 217 is a line width determination unit, 218 is a horizontal pattern composition unit, 219 is a horizontal composition pattern memory, 220 is a vertical pattern composition unit, 221 is a vertical pattern pattern memory, 222
Is a right diagonal pattern synthesis unit, 223 is a right diagonal synthesis pattern memory, 224 is a left diagonal pattern synthesis unit, 225 is a left diagonal synthesis pattern memory, 226 is a loop counter, 227 is an output control unit, 228 is a feature extraction unit, 229 Is an identification unit, 230
Is a dictionary memory and 231 is a recognition result.

Here, differences from the first embodiment will be mainly described. First, 201, 202, 203, 20
In No. 4, according to the first embodiment, only the data within the circumscribing frame of the binary image of the pattern register 203 is transferred to the register 205. As will be described later, not only the binary data of the character pattern but also the blurred pattern are sequentially overwritten in the register 205. The line width calculation unit 206 uses this register 2
The line width is calculated for the data in 05. Since the binary data of the character pattern is currently stored, the average line width of the character pattern is calculated. The method of Example 1 is also applied to the calculation of the line width.

Next, as in the first embodiment, this register 2
For the binary data in 05, the horizontal scanning section 207,
The vertical scanning unit 209, the right diagonal scanning unit 211, and the left diagonal scanning unit 213 scan horizontally, vertically, diagonally to the right, and diagonally to the left, respectively, and the subpatterns are extracted with the line width as a threshold. Horizontal sub-pattern memory 208,
The data is stored in the vertical sub pattern memory 210, the right diagonal sub pattern memory 212, and the left diagonal sub pattern memory 214.

Next, the blur pattern extracting unit 215 extracts the blur pattern from the binary data of the character pattern of the register 205 and the sub patterns stored in the memories 208, 210, 212 and 214 and transfers it to the register 205. To do. At this time, the extraction of the faint pattern is performed as shown in FIG.
This method is used as the fading pattern 1 for the sake of convenience. Then, the minute segment removing unit 216 removes the minute segment of the blurred pattern 1, and after checking the number of remaining segments, the line width calculating unit 206 calculates the line width of the blurred pattern 1 and further determines the line width. The unit 217 determines whether to scan the fading pattern 1 based on the line width value. However, the minute segment removal unit 216 or the line width determination unit 21
The determination of 7 applies mutatis mutandis to Example 1. If it is determined that the blur pattern 1 does not need to be scanned, the sub patterns stored in the memories 208, 210, 212, and 214 are output to the feature extraction unit 228 via the output control unit 227.
When it is determined that scanning is necessary, scanning is performed to the horizontal composition sub-pattern memory 219, the vertical composition sub-pattern memory 221, the right diagonal composition sub-pattern memory 223, and the left diagonal composition sub-pattern memory 225, respectively. To be done. Next, similarly to the first embodiment, the scanning patterns 207, 209, 2 in each direction are applied to the blur pattern 1 in the register 205.
11 and 213, the sub-pattern of the blurred pattern 1, that is, the blurred sub pattern 1, is extracted based on the line width of the blurred pattern 1, and the memories 208 and 2 are respectively extracted.
10, 212, and 214 are stored.

Next, the horizontal sub-pattern combining section 218, the vertical sub-pattern combining section 220, and the right diagonal sub-pattern combining section 22.
2. In the left diagonal sub-pattern combining unit 224, the memory 2
08, 210, 212, 214 and the sub-pattern 1 stored in the memories 219, 221, 223, 225 are combined, and the combined sub-pattern 1 is output to the memories 219, 221, 223, 225 again. To be done.

The synthetic sub-pattern 1 is the same as that synthesized as a result of the two sub-pattern extractions in the first embodiment, but in the second embodiment, the blurring pattern extraction unit 215 further registers the register 205 at the present time. The blur pattern 1 and the memories 208, 210, 212 stored in
By using the faint sub-pattern 1 stored in 214, 2
The stroke component which is not detected by the second scanning is extracted and is set as the blur pattern 2 in the register 205.
To store. Here, in FIG. 3, the memories 107, 1
09, 111, and 113 are the memories 208,
Character pattern memory 3 corresponding to 210, 212, and 214
04 corresponds to the register 205.

Next, with respect to the blur pattern 2, the blur sub pattern 2 is obtained by scanning with the line width of the blur pattern 2 as a threshold value, and the memories 219, 221, 223, 225 in the combining units 218, 220, 222, 224. The composition with the composition sub-pattern 1 stored in
It is output to 19, 221, 223, and 225 as a combined sub-pattern 2. In exactly the same manner, for the blur pattern K, the blur sub-pattern K is obtained by scanning with the line width of the blur pattern K as a threshold, and the combining units 218, 220, 22 are obtained.
2, 224, memories 219, 221, 223,
The composite sub-pattern K-1 stored in 225 is combined, and the composite sub-pattern K-1 is output to the memories 219, 221, 223 and 225 as the composite sub-pattern K again.

The loop counter 226 counts the number of times K of sub-pattern combination, and when K reaches a predetermined threshold value M, it notifies the output control unit 227 of this. At that time,
In the output control unit 227, the memories 219, 221, 22
The combined sub-pattern M stored in Nos. 3 and 225 is transferred to the feature extraction unit 228. Even if the number of times K of synthesis does not reach M, if the minute segment removal unit 216 or the line width determination unit 217 determines that it is not necessary to scan the fading pattern K, the combination sub-pattern K at that time is a feature. It is transferred to the extraction unit 228.

Hereinafter, the feature extracting section 228, the identifying section 229,
Since the dictionary memory 230 and the recognition result 231 are all the same as those in the first embodiment, the description thereof is omitted here.

As described above, according to the second embodiment, by scanning M times, it is possible to create a sub-pattern that reflects M kinds of stroke components having different line widths. Highly accurate recognition performance can be stably maintained even for complicated Chinese characters and figures. Further, the first embodiment is equivalent to the case where M = 1 in the second embodiment, and corresponds to the special case of the second embodiment.

The first and second embodiments are not limited to the above-mentioned examples. For example, the blur pattern extracting means in the blur pattern extracting unit 114 or 215 is not limited to the method shown in FIG.
The same result is obtained by combining the ND, NAND, and NOT circuits, and by counting the original binary image and the black pixels of four types of sub-patterns for each pixel and extracting the ones with the count number of 1. There are several possible output methods, but all of them belong to the present invention as long as the blur pattern defined in this embodiment can be extracted.

Further, as a method for removing an extrinsic stroke component, removal of a minute segment, determination by a line width, etc. were prepared. However, these conditional expressions and threshold values are set within the scope of the present invention. It can be changed arbitrarily. The line width determination unit 13
4 and the minute segment removing unit 139 determines whether or not the portion left as the blurred pattern represents the essential characteristics of the character by using the characteristics such as the size and line width of the segment of the blurred pattern. Since both have the same function, either one may be implemented. Further, any feature other than the above feature, such as the distribution state of the segment or the black pixel density, can be used as long as it can obtain the same effect as the above determination, and belongs to the present invention. Also, if it is known in advance that the characters to be input are of high quality and the local line width is constant, the blur pattern is not extracted, and conversely, the character quality is low and there are many blur characters. If it is known in advance, an embodiment in which the faint pattern is extracted and combined constantly or while performing the determination also belongs to the present invention.

Further, the configuration of the pattern register and each memory, the line width calculation method, the feature matrix extraction method, the circumscribing frame division method, etc. can be appropriately changed within the scope of the present invention. Further, in the block diagram of FIG. 1, the processing and operation assigned to each component, the flow of input / output signals, the number of installations, the position, and other conditions can be arbitrarily changed.

[0051]

As described above in detail, according to the present invention, the input character pattern is converted into a binary image, and the line width of the binary image in the circumscribing frame of the binary image is calculated. The binary image in the circumscribing frame is scanned horizontally, vertically, diagonally to the right, and diagonally to the left to extract four types of sub-patterns that reflect the stroke distribution state based on the line width. Using the binary image and the four types of sub-patterns, a blur pattern composed of pixel portions not extracted as a sub-pattern is detected, and the line width of this blur pattern is calculated. Further, with respect to the blur pattern, using four threshold values set on the basis of the line width, scanning is performed in the horizontal, vertical, right diagonal, and left diagonal directions, and there are four types that reflect the distribution state of the detected strokes. A sub-pattern is extracted, a composite sub-pattern is created by combining the sub-pattern and the blur sub-pattern for each type, a feature matrix is extracted based on the composite sub-pattern, and the feature matrix and the dictionary are collated. As a result, the recognition result is output.Therefore, some strokes that make up the character pattern, which play an essential role in recognition, are compared with the local line width with other portions. Even when it gets smaller,
It is extracted as a part of the sub-pattern, and therefore poor quality character patterns and various scales with large differences in local line widths.
It is possible to realize a character recognition device capable of stably maintaining high-precision recognition performance even for complicated Kanji characters and figures composed of strokes of a ball.

[Brief description of drawings]

FIG. 1 is a block diagram showing a first embodiment of a character recognition device according to the present invention.

FIG. 2 is a block diagram showing a second embodiment of the character recognition device according to the present invention.

FIG. 3 is a block diagram showing a configuration of a blur pattern extracting unit.

FIG. 4 is an example of a pattern in which a blurred portion exists.

FIG. 5 is an example of a pattern in which there is a portion that is not extracted as a sub-pattern due to crushing.

FIG. 6 is a diagram showing an example of an application example of the present invention.

[Explanation of symbols]

 101 optical signal 102 photoelectric conversion unit 103 pattern register 104 circumscribing frame detection unit 105 character pattern line width calculation unit 106 horizontal direction scanning unit 107 horizontal sub pattern 1 memory 108 vertical direction scanning unit 109 vertical sub pattern 1 memory 110 right diagonal direction scanning unit 111 right Diagonal sub-pattern 1 memory 112 Left diagonal sub-pattern scanning unit 113 Left diagonal sub-pattern 1 memory 114 Blurred pattern extraction unit 115 Blurred pattern memory 116 Blurred pattern line width calculation unit 117 Horizontal scanning unit 118 Horizontal blur sub-pattern memory 119 Vertical scanning unit 120 Vertical Faint sub-pattern memory 121 Right diagonal scan section 122 Right diagonal sub-pattern memory 123 Left diagonal scan section 124 Left diagonal sub-pattern memory 125 Horizontal sub-pattern combining section 126 Horizontal sub-pattern 2 Memory 127 Vertical sub-pattern combining unit 128 Vertical sub-pattern 2 memory 129 Right oblique sub-pattern combining unit 130 Right oblique sub-pattern 2 memory 131 Left oblique sub-pattern combining unit 132 Left oblique sub-pattern 2 memory 133 Output control unit 134 Line width determination unit 135 Feature extraction unit 136 Identification Part 137 dictionary memory 138 recognition result 139 minute segment removal part

Claims (1)

[Claims] 1. A photoelectric conversion unit for optically scanning a character pattern written on a form or the like to convert it into a binary image which is a quantized electric signal, and a character pattern converted into the binary image. A pattern register for storing, a circumscribing frame detecting unit for detecting a circumscribing frame of the character pattern in the pattern register, a character pattern line width calculating unit for calculating a line width of the character pattern in the circumscribing frame of the pattern register, For the character pattern in the circumscribed frame of the pattern register,
When the number of consecutive black pixels obtained by scanning horizontally, vertically, diagonally to the right, and diagonally to the left exceeds a threshold value determined based on the line width, it is detected as a stroke, and these strokes are detected. -A sub-pattern extraction unit that extracts four types of sub-patterns representing the distribution of black and white in each direction, and a binary image in the circumscribed frame of the pattern register and the four types of sub-patterns among the black pixels that form a character pattern. A blur pattern extracting unit that extracts a black pixel that does not belong to any of the four types of sub patterns as a blur pattern, and a fine segment removing unit that removes a fine segment from each of the independent segments that form the blur pattern. It is necessary to extract the sub-pattern for the faint pattern line width calculation unit that calculates the faint pattern line width and the faint pattern from which minute segments have been removed. If it is determined that the blur pattern sub-pattern that represents the stroke distribution of the blur pattern based on the line width of the blur pattern is extracted in the same manner as above, four types of blur sub patterns are extracted in each scanning direction; A control unit that outputs to the feature extraction unit one of the sub-patterns of the combined sub-pattern and the blurred sub-pattern for each scanning direction, and a feature extraction unit that extracts the feature matrix of the sub-pattern or the combined sub-pattern, A character recognition device comprising: an identification unit that outputs a recognition result based on a result obtained by comparing the feature matrix with a dictionary matrix prepared in advance.