JP5176763B2 - Low quality character identification method and apparatus - Google Patents

Description

  BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method and an apparatus for identifying a degraded character, and in particular, character identification for identifying a print type of a large size class (also referred to as a super multi class), for example, Chinese, Japanese, or Korean. The present invention relates to a method and an apparatus.

  In recent years, research on character identification has progressed greatly, and a character identification method for accurately identifying a character image with a clean background has already been developed. However, no character identification method has yet been developed that can provide satisfactory results for low quality characters with a complex background and reduced image quality.

  The following two causes are conceivable as the character identification difficulty and the application bottleneck. The first is character identification when there is a complicated background. In such a case, it is necessary to detect and extract characters from the complex background of the image by an automatic character detection and character extraction system. The second is a problem of identifying low-quality characters when the image quality is low. In such a case, there are situations where the handwriting is blurred in the character image, the strokes constituting one stroke of the character are stuck together, the strokes are cut off, or the resolution is lowered, and an effective solution is available. It is necessary. As a prior art, two types of solutions are provided.

  The first solution is an identification method based on stroke edge features. In this identification method, first, a character to be identified is binarized, and then normalization of the character is performed by a non-linear algorithm. Finally, character stroke characteristics are extracted and character identification is performed by statistical classification.

  FIG. 4 shows a statistical classification procedure in the character identification method. The statistical classification procedure based on stroke edge features includes two steps: coarse classification and reclassification. In such an identification method including rough classification and reclassification, the reclassification algorithm is complicated, so that the calculation speed is slow. Therefore, the classification efficiency is improved by performing coarse classification and greatly narrowing down the identification candidates of the reclassification algorithm.

The details of the algorithm of this method can be referred to the following documents.
"Improvement of Handwritten Japanese Character Recognition Using Weighted Direction Code Histogram", F. Kimura, T. Wakabayashi, S. Tsuruoka, Y. Miyake, Pattern Recognition, v.30, n.8, pp.1329-1337, 1997

  The second solution is an identification method based on the image pattern feature of the extracted character in the gradation character image. In this identification method, first, gradation normalization is performed by removing the background of the character to be identified. Then, normalization of characters is performed by a linear algorithm. Finally, the image pattern features of the characters are extracted, and the characters are identified by statistical classification.

  FIG. 4 described above can also be applied to the statistical classification procedure of the identification method based on the image pattern feature, and the statistical classification procedure based on the image pattern feature also includes two steps of coarse classification and reclassification.

The details of the algorithm of this method can be referred to the following documents.
"Camera Based Degraded Text Recognition Using Grayscale Feature", Jun Sun, Yoshinobu Hotta, Yutaka Katsuyama, Satoshi Naoi, 8th International Conference on Document Analysis and Recognition (ICDAR2005), August 2005, Seoul Korea, pp.182-186
"A Grayscale Image Based Character Recognition Algorithm to Low Quality and Low Resolution Images", Wang XW, et al., Document Recognition and Retrieval VIII, Electronic Imaging 2001 [C]. San Jose, IS & T / SPIE, 2001

  However, each of the above two solutions has advantages and disadvantages. That is, as shown in FIG. 1, the identification method based on the stroke edge feature shows good adaptability to the deformation of the character stroke due to the change of the font style or the like. On the other hand, the identification method that extracts the image pattern feature of the character from the gradation character image shows good adaptability to the image degradation caused by various noises.

  In the identification method based on the stroke edge feature, since the stroke edge feature is usually extracted from the binary image, if the quality of the binary image is not good, the estimation of the stroke edge is not performed accurately, and the feature identification effect is obtained. There is a problem that decreases. The effect of binarization greatly depends on the degree of image degradation as well as depending on the algorithm. There are a plurality of methods for expressing the degree of image degradation, such as the degree of edge blurring and the resolution. As for a character image, the size of the character image can be cited as a more direct image deterioration index. In the case of a binary character image, the size of the image is defined by the length and width of a rectangle circumscribing the binarized image. As shown in FIG. 3, the result of binarizing images of different sizes is different, and the loss of pixels corresponding to the stroke of a character increases as the image size is reduced. In addition, when the amount of loss of pixels corresponding to a character stroke reaches a certain level, an identification engine that operates based on the stroke edge feature cannot obtain an accurate identification result.

  In addition, the identification method for extracting the image pattern feature from the gradation character image does not have a binarization procedure, so that the stroke information is kept to the maximum, and for a small size image. It shows a good discrimination effect. However, since the image pattern features of characters are sensitive to stroke deformation due to differences in font styles as shown in FIG. 2, when the stroke is deformed to some extent, the identification engine that extracts image pattern features from a gradation character image is An accurate identification result cannot be obtained.

  Thus, it can be seen that the above two solutions have different advantages and at the same time have different problems. In order to combine the advantages of the above two methods to obtain a higher performance identification method, a method for identifying low quality characters by selecting one of the above two methods according to different features is disclosed. .

JP-A-11-66240 JP 2000-82113 A US Pat. No. 4,551,851 "Robust Chinese Character Recognition by Selection of Binary-Based and Grayscale-Based Classifier", Yoshinobu Hotta, Jun Sun, Yutaka Katsuyama, Satoshi Naoi, Document Analysis System 2006, 553−563

  However, as shown in FIG. 5, in Patent Documents 1 and 2 and Non-Patent Document 1, an appropriate identification method is selected by estimating the cause and degree of deterioration, and the identification result by the selected identification method is finalized. Output as a result. Such a method has two drawbacks. First, in practice, the cause and degree of character deterioration are very complex, and the reliability of algorithms that estimate the cause and degree of character deterioration has been a problem in the past. Reliability is also a question. Second, even if the cause and degree of character degradation is correctly estimated, there is no guarantee that one of the above two solutions is optimal and the other is completely unsuitable; The method is only more suitable than the other method. For this reason, the methods disclosed in Patent Documents 1 and 2 and Non-Patent Document 1 are not sufficient, and it is considered that they do not satisfy the demand for practical use.

  In addition, as shown in FIG. 5, the method described in Patent Document 3 uses a discrimination mechanism to group input characters, and based on the grouped result, the input characters are divided into several sub-classifiers (Sub- classifier) for identification. Each sub-classifier can identify only one group of characters. The disadvantage of this method is that the first discriminator cannot achieve effective grouping for low-quality characters, and if the grouping is incorrect, it will cause direct misidentification.

  The present invention has been made to solve the above-described problems of the prior art, and is effective in an identification method based on stroke edge features and an identification method that extracts the entire image pattern features from a gradation character image. By combining with, it prevents the deterioration of the recognition rate due to the character deterioration due to the difference in font style and the image deterioration due to the deterioration of image quality, and improves the recognition performance of low quality characters in large size class character recognition The purpose is to let you.

  In order to achieve the above object, the low quality character identification method corresponds to the stroke edge feature and the image pattern feature by identifying the character to be identified based on the stroke edge feature and the image pattern feature, respectively. An identification step for obtaining two candidate character code groups, a merge step for determining a merge candidate character code group by merging candidate character codes in the two candidate character code groups, and a merge candidate character code group On the other hand, a calculation step of acquiring two types of identification distances by performing calculations related to each of the stroke edge feature and the image pattern feature, and a determination step of determining an optimum candidate character code based on the two types of identification distances. ,including.

  According to this identification method, the range of candidate character codes can be effectively expanded by an organic combination of the two identification methods, so by outputting a merge candidate character code group as a result of character identification, The accuracy of the candidate character code can be effectively improved. Furthermore, the convenience of manual identification of low quality characters and the accuracy of automatic identification can be further improved.

  Further, the low quality character identification method is the above method, wherein the merging step extracts all or part of the candidate character codes from each of the two candidate character code groups, and the all or part of the candidate characters. A merge candidate character code group is determined by taking a union of codes.

  In addition, the low quality character identifying method is the above method, wherein the merging step extracts all or part of the candidate character codes by determining the number of candidate character codes.

  In the above-described method, the number of candidate character codes extracted from each candidate character code group is directly proportional to the degree of deterioration of the character image.

  Further, the low quality character identification method is the above method, wherein the number of candidate character codes extracted from each candidate character code group is the first candidate character code and the second candidate character code in the candidate character code group. When the difference in identification distance between the Kth (K is a natural number) and the K + 1th candidate character code is less than a predetermined number times the reference, K candidates from the candidate character code group It is determined by the algorithm that extracts the character code.

Further, in the above method, the number of candidate character codes extracted from each candidate character code group is the same as that of the low quality character when the candidate character code group includes N candidate character codes. When the larger one of the number of pixels and the number of pixels of the width is less than n pixels, the number of candidate character codes to be extracted from the candidate character code group is N, and the number of pixels and the width of the height of the character image If the larger of the number of pixels is greater than m pixels, the number of candidate character codes to be extracted from the candidate character code group is N ′, and the number of pixels of the height of the character image and the number of pixels of the width are When the larger one is n pixels or more and m pixels or less, the number of candidate character codes to be extracted from the candidate character code group is determined.
(Where S is the larger number of pixels in the height and width of the character image, and m, n, N, and N ′ can be set by the user, And m> n).

The low quality character identification method is the above method, wherein the number of candidate character codes extracted from each candidate character code group is such that the candidate character code group includes N candidate character codes and the merge candidate character When the minimum number of candidate character codes to be extracted as candidate character codes belonging to the code group is N ′, the area of the region between the pixel value indicating the stroke and the pixel value indicating the background in the histogram of the character image; When the ratio of the total area of the histogram to R is R, the number of candidate character codes to be extracted from the candidate character code group is expressed as N−R × (N−N ′)
(Where 0 ≦ R ≦ 1, N and N ′ can be set by the user).

  The low quality character identification method is the above method, wherein the determining step determines the optimum candidate character code according to the accuracy of the merge candidate character code group obtained by integrating two types of identification distances. To do.

The low-quality character identification method is the above-described method, wherein the accuracy of the merge candidate character code group is obtained by a total reliability (C (I)), and the total reliability (C (I)) is: Sum of reliability (C _l (I)) based on stroke edge features and reliability (C _g (I)) based on image pattern features,
C (I) = C _l (I) + C _g (I)
It is.

  Further, the low-quality character identification device extracts a stroke edge feature of a character to be identified, and performs a discrimination based on the stroke edge feature, thereby obtaining a first candidate character code group. Image pattern feature processing means for extracting a second candidate character code group by extracting an image pattern feature of a character to be identified, and performing identification based on the image pattern feature, and the first candidate character Merging means for generating a merge candidate character code group by merging the code group and the second candidate character code group, and identification of each candidate character code of the merge candidate character code group based on a stroke edge feature Stroke edge feature reprocessing means for calculating distance, based on image pattern feature The image pattern feature reprocessing means for calculating the identification distance of each candidate character code of the merge candidate character code group, and the two types of identification distances calculated by the stroke edge feature reprocessing means and the image pattern feature reprocessing means And integrating means for acquiring the overall reliability of each candidate character code, and output means for outputting the optimum candidate character code from the overall reliability.

  According to the low quality character identification method and apparatus disclosed in the present specification, an identification effect superior to that of the prior art can be obtained. Using the discrimination test for 4238 Chinese characters, the method proposed by the present invention was compared with the case using only the method based on the stroke edge direction feature and the case using only the method based on the image pattern feature. When the image size is 8 × 8, the identification rate of the method based on the stroke edge direction feature is 58.80%, and the identification rate of the method based on the image pattern feature is 79.90%. The proposed method has an identification rate of 88.15%. Therefore, according to the present invention, it can be seen that the accuracy of the identification of the low-quality character image is greatly improved.

  Hereinafter, a detailed embodiment of a low quality character identification method and apparatus according to the present invention will be described with reference to the drawings.

  As shown in FIG. 6, the low-quality character identification method according to the embodiment performs identification based on the stroke edge feature and the image pattern feature for each character to be identified, and the stroke edge feature and the image pattern. Step S1 for obtaining two candidate character code groups based on characteristics, Step S2 for merging candidate character codes in the two candidate character code groups to determine a merge candidate character code group, and the merge candidate character Step S3 for obtaining two types of identification distances by performing an operation based on the stroke edge feature and the image pattern feature for each of the code groups, and an optimal candidate based on the above two types of identification distances And step S4 for determining a character code.

  According to this identification method, since the range of candidate character codes can be effectively expanded by organically combining the two identification methods, a merge candidate character code group can be output as a result of character identification. The accuracy of the candidate character code can be effectively improved. Hereinafter, each example will be described in detail.

  With reference to FIG. 7, FIG. 8, and FIG. 9, the details of the first embodiment of the low-quality character identifying method will be described in the order of steps. FIG. 7 is a flowchart showing a low quality character identification method according to the first embodiment. FIG. 8 is a flowchart showing a detailed procedure of step S1 in the low quality character identification method according to the first embodiment. FIG. 9 is a flowchart illustrating the detailed procedure of step S2 in the low quality character identification method according to the first embodiment.

  The low quality character identification method according to the first embodiment includes steps S1 and S2. Hereinafter, with reference to the drawings, the first embodiment will be described in detail using an identification method based on coarse classification as an example.

  In step S1, the stroke edge feature and the image pattern feature of the character to be identified are extracted, and rough classification is performed based on the two features, respectively, so that two coarse classification candidates based on the stroke edge feature and the image pattern feature are obtained. Get the character code group. Step S1 has two sub-steps S11 and S12.

  In step S11, a stroke edge feature and an image pattern feature are extracted for each character to be identified.

  As shown in FIG. 8, step S11 is performed in two parts. Specifically, for one character image to be identified, a stroke edge feature and an image are respectively obtained by using an identification method based on a stroke edge feature and an identification method for extracting an image pattern feature of a character from a gradation character image. Extract pattern features. The extraction of the stroke edge feature and the image pattern feature can also be performed using the methods described in the background art of this specification. In step S12, each of the stroke edge feature and the image pattern feature is roughly classified, and a coarse classification candidate character code group N1 based on the stroke edge feature and a coarse classification candidate character code group N2 based on the image pattern feature are obtained.

  As shown in FIGS. 7 and 8, this step S12 is also performed in two parts. Specifically, the rough classification for the input character image is performed for the stroke edge feature extracted in step S11, and the rough classification for the input character image is performed for the image pattern feature extracted in step S11. The purpose of roughly classifying these features is to output a plurality of candidate character codes having a high possibility by performing preliminary identification on the input character image.

  The rough classification procedure can be implemented by a plurality of algorithms, and all the algorithms listed in the background art of this specification can be applied to this embodiment. For example, in this embodiment, after calculating the average value of features of each character type included in the training sample, the distance between the feature obtained from the input character and the average value of the feature of the character type is calculated. Calculated as the identification distance. Since the identification distance represents the distance between the candidate character code and the input character, the accuracy of the candidate character code can be estimated from the data of the identification distance. Therefore, an algorithm is used in which the candidate character codes are sorted in ascending order of the identification distance, and a predetermined number of candidate character codes are selected from the head and output as the result of rough classification.

  Thus, in this embodiment, one coarse classification candidate character code group N1 based on the stroke edge feature and another coarse classification candidate character code group N2 based on the image pattern feature are output. However, the candidate character codes in the two coarse classification candidate character code groups N1 and N2 are sorted in ascending order of the identification distance.

  In step S2, the two coarse classification candidate character code groups N1 and N2 are merged to generate a merge candidate character code group M.

  Since the stroke edge feature and the image pattern feature are suitable for different low quality characters, the stroke edge feature coarse classification candidate character code group N1 and the image pattern feature coarse classification candidate character code output in step S12 shown in FIG. It is very different from group N2. Therefore, it is necessary to output one merge candidate character code group M by combining the two rough classification candidate character code groups. This merge candidate character code group is output as a character identification result of the identification method according to the present embodiment. This merging procedure can be performed by a plurality of algorithms such as a union set, a product set, and the like. In this embodiment, the union of two candidate character code groups is taken.

  As shown in FIG. 9, step S2 is also divided into two sub-steps S21 and S22.

  In step S21, the number of candidate character codes to be extracted from the candidate character code group is determined based on two types of data, ie, the increase in the identification distance of the coarse classification candidate character codes or the pixel value of the character image. The merge candidate character code group M is obtained by taking the union of the candidate character code groups N1 ′ and N2 ′. However, the number of candidate character codes extracted from the coarse classification candidate character code group is directly proportional to the degree of deterioration of the character image. That is, as the degree of deterioration of the character image becomes severe, the number of candidate character codes extracted from the coarse classification candidate character code group increases.

  In the two coarse classification candidate character code groups N1 and N2 shown in step S12, since the candidate character codes are sorted in ascending order of the identification distance, the two coarse classification candidate character code groups N1 are determined after the number is determined. And candidate character codes are extracted from N2 in ascending order of the identification distance.

  Hereinafter, three different calculation methods for determining the number of candidate character codes extracted from the candidate character code group in step S21 will be described.

  In the first method, the number of candidate character codes to be extracted from the coarse classification candidate character code groups is determined based on the degree of increase in the identification distance between the two coarse classification candidate character code groups N1 and N2. Specifically, in this algorithm, the Kth (K is a natural number) and the K + 1th, based on the difference in identification distance between the first candidate character code and the second candidate character code in each coarse classification candidate character code group K candidate character codes are extracted from this coarse classification candidate character code group when the difference in identification distance between the candidate character codes is less than a predetermined number of times the reference.

  Therefore, after determining the number of candidate character codes of the coarse classification candidate character code group using this method, two new coarse classification candidate character code groups N1 'and N2' are generated.

  In the second method, the number of candidate character codes to be extracted from the coarse classification candidate character code group is determined by pixels of the character image. Specifically, in this algorithm, the number of candidate character codes in each coarse classification candidate character code group is determined by performing statistical processing on the size of character images or character image pixels. In the coarse classification, the ability to identify a character with a severe deterioration level is low. Therefore, when the character image pixel is very small or the character is very blurred, the correct character code corresponding to the input image is the coarse classification candidate character code group N1. Therefore, it is necessary to increase the number of candidate character codes extracted from each coarse classification candidate character code group.

Assuming that one candidate character code group includes N character codes and the number of candidate character codes in the coarse classification candidate character code group is at least 3 or more, coarse classification candidate characters are represented by character image pixels. As an example of the algorithm for determining the number of candidate character codes of the code group, the following can be cited. That is, when the larger one of the number of pixels of the height of the character image and the number of pixels of the width is less than 15 pixels, for example, N which is the largest candidate number is selected and the number of pixels of the height of the character image If the larger of the number of pixels and the number of pixels of the width is larger than 30 pixels, for example, the smallest candidate number 3 is selected and the larger of the number of pixels of the height of the character image and the number of pixels of the width is selected. Is 15 pixels or more and 30 pixels or less, the number of candidate character codes is
And

  However, S is the larger number of pixels of the height pixel number and the width pixel number of the binary image of the character, and the minimum number of candidate character codes of each candidate character code group (3 in the above example) ) May be set by the user, and thresholds for the number of pixels (15 and 30 in the above example) may also be set by the user. By using this method to determine the number of candidate character codes of the coarse classification candidate character code group, two new coarse classification candidate character code groups N1 'and N2' are generated.

In the third method, the number of candidate character codes to be extracted from the ancestor classification candidate character code group is determined based on the histogram of the character image. Specifically, assuming that one candidate character code group includes N character codes and the number of candidate character codes to be extracted from the coarse classification candidate character code group is at least three, An example of an algorithm for selecting the number of candidate character codes in the coarse classification candidate character code group by using a histogram is as follows. That is, first, a histogram of a character image (frequency at which each pixel value appears in the image) is obtained by statistical processing, and when the pixel value of the character stroke is smaller than the pixel value of the background, Find the first peak by searching in the direction of the pixel value. The pixel value corresponding to this peak is a typical pixel value of the stroke. In the histogram, the first peak is searched for by searching from the largest pixel value to the smallest pixel value. The pixel value corresponding to this peak is a typical pixel value of the background. The area of the region existing between the two peaks is small when the character image is clear, and is large when the character image is blurred. Therefore, the ratio of the area of the region existing between the two peaks and the total area of the histogram is R, and the number of candidate character codes is
N-R × (N-3)
Calculated by

  However, the range of R satisfies 0 ≦ R ≦ 1, the minimum number of candidate character codes to be extracted from the coarse classification candidate character code group (3 in the above example) may be set by the user, and the pixel range is also set by the user. May be set. By using this method to determine the number of candidate character codes of the coarse classification candidate character code group, two new coarse classification candidate character code groups N1 'and N2' are generated.

  Next, Step S22 is executed to merge the two new coarse classification candidate character code groups N1 'and N2' generated by the extraction. As described above, this merging procedure can be performed by a plurality of algorithms such as union, intersection, etc. In this embodiment, the two rough classification candidate character code groups N1 ′ extracted in step S21 are used. By taking the union of N2 ′ and N2 ′, the merge candidate character code group M that is the identification result of the first embodiment is obtained.

  Below, the effect by the procedure which identifies the low quality character which concerns on Example 1 is demonstrated in detail.

As shown in FIG. 10, [Character 1] which is a recognition target character by the identification method according to the first embodiment.
Will be identified. First,
[Character 2]
The stroke edge feature and the image pattern feature are respectively extracted from the character "." Thereafter, the rough classification candidate character code group N1 based on the stroke edge feature and the rough classification candidate character code group N2 based on the image pattern feature are acquired by performing rough classification on the stroke edge feature and the image pattern feature, respectively. However, the rough classification candidate character code group N1 based on the stroke edge feature is sorted in the order of the identification distance [character 3].
Etc. The coarse classification candidate character code group N2 based on the image pattern feature is sorted in the order of the identification distance [character 4].
Etc.

As can be seen from the above, the rough stroke candidate character code group N1 for the stroke edge feature and the coarse classification candidate character code group N2 for the image pattern feature are greatly different. Therefore, it is necessary to output one merge candidate character code group M by combining the two coarse classification candidate character code groups. This merge candidate character code group is output as a character identification result of this identification method. In this embodiment, the number of candidate character codes to be extracted from the candidate character code group is determined according to the degree of increase in the identification distance of each coarse classification candidate character code. By this step, two extracted rough classification candidate character code groups N1 ′ and N2 ′ are obtained. However, N1 ′ is sorted in the order of the identification distance [character 5].
N2 ′ is the sorted “receive, 采” sorted in order of identification distance.
The two candidate character codes are included.

If the union is taken for both of these, the merge candidate character code group M is
[Character 6]
Is obtained. As can be seen from this example, when the identification method based solely on the image pattern feature is used (N2 ′ above), an accurate identification result cannot be obtained. Since the rough classification candidate character code group obtained by the effective combination effectively expands the range of candidate character codes, the accuracy of identification can be improved and the object of the present invention can be achieved.

  As shown above, the merge candidate character code group M is an integration of the stroke edge feature coarse classification candidate character code group N1 and the image pattern feature coarse classification candidate character code group N2 sorted in order of identification distance. Therefore, the range of candidate character codes can be effectively expanded. In addition, because candidate character codes are extracted based on the degree of image degradation, the merge candidate character code group M can be output as a character identification result, and the accuracy of the candidate character codes can be effectively improved. The object of the present invention can be achieved. In Example 1, since the two candidate character code groups N1 and N2 are sorted by the identification distance, it is possible to appropriately extract candidate characters having a small identification distance. Therefore, in addition to achieving the object of the present invention, the first embodiment can improve the convenience of manual identification of low-quality characters and the accuracy of automatic identification.

  Embodiment 2 according to the present invention is an embodiment that further improves the convenience of manual identification of low-quality characters and the accuracy of automatic identification. Hereinafter, with reference to FIGS. 11 and 12, the details of the second embodiment of the low quality character identification method will be described in the order of steps. FIG. 11 is a flowchart illustrating a low quality character identification method according to the second embodiment. FIG. 12 is a flowchart illustrating a detailed procedure of step S5 in the low quality character identification method according to the second embodiment.

  The low quality character identification method according to the second embodiment includes six steps of step STEP1, step STEP2, step STEP3, step STEP4, step STEP5, and step STEP6.

  In step STEP1, a stroke edge feature and an image pattern feature are extracted for each character image to be identified. Step STEP1 is performed in two parts. Specifically, for each character image to be identified, a stroke edge feature is obtained by using an identification method based on the stroke edge feature and an identification method for extracting a character image pattern feature from a gradation character image, respectively. Extract image pattern features. The extraction of the stroke edge feature and the image pattern feature can also be performed by the method described in the background art of this specification. Since step STEP1 is the same as step S11 of the first embodiment, detailed description thereof is omitted.

  In step STEP2, by roughly classifying the stroke edge feature and the image pattern feature, respectively, the coarse classification candidate character code group N1 based on the stroke edge feature and the coarse classification candidate character code group N2 based on the image pattern feature are obtained. To do.

  This step STEP2 is also performed in two parts. Specifically, the input character image is roughly classified with respect to the stroke edge feature extracted at step STEP1, and the input character image is roughly classified with respect to the image pattern feature extracted at step STEP1. The purpose of roughly classifying these features is to output a plurality of candidate character codes having a high possibility by performing preliminary identification on the input character image. Since step STEP2 is the same as step S12 of the first embodiment, detailed description thereof is omitted.

  By step STEP2, coarse classification candidate character code group N1 based on the stroke edge feature and coarse classification candidate character code group N2 based on the image pattern feature are obtained. It is assumed that the candidate character codes in the two coarse classification candidate character code groups N1 and N2 are sorted in ascending order of the identification distance.

  In step STEP3, the two rough classification candidate character code groups N1 and N2 are merged to generate a merge candidate character code group M.

  Since the stroke edge feature and the image pattern feature are suitable for different low-quality characters, the stroke edge feature rough classification candidate character code group N1 and the image pattern feature coarse classification candidate character code group N2 output in step STEP2 are It is very different. Therefore, it is necessary to output one merge candidate character code group M by combining the two rough classification candidate character code groups. This merge candidate character code group is output as a character identification result of the identification method according to the present embodiment. This merging procedure can be performed by a plurality of kinds of algorithms such as a union set, a product set, and the like. In this embodiment, the union of two candidate character code groups is taken.

  Since step STEP3 is the same as step S2 of the first embodiment, detailed description thereof is omitted.

  In step STEP 3, the number of candidate character codes to be extracted from the candidate character code group is determined based on the two types of data, ie, the increase in the identification distance of the coarse classification candidate character codes or the pixel value of the character image. The merge candidate character code group M is obtained by taking the union of the candidate character code groups N1 ′ and N2 ′. However, the number of candidate character codes extracted from the coarse classification candidate character code group is directly proportional to the degree of deterioration of the character image. That is, as the degree of deterioration of the character image becomes severe, the number of candidate character codes extracted from the coarse classification candidate character code group increases.

  In the second embodiment, the three algorithms described in the first embodiment can be applied when determining the number of candidate character code extractions, and thus the description thereof is omitted here.

  Since the candidate character codes in the two coarse classification candidate character code groups N1 and N2 are sorted in ascending order of the identification distance, after the number is determined, the two coarse classification candidate character code groups N1 and N2 Candidate character codes are extracted from each in ascending order of identification distance.

  The merge candidate character code group M is obtained by taking the union of the two coarse classification candidate character code groups N1 'and N2' extracted by the above method. The merge candidate character code group M is a combination of the rough classification candidate character code group N1 of stroke edge features and the rough classification candidate character code group N2 of image pattern features sorted in order of identification distance. Can be effectively expanded. Further, since the candidate character code is extracted based on the degree of image degradation, the merge candidate character code group M can be output as a character identification result, and the accuracy of the candidate character code can be effectively improved. Can achieve the purpose.

  In order to further improve the convenience of manual identification of low-quality characters and the accuracy of automatic identification and to further confirm the accuracy of merge candidate character code groups, in this embodiment, merge candidate character code groups M Based on the above, the following steps are performed to perform reclassification using an identification method that improves the identification performance although the identification speed is slow.

  In step STEP4, two sets of identification distances M1 and M2 are obtained by performing reclassification on the merge candidate character code group M based on the stroke edge feature and the image pattern feature.

  Step STEP4 shown in FIG. 11 is performed in two parts. Specifically, as a first part, the identification distance of each candidate character code in the merge candidate character code group M is calculated based on the stroke edge feature to obtain a set M1 of identification distances based on the stroke edge feature. Further, as a second part, the identification distance of each candidate character code in the merge candidate character code group M is calculated based on the image pattern feature to obtain a set M2 of identification distances based on the image pattern feature.

  For this reason, after the reclassification procedure is completed, two identification distances correspond to each candidate character code in the merge candidate character code group M. Further, as described above, the smaller the identification distance, the higher the accuracy of the candidate character code.

  In step STEP5, the total reliability set C is generated by calculating the total reliability of each candidate character code in the merge candidate character code group M from the two sets of identification distances M1 and M2.

  As shown in FIG. 12, this step STEP5 uses the concept of reliability. The reliability is a degree of certainty that an estimation result based on a sample falls within a predetermined range indicating a correct answer. Confidence is usually expressed as a percentage (%). For example, the reliability of 95% indicates that the possibility that the estimation result by sampling coincides with the overall accuracy is 95%, and the possibility that it does not coincide is 5%.

  Since each candidate character code in the merge candidate character code group M has two identification distances, two sets of identification distances M1 and M2 are generated. In step STEP5, one total reliability is calculated for each candidate character code from the two sets of identification distances M1 and M2, and the set of the total reliability of the candidate character code is C.

  The total reliability C is an overall reliability because the reliability based on the stroke edge feature and the reliability based on the image pattern feature are comprehensively considered. This total reliability C is obtained by a plurality of methods such as addition and integration, for example. In this embodiment, the total reliability C is the sum of the reliability based on the stroke edge feature and the reliability based on the image pattern feature.

For example, the total reliability of the I-th candidate character code is given by the following formula C (I) = C _l (I) + C _g (I)
Is obtained. Where C _l (I) is the reliability of the I-th candidate character code calculated by the stroke edge feature, and C _g (I) is the reliability of the I-th candidate character code calculated by the image pattern feature. is there.

The two reliability levels C ₁ (I) and C _g (I) are calculated from a set of identification distances M1 based on stroke edge features and a set of discrimination distances M2 based on image pattern features, respectively. A plurality of methods in statistics can be applied to the calculation of the reliability.

In this embodiment, if the identification distance of the I-th candidate character code generated based on the stroke edge feature is M1 (I) by this identification method, the smaller the identification distance is, the more the input character image becomes the stroke edge. There is a high possibility that it corresponds to one of the candidate character codes based on the features. In this case, the reliability calculated based on the stroke edge feature is
It becomes.

Further, when the identification distance of the I-th candidate character code generated based on the image pattern feature by this identification method is M2 (I), the input character image is a candidate based on the image pattern feature as the identification distance is smaller. Most likely one of the character codes. In this case, the reliability calculated based on the image pattern feature is
It becomes.

As can be seen, the numerical range of the reliability is between 0 and 1, and the higher the reliability, the higher the accuracy of the identification result for the candidate character code by the above-described identification method.
Then, the total reliability C (I) of the I-th candidate character code is calculated by adding the two reliability C _l (I) and C _g (I) of the I-th candidate character code.

  In step STEP6, the merge candidate character code groups M are sorted in descending order of the comprehensive reliability set C, and the optimum candidate character code group M 'is output.

  Finally, as shown in FIG. 11, all candidate character codes are newly sorted in descending order of the overall reliability C (I), and the sorted candidate character code groups are output as final results. The candidate character code group sorted in descending order of the overall reliability is M ′.

  Below, the effect by the procedure which identifies the low quality character which concerns on Example 2 is demonstrated in detail.

As illustrated in FIG. 13, [Character 7], which is a character to be identified, by the identification method according to the second embodiment.
Will be identified. First,
[Character 8]
The stroke edge feature and the image pattern feature are respectively extracted from the character "." Thereafter, the rough classification candidate character code group N1 based on the stroke edge feature and the rough classification candidate character code group N2 based on the image pattern feature are acquired by performing rough classification on the stroke edge feature and the image pattern feature, respectively. However, the rough classification candidate character code group N1 based on the stroke edge feature is sorted in the order of the identification distance [character 9].
Etc. The coarse classification candidate character code group N2 based on the image pattern feature is sorted in the order of the identification distance [character 10].
Etc.

In this embodiment, the number of candidate character codes to be extracted from the candidate character code group is determined according to the degree of increase in the identification distance of each coarse classification candidate character code. By this step, two extracted rough classification candidate character code groups N1 ′ and N2 ′ are obtained. However, N1 ′ is sorted in order of identification distance [character 11]
N2 ′ is sorted in the order of the identification distances.
including. When the union is taken for both of these, the merge candidate character code group M is [character 12].
Is obtained.

In order to further improve the accuracy of the candidate character code, the above-described merge candidate character code group is reclassified based on the stroke edge feature and the image pattern feature, respectively, thereby providing two types of identification distances M1 and M2. Is obtained. According to the calculation method of the total reliability according to the present embodiment, first, the candidate character code group M merged based on the two types of identification distances M1 and M2 [character 13].
After calculating two reliability C _l (I) and C _g (I) of each candidate character code, add C _l (I) and C _g (I) corresponding to each candidate character code Then, the overall reliability C (I) of the candidate character code is acquired. [Character 14] which is the merge candidate character code group M
Are sorted in descending order of the value of the total reliability C (I), and are the optimum candidate character code group M ′ [character 15].
Is output.

  In the second embodiment, based on the candidate character code group M obtained in the first embodiment, two types of identification distances M1 and M2 of the candidate character code group M are calculated by reclassification, and the identification distance is converted into reliability. Then sort by total confidence. Therefore, in the second embodiment, sorting is performed so that the correct identification result is the first candidate character code in the optimum candidate character code group M ′. As can be seen, the classification distance is calculated by reclassification, converted into the overall reliability, and the candidate character codes are sorted by the overall reliability. The convenience of identification and the accuracy of automatic identification can be improved.

[Low-quality character recognition device]
Hereinafter, with reference to the drawings, details of the low-quality character identifying apparatus according to the above embodiment will be described.

  As shown in FIG. 14, in addition to the character input unit 0, the low-quality character identification device mainly includes a stroke edge feature processing unit 1, an image pattern feature processing unit 2, and a coarse classification result merging unit 3. .

  Further, the low-quality character identification device may include a first reclassification unit 4, a second reclassification unit 5, a reclassification result integration unit 6, and a result output unit 7.

  The stroke edge feature processing means 1 includes a stroke edge feature extraction module 11 and a stroke edge feature rough classification module 12, and the image pattern feature processing means 2 includes an image pattern feature extraction module 21 and an image pattern feature rough classification module 22. Including.

  In this low quality character identification device, after receiving the character image to be identified through the character input means 0, the stroke edge feature extraction module 11 performs an identification method based on the stroke edge feature on the input character image. To extract stroke edge features.

  The image pattern feature extraction module 21 extracts an image pattern feature from the input character image using an identification method for extracting a character image pattern feature from a gradation character image.

  Various algorithms can be used to extract the stroke edge feature and the image pattern feature, but all of these algorithms can be applied to this embodiment. Note that the algorithm described in the background art and examples of this specification can be used as an algorithm for extracting the stroke edge feature.

  The stroke edge feature extraction module 11 transmits the extracted stroke edge feature to the stroke edge feature rough classification module 12. After receiving the stroke edge feature, the stroke edge feature coarse classification module 12 performs rough classification on the input character image based on the stroke edge feature, thereby obtaining a rough classification candidate character code group N1 based on the stroke edge feature. The candidate character codes in the coarse classification candidate character code group N1 are sorted according to the identification distance.

  The image pattern feature extraction module 21 transmits the extracted image pattern features to the image pattern feature rough classification module 22. After receiving the image pattern feature, the image pattern feature coarse classification module 22 performs rough classification on the input character image based on the image pattern feature, thereby obtaining a rough classification candidate character code group N2 based on the image pattern feature. The candidate character codes in the coarse classification candidate character code group N2 are sorted according to the identification distance.

  Since the stroke edge feature and the image pattern feature have a complementary relationship when different low quality characters are handled, the coarse classification candidate character code group N1 based on the stroke edge feature and the coarse classification candidate character code group based on the image pattern feature It is very different from N2. Therefore, the coarse classification result merge means 3 combines two coarse classification candidate character code groups and outputs one merge coarse classification candidate character code group M.

  Before the merging procedure, the number of candidate character codes to be extracted from the candidate character code group is determined according to the increase in the identification distance of the coarse classification candidate character codes or the pixel value of the character image, and the two extracted The merge candidate character code group M is obtained by taking the union of the coarse classification candidate character code groups N1 ′ and N2 ′. However, the number of candidate character codes to be extracted from the determined rough classification candidate character code group is directly proportional to the degree of deterioration of the character image. That is, the number of candidate character codes extracted from the coarse classification candidate character code group increases as the deterioration of the character image becomes severe.

  The coarse classification result merging means 3 takes the union of two coarse classification candidate character codes obtained by extracting the determined number of candidate character codes from the respective candidate character code groups N1 and N2. The merge candidate character code group M is acquired.

  This merging procedure can be performed by a plurality of algorithms such as union, intersection, etc., and in this embodiment, a method of taking the union of two coarse classification candidate character code groups is adopted. .

  The coarse classification result merging means 3 transmits the merge candidate character code group M to the first reclassifying means 4 and the second reclassifying means 5, respectively.

  After the first reclassification unit 4 and the second reclassification unit 5 receive the merge candidate character code group M, the first reclassification unit 4 uses the stroke edge feature to each of the merge candidate character code groups M. By calculating the identification distance of the candidate character code, an identification distance set M1 based on the stroke edge feature is acquired, and the identification distance set M1 is transmitted to the reclassification result integrating means 6.

  The second reclassifying unit 5 obtains a set M2 of identification distances based on the image pattern features by calculating the identification distances of the candidate character codes in the merge candidate character code group M using the image pattern features. Then, the set M2 of identification distances is transmitted to the reclassification result integrating means 6.

  Through the processing by the first reclassification means 4 and the second reclassification means 5, each candidate character code in the merge candidate character code group corresponds to two identification distances. The reclassification result integration unit 6 calculates one total reliability for each candidate character code based on the two identification distances of each candidate character code. A set of total reliability of the merge candidate character code is C. The algorithm of the total reliability is as described in the above embodiment.

  Finally, the reclassification result integration unit 6 transmits the total reliability for each candidate character to the result output unit 7, and the result output unit 7 newly sorts all candidate character codes in descending order of the total reliability. The sorted candidate character code group M ′ is output as the final result.

  Although the details of the embodiments of the present invention have been described above, the above contents and specific embodiments are intended to show the practicality of the technology according to the present invention, and the claims according to the present invention It is not intended to limit the scope. It is obvious that those skilled in the art can implement various changes, equivalent conversions or improvements without departing from the spirit and principle of the present invention. The scope of the claims according to the present invention conforms to the scope of the appended claims.

  The following additional notes are disclosed with respect to the above embodiments.

(Appendix 1) An identification step of acquiring two candidate character code groups respectively corresponding to the stroke edge feature and the image pattern feature by identifying the character to be identified based on the stroke edge feature and the image pattern feature;
Merging to determine a merge candidate character code group by merging candidate character codes in the two candidate character code groups;
A calculation step for obtaining two types of identification distances by performing a calculation for each of the stroke edge feature and the image pattern feature for the merge candidate character code group;
A determination step of determining an optimal candidate character code based on the two types of identification distances;
A method for identifying low quality characters, comprising:

(Supplementary Note 2) The merge step includes
Extracting all or part of the candidate character codes from each of the two candidate character code groups, and determining a merge candidate character code group by taking a union of all or part of the candidate character codes. The method for identifying low-quality characters according to Supplementary Note 1, which is a feature.

(Supplementary Note 3) The merge step includes
The low-quality character identification method according to appendix 2, wherein all or part of the candidate character codes are extracted by determining the number of candidate character codes.

(Supplementary note 4) The low quality character identification method according to supplementary note 3, wherein the number of candidate character codes extracted from each candidate character code group is directly proportional to the degree of deterioration of the character image.

(Supplementary Note 5) The number of candidate character codes extracted from each candidate character code group is based on the difference in identification distance between the first candidate character code and the second candidate character code in the candidate character code group. (K is a natural number) If the difference in identification distance between the Kth and K + 1th candidate character code is less than a predetermined number times the reference, it is determined by an algorithm that extracts K candidate character codes from the candidate character code group. The identification method of the low quality character of the additional statement 3 characterized by the above-mentioned.

(Supplementary Note 6) The number of candidate character codes extracted from each candidate character code group is as follows. When the candidate character code group includes N candidate character codes,
If the larger of the number of pixels in the height of the character image and the number of pixels in the width is less than n pixels, the number of candidate character codes to be extracted from the candidate character code group is N,
If the larger of the number of pixels in the height of the character image and the number of pixels in the width is more than m pixels, the number of candidate character codes to be extracted from the candidate character code group is N ′,
When the larger of the number of pixels of the height of the character image and the number of pixels of the width is n pixels or more and m pixels or less, the number of candidate character codes to be extracted from the candidate character code group is
(Where S is the larger number of pixels in the height and width of the character image, and m, n, N, and N ′ can be set by the user, The identification method of low quality characters according to supplementary note 3, wherein the identification method is determined by an algorithm (where m> n).

(Supplementary Note 7) The number of candidate character codes extracted from each candidate character code group is extracted as a candidate character code that includes N candidate character codes and that belongs to the merge candidate character code group. When the minimum number of power candidate character codes is N ′,
Candidates to be extracted from the candidate character code group when the ratio of the area between the pixel value indicating the stroke and the pixel value indicating the background in the histogram of the character image and the area of the entire area of the histogram is R The number of character codes is N−R × (N−N ′)
(Note that 0 ≦ R ≦ 1, N and N ′ can be set by the user). 4. The method for identifying low-quality characters according to supplementary note 3, wherein:

(Appendix 8) The determination step includes
The low quality character identification method according to appendix 1, wherein the optimum candidate character code is determined according to the accuracy of the merge candidate character code group obtained by integrating two types of identification distances.

(Supplementary Note 9) The accuracy of the merge candidate character code group is obtained by the overall reliability (C (I)), and the overall reliability (C (I)) is the reliability based on the stroke edge feature (C _l (I)) and the reliability based on image pattern features (C _g (I)),
C (I) = C _l (I) + C _g (I)
The method for identifying low-quality characters according to supplementary note 8, wherein:

(Supplementary Note 10) Stroke edge feature processing means for extracting a stroke edge feature of a character to be identified and obtaining a first candidate character code group by performing identification based on the stroke edge feature;
Image pattern feature processing means for obtaining a second candidate character code group by extracting an image pattern feature of a character to be identified and performing identification based on the image pattern feature;
Merging means for generating a merge candidate character code group by merging the first candidate character code group and the second candidate character code group;
Stroke edge feature reprocessing means for calculating an identification distance of each candidate character code of the merge candidate character code group based on a stroke edge feature;
Image pattern feature reprocessing means for calculating an identification distance of each candidate character code of the merge candidate character code group based on an image pattern feature;
An integration unit that acquires the total reliability of each candidate character code by integrating two types of identification distances calculated by the stroke edge feature reprocessing unit and the image pattern feature reprocessing unit;
An output means for outputting the optimum candidate character code from the overall reliability;
An apparatus for identifying low-quality characters, comprising:

It is a figure which shows the complementary relationship of two identification methods with respect to a different deterioration degree. It is a figure which shows the difference of the character image by a font. It is a figure which shows the specific example of the character image of a different size, and a binary image. It is a flowchart which shows the process of the identification method based on a stroke edge feature and an image pattern feature. It is a flowchart which shows the process which selects one of two identification methods and identifies a low quality character. It is a flowchart which shows the identification method of the low quality character which concerns on an Example. 3 is a flowchart illustrating a method for identifying low-quality characters according to the first embodiment. 5 is a flowchart illustrating a detailed procedure of step S1 according to the first embodiment. It is a flowchart which shows the detailed procedure of step S2 which concerns on Example 1. FIG. It is a figure which shows the specific example of the procedure of the identification method of the low quality character which concerns on Example 1. FIG. 10 is a flowchart illustrating a method for identifying low-quality characters according to a second embodiment. It is a flowchart which shows the detailed procedure of step STEP5 which concerns on Example 2. FIG. It is a figure which shows the specific example of the procedure of the identification method of the low quality character which concerns on Example 2. FIG. It is a block diagram which shows the structure of the identification apparatus of the low quality character which concerns on an Example.

Explanation of symbols

0 Character input means 1 Stroke edge feature processing means 11 Stroke edge feature extraction module 12 Stroke edge feature rough classification module 2 Image pattern feature processing means 21 Image pattern feature extraction module 22 Image pattern feature coarse classification module 3 Coarse classification result merge means 4 One reclassification means 5 Second reclassification means 6 Reclassification result integration means 7 Result output means

Claims (10)

An identification step of acquiring two candidate character code groups respectively corresponding to the stroke edge feature and the image pattern feature by identifying the character to be identified based on the stroke edge feature and the image pattern feature;
Merging to determine a merge candidate character code group by merging candidate character codes in the two candidate character code groups;
A calculation step for obtaining two types of identification distances by performing a calculation for each of the stroke edge feature and the image pattern feature for the merge candidate character code group;
A determination step of determining an optimal candidate character code based on the two types of identification distances;
A method for identifying low quality characters, comprising: The merging step includes
Extracting all or part of the candidate character codes from each of the two candidate character code groups, and determining a merge candidate character code group by taking a union of all or part of the candidate character codes. The low quality character identification method according to claim 1, wherein: The merging step includes
The method for identifying low-quality characters according to claim 2, wherein all or part of the candidate character codes are extracted by determining the number of candidate character codes.   4. The low quality character identification method according to claim 3, wherein the number of candidate character codes extracted from each candidate character code group is directly proportional to the degree of deterioration of the character image.   The number of candidate character codes extracted from each candidate character code group is based on the difference in identification distance between the first candidate character code and the second candidate character code in the candidate character code group, and K (K is a natural number) ) When the difference between the identification distances of the (th) and (K + 1) th candidate character code is less than a predetermined number of times the reference, it is determined by an algorithm that extracts K candidate character codes from the candidate character code group. The method for identifying low-quality characters according to claim 3. The number of candidate character codes extracted from each candidate character code group is as follows when the candidate character code group includes N candidate character codes:
If the larger of the number of pixels in the height of the character image and the number of pixels in the width is less than n pixels, the number of candidate character codes to be extracted from the candidate character code group is N,
If the larger of the number of pixels in the height of the character image and the number of pixels in the width is more than m pixels, the number of candidate character codes to be extracted from the candidate character code group is N ′,
When the larger of the number of pixels of the height of the character image and the number of pixels of the width is n pixels or more and m pixels or less, the number of candidate character codes to be extracted from the candidate character code group is
(Where S is the larger number of pixels in the height and width of the character image, and m, n, N, and N ′ can be set by the user, 4. The method for identifying low-quality characters according to claim 3, wherein the determination is performed by an algorithm (where m> n). The number of candidate character codes extracted from each candidate character code group is such that the candidate character code group includes N candidate character codes and should be extracted as a candidate character code belonging to the merge candidate character code group. When the minimum number of N ′ is N ′,
Candidates to be extracted from the candidate character code group when the ratio of the area between the pixel value indicating the stroke and the pixel value indicating the background in the histogram of the character image and the area of the entire area of the histogram is R The number of character codes is N−R × (N−N ′)
4. The method for identifying low-quality characters according to claim 3, wherein 0 ≦ R ≦ 1, N and N ′ can be set by a user. The determining step includes
2. The low quality character identification method according to claim 1, wherein the optimum candidate character code is determined according to the accuracy of the merge candidate character code group obtained by integrating two types of identification distances. The accuracy of the merge candidate character code group is obtained by the overall reliability (C (I)), and the overall reliability (C (I)) is the reliability based on the stroke edge feature (C _l (I)). And the reliability based on image pattern characteristics (C _g (I)),
C (I) = C _l (I) + C _g (I)
The low quality character identification method according to claim 8, wherein: Stroke edge feature processing means for obtaining a first candidate character code group by extracting a stroke edge feature of a character to be identified and performing identification based on the stroke edge feature;
Image pattern feature processing means for obtaining a second candidate character code group by extracting an image pattern feature of a character to be identified and performing identification based on the image pattern feature;
Merging means for generating a merge candidate character code group by merging the first candidate character code group and the second candidate character code group;
Stroke edge feature reprocessing means for calculating an identification distance of each candidate character code of the merge candidate character code group based on a stroke edge feature;
Image pattern feature reprocessing means for calculating an identification distance of each candidate character code of the merge candidate character code group based on an image pattern feature;
An integration unit that acquires the total reliability of each candidate character code by integrating two types of identification distances calculated by the stroke edge feature reprocessing unit and the image pattern feature reprocessing unit;
An output means for outputting the optimum candidate character code from the overall reliability;
An apparatus for identifying low-quality characters, comprising: