JP3370934B2 - Optical character reading method and apparatus - Google Patents

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an optical character reading method and apparatus for reading handwritten or printed characters on a form or document.

[0002]

2. Description of the Related Art As a similarity calculation method in a conventional optical character reading apparatus, a similarity calculation method based on a distance scale is generally used. The distance scale is roughly classified into a distance scale that focuses on individual dimensions of the pattern vector and a distance scale that considers the distribution of the entire dimension. The IEICE Transactions D-II Vol.J79-D-II No .1 pp.45-52 As shown in "High-accuracy handwritten character recognition using improved Mahalanobis distance", the latter similarity calculation method based on the distance scale that considers the distribution of the entire dimension It is known to be effective. Above all, the composite similarity method is one of the general similarity calculation methods that have already been put to practical use.

There is one similarity calculation method of the composite similarity method.
For one character category, a principal component analysis is performed on the distribution in the pattern space created by the learning character patterns contained therein, and N eigenvectors ψi (i = 1, 2, ..., N) obtained as the principal components are analyzed. ) As a dictionary of the category, the similarity to the input character pattern vector f is calculated based on the composite similarity S defined below.

[Equation 1] Here, λi is an eigenvalue. As a recognition result, the presence or absence of a reject is determined based on the magnitude of the ratio of the second largest similarity to the maximum similarity. If no rejection occurs, the category with the highest similarity is generally used as the recognition result. Target.

As a recognition method using the structural feature in the conventional optical character reading apparatus, see Toshiba Review, Vol.
41, No. 12, pp. 1012-1015 (1986) The contour structure matching method shown in "Free handwritten character recognition" is known.

According to this method, a Freeman code chain coded in eight directions is extracted as a feature amount in the first stage, and is divided by the smoothed curvature. The divided partial contours are called segments and are classified into three types: convex, straight, and concave.Each segment has a length, curvature, position,
Attribute information such as direction is given.

On the other hand, a similar sequence of contour segments is described in the dictionary, and the upper and lower limit values of the attribute information are set in these sequences.
By comparing the input pattern and the attribute information described in the dictionary, the correspondence between the segment of the input pattern and the dictionary segment is checked and recognition is executed.

[0007]

However, since the conventional optical character reader uses only the statistical features obtained by the principal component analysis, a deviation from the human recognition characteristic occurs, and the character shape is changed. There is a problem that similar categories, which can be easily separated by humans, cannot be separated by the recognition method using statistical features because structural differences cannot be reflected. In addition, the recognition method using only structural features has a problem that the efficiency of dictionary creation is low.

When a character written on a form or the like having a frame line (non-dropout) is targeted, a part of the frame line is touched by the character contacting the frame line or protruding the frame / correcting the amount of money. It is easy to generate character cut-out errors such as combining with characters, dropping a part of a character, cutting out a correction part or a part of a correction mark as a character, recognizing those patterns and only calculating the similarity ratio. There is also a problem that when a reject is judged, it is not rejected but is often misread.

In view of the above-mentioned conventional problems, the present invention is based on a conventional statistical feature-based recognition method, and extracts structural features by limiting extraction to those necessary for separating similar categories. By improving the efficiency of creation and by using structural features to identify similar categories that are prone to error in the recognition method using statistical features, the ability to separate similar categories, that is, the recognition performance, is improved, and the similarity of character patterns is improved. An object of the present invention is to provide an optical character reading method and an apparatus for efficiently rejecting character cut-out error characters by using not only the ratio but also the degree of similarity.

[0010]

In order to achieve the above object, the present invention separates a character frame line and a character string pattern from character string data including a photoelectrically converted character frame line, and separates them. A character cutout unit that cuts out individual character patterns from a character string pattern, and outputs a size of a circumscribed rectangle for each character pattern and a cutout character image surrounded by the circumscribed rectangle, and individual cutout characters cut out by the character cutout unit. Basic feature extraction means for obtaining contour information and background structure information for the image; statistical feature extraction means for obtaining at least one kind of statistical feature based on the contour direction component and background distance output from the basic feature extraction means; Structural feature extracting means for obtaining at least one type of structural feature based on the basic feature output from the basic feature extracting means; Characteristics determine the similarity of each category of character images cut out from an output statistical features from the extraction means, at least one chosen in descending order of similarity
Similarity between two categories, recognition means for outputting at least one category selected in the same order, categories output from the recognition means, degrees of similarity to the categories, and structural features output from the structural feature extraction means And a first structure identifying means for determining the category to which the final character pattern belongs based on the.

With this configuration, the basic feature extracting means extracts the basic features that can be used as both the statistical features and the structural features to make the configuration common, and the structural feature extracting means identifies them by the recognizing means. By limiting the extraction of the structural features that are effective for separating difficult similar categories, the dictionary creation is made efficient, and the structure identification means is based on the recognition results of the recognition means using statistical features. By selecting the features required for separating similar categories and determining the final category, both efficient processing and improvement of recognition performance of similar categories are possible.

[0012]

BEST MODE FOR CARRYING OUT THE INVENTION The invention according to claim 1 of the present invention separates a character frame line and a character string pattern from character string data including a photoelectrically converted character frame line, and separates the separated character. Each character pattern is cut out from the column pattern, the size of the circumscribing rectangle for each character pattern and the cut-out character image surrounded by the circumscribing rectangle are obtained, and the contour information and the background structure information for the cut out individual cut-out character image are obtained. Extracted as a basic feature, at least one type of statistical feature is extracted based on the extracted basic feature, at least one type of structure based on the clipped individual cut-out character image and the extracted basic feature Characteristics are extracted, the degree of similarity to each category of the cut-out character image is obtained from the extracted statistical characteristics, and the category with the highest degree of similarity is individually calculated. Each of the character patterns to which the final character pattern belongs is determined based on the category to which the determined character pattern belongs, the degree of similarity to the category, and the extracted structural feature. An optical character reading method comprising steps.

With this configuration, it is possible to standardize the configuration by extracting basic features that can be used for both statistical features and structural features. Similar categories that are difficult to identify by recognition processing using statistical features It is necessary in the structure identification means based on the recognition result output from the recognition means that makes the dictionary efficient by extracting only the structural features effective for separating By selecting various structural features and determining the final category, it is possible to improve processing efficiency and improve recognition performance of similar categories.

According to a second aspect of the present invention, in the character string data including the photoelectrically converted character frame line, the character frame line and the character string pattern are separated, and the separated character string pattern is individually separated. , The size of the circumscribing rectangle for each character pattern and the clipped character image surrounded by the circumscribing rectangle are obtained, and the outline information and background structure information for each of the clipped individual clipped character images are extracted as basic features. Then, at least one type of statistical feature is extracted based on the extracted basic feature, and at least one type of structural feature is extracted based on the extracted individual cut-out character image and the extracted basic feature. Then, the degree of similarity to each category of the cut-out character image is obtained from the extracted statistical characteristics, and the category with the highest degree of similarity is set to an individual character. It is determined that belongs to turn category,
The similarity is corrected based on at least one of the size of the circumscribed rectangle for each of the cut out character patterns, the category to which the determined character pattern belongs, and the extracted structural feature, and the similarity after correction is corrected. It is an optical character reading method characterized by comprising respective steps of judging a category to which a final character pattern belongs based on a degree.

With this configuration, it is possible to reflect a subtle difference in the degree of similarity of the cut-out character image with respect to each category in the final category determination, and quantitatively correct the degree of similarity in the detection result of the structural feature. Since it can be reflected in the value, it has an effect of improving the recognition performance.

According to a third aspect of the present invention, the type of handwritten characters or handwritten characters is determined based on the size of the circumscribed rectangle with respect to the extracted character pattern, and the result of the character type determination and the cutout are performed. Determining the category to which the final character pattern belongs based on at least one of the size of the circumscribed rectangle for the character pattern, the category to which the determined character pattern belongs, the similarity to the category, and the extracted structural feature The optical character reading method according to claim 1 or 2, further comprising respective steps.

With this configuration, by performing specialized processing according to the character type, the structural characteristics peculiar to print / handwriting can be used without confusion, the recognition performance between similar categories can be improved, and the processing efficiency can be improved. It has the effect of being able to achieve.

The invention according to claim 4 of the present invention is based on at least one of the size of a circumscribing rectangle for the cut-out character pattern, the similarity to the determined category, and the extracted structural feature. The optical character reading method according to claim 1, 2 or 3, further comprising: a step of determining a cutout error, and re-cutting out the character pattern when the cutout error is determined.

With this configuration, when the circumscribed rectangle is significantly different from the circumscribed rectangles of other characters, is out of the prescribed circumscribed rectangle, or the similarity is extremely low compared to the similarity of other characters, , If the difference between the similarity of the first rank and the similarity of the second rank or less is small, it is judged that there is a cutout error and the cutout is performed again to reduce the cutout error,
This has the effect of reducing erroneous recognition due to clipping errors.

According to a fifth aspect of the present invention, the maximum similarity of the cut-out character pattern, the value of the ratio between the second similarity and the maximum similarity, and the complexity of the background of the cut-out character pattern. On the basis of the degree, the reject processing is performed when the reliability of the category to which the determined character pattern belongs is low, and in other cases, the category to which the determined character pattern belongs is set as the final recognition result. The optical character reading method according to any one of claims 1 to 4, further comprising steps.

With this configuration, the magnitude of the maximum similarity and the background complexity are also used for the rejection determination, so that the preprocessing error character that is difficult to reject only by the ratio of the similarities can be efficiently rejected.

According to a sixth aspect of the present invention, in the character string data including the photoelectrically converted character frame line, the character frame line and the character string pattern are separated, and the character string pattern is separated from the separated character string pattern. A character cutting means for cutting out the character pattern of, and outputting a size of a circumscribing rectangle for each character pattern and a cut out character image surrounded by the circumscribing rectangle, and a basic feature for each cut out character image cut out by the character cutting means. Basic feature extracting means for obtaining contour information and background structure information, statistical feature extracting means for obtaining at least one kind of statistical feature based on the basic features output from the basic feature extracting means, and character cutting means for cutting. At least one structural feature based on the extracted individual character images and the basic features output from the basic feature extraction means. And a similarity of at least one category selected in descending order of similarity, from the structural feature extraction means for obtaining , A recognition unit that outputs at least one category selected in the same order, and a final output based on the category output from the recognition unit, the similarity to the category, and the structural feature output from the structural feature extraction unit. And a first structure identifying means for determining a category to which a specific character pattern belongs.

With this configuration, the basic feature extracting means can make the configuration common by extracting the basic features that can be used for both the statistical feature and the structural feature, and the structural feature extracting means can recognize the feature. It is possible to improve the efficiency of dictionary creation by extracting mainly structural features that are effective for separating similar categories that are difficult to identify by means, and use the statistical features to identify the recognition results output from the recognition Based on this, the structure identification means selects necessary structural features and determines the final category, which has the effect of making it possible to improve processing efficiency and improve recognition performance of similar categories.

According to a seventh aspect of the present invention, the first structure identifying means outputs at least one structural element output from the structural feature extracting means based on the category output from the recognizing means. 7. The recognition-dependent feature selecting means for selecting an effective structural feature from the features, and the selected structural feature identifying means for determining a category based on the selected structural feature. The optical character reading device described above.

With this configuration, the structural identification means selects necessary structural features and determines the final category based on the recognition result output from the recognition means that performs the identification using the statistical features. This has the effect of further improving the efficiency of processing and improving the recognition performance of similar categories.

The invention according to claim 8 of the present invention comprises a character type judging means for judging a printed character or handwriting based on the size of the circumscribing rectangle outputted from the character cutting means,
The first structure identification means determines a category to which each character pattern belongs to a character determined to be a print type by the character type determination means, and a handwritten character by the character type determination means. 8. The optical character reading device according to claim 6 or 7, further comprising: first handwriting processing means for determining a category to which each character pattern belongs for the character determined to be.

With this configuration, by performing specialized processing according to the character type, the structural characteristics peculiar to print / handwriting can be used without confusion, the recognition performance between similar categories can be improved, and the processing efficiency can be improved. It has the effect of being able to achieve.

According to a ninth aspect of the present invention, the first type processing means is structurally characterized based on the size of the circumscribed rectangle output from the character cutting means and the category output from the recognition means. Recognition-dependent feature selection means for selecting an effective structural feature from at least one structural feature output from the extraction means, and selection for performing category determination based on the structural feature selected by the recognition-dependent feature selection means 9. The optical character reading device according to claim 8, further comprising a structural feature identifying means.

With this configuration, the structural identification means selects necessary structural features and determines the final category based on the recognition result output from the recognition means that performs the identification using the statistical features. It has an effect that the efficiency of the processing can be further improved and the recognition performance of the similar category can be improved.

According to a tenth aspect of the present invention, in the first character processing means, based on the size of the circumscribing rectangle output from the character cutting means and the category output from the recognizing means, A circumscribing rectangle category determining means for making a final category determination for the category, and at least one of the structural feature extracting means based on the category output from the recognizing means for a character that cannot be determined by the circumscribing rectangular category determining means. A recognition dependent feature selecting means for selecting an effective structural feature from the structural features, and a selected structural feature identifying means for performing a category determination based on the structural feature selected by the recognition dependent feature selecting means. 9. The optical character reading device according to claim 8, wherein

With this configuration, the circumscribing rectangle category determining means first determines a category that can be discriminated only by the size of the circumscribing rectangle, so that the categories are efficiently separated, and the recognition-dependent feature selecting means determines the result of the recognizing means. Depending on the structural features required for similar category separation, the final category judgment can be made according to the structural features selected by the selected structural feature identification means, enabling efficient and effective similar category separation. Has the action of

The invention according to claim 11 of the present invention is a structural feature, wherein the first handwriting processing means is based on the size of the circumscribed rectangle output from the character cutting means and the category output from the recognition means. Recognition-dependent feature selection means for selecting an effective structural feature from at least one structural feature output from the extraction means, and category determination based on the structural feature selected by the recognition-dependent feature selection means 9. The optical character reader according to claim 8, further comprising a selective structural feature identifying means.

With this configuration, the structural feature required by the structure identifying means is selected and the final category is determined based on the recognition result output from the recognizing means that performs the identification using the statistical feature. This has the effect of further improving the efficiency of processing and improving the recognition performance of similar categories.

According to a twelfth aspect of the present invention, in the first handwriting processing means, based on the size of the circumscribing rectangle output from the character cutting means and the category output from the recognizing means, A circumscribing rectangle category determining means for making a final category determination for the category, and at least one of the structural feature extracting means based on the category output from the recognizing means for a character that cannot be determined by the circumscribing rectangular category determining means. A recognition dependent feature selecting means for selecting an effective structural feature from the structural features, and a selected structural feature identifying means for performing a category determination based on the structural feature selected by the recognition dependent feature selecting means. 9. The optical character reading device according to claim 8, wherein

With this configuration, the circumscribing rectangle category determining means first determines a category that can be discriminated only by the size of the circumscribing rectangle, so that the categories are efficiently separated, and the recognition-dependent feature selecting means determines the result of the recognizing means. Depending on the structural features required for similar category separation, the final category judgment is performed according to the structural features selected by the selected structural feature identification means, which enables efficient and effective similar category separation. Has the effect of being.

According to a thirteenth aspect of the present invention, the character frame line and the character string pattern are separated for the character string data including the photoelectrically converted character frame line, and the character string pattern is separated from the separated character string pattern. A character cutting means for cutting out the character pattern of, and outputting a size of a circumscribing rectangle for each character pattern and a cut out character image surrounded by the circumscribing rectangle, and a basic feature for each cut out character image cut out by the character cutting means. Basic feature extracting means for obtaining contour information and background structure information, statistical feature extracting means for obtaining at least one kind of statistical feature based on the basic features output from the basic feature extracting means, and character cutting means for cutting. At least one structural type based on the extracted individual character images and the basic features output from the basic feature extraction means. And structural features extracting means for obtaining the symptoms,
Obtaining the degree of similarity to each category of the cut-out character image from the statistical features output from the statistical feature extraction means,
Similarity of at least one category selected in descending order of similarity, recognition means for outputting at least one category selected in the same order, size of the circumscribed rectangle of the character pattern output from the character cutting means, and the recognition means The similarity output from the recognition means based on at least one of the category output from the structural feature extraction means and the structural feature output from the structural feature extraction means, and the character pattern based on the corrected similarity. And a second structure identifying means for finally determining a category to which the character belongs.

With this configuration, a subtle difference in the similarity output from the recognition means can be reflected in the final category determination, and the detection result of the structural feature can be quantitatively used as a correction value for the similarity. Since it can be reflected, it has an effect of improving recognition performance.

According to a fourteenth aspect of the present invention, in the second structure identifying means, the size of the circumscribed rectangle of the character pattern output from the character cutting means and at least one output from the recognizing means. Similarity correction means for correcting the similarity output from the recognition means based on the category and the structural feature output from the structural feature extraction means, and similarity of at least one category output from the recognition means Degree and the similarity degree corrected by the similarity degree correction means to calculate a corrected similarity degree, and a correction category for determining a category to which each character pattern belongs from the corrected similarity degree output from the adder. 14. The optical character reading device according to claim 13, further comprising a determining unit.

With this configuration, the subtle difference in the similarity output from the recognition means can be reflected in the final category determination, and the detection result of the structural feature can be quantitatively used as the correction value of the similarity. Since it can be reflected, it has an effect of improving recognition performance.

According to a fifteenth aspect of the present invention, there is provided a character type determining means for determining a printed character or handwriting based on the size of the circumscribing rectangle output from the character cutting means, and the second structure identification. A second character processing unit that determines a category to which each character pattern belongs to a character that is determined to be a print character by the character type determination unit; and a character that is determined to be handwritten by the character type determination unit. On the other hand, the optical character reader according to claim 13 or 14, further comprising a second handwriting processing means for determining a category to which each character pattern belongs.

With this configuration, by performing the specialized processing according to the character type, the structural characteristics peculiar to print / handwriting can be used without confusion, the recognition performance between similar categories can be improved, and the processing efficiency can be improved. It has the effect of being able to achieve.

According to a sixteenth aspect of the present invention, in the second character processing means, the size of the circumscribing rectangle of the character pattern output from the character cutting means and at least one output from the recognizing means. Similarity correction means for correcting the similarity output from the recognition means based on the category and the structural feature output from the structural feature extraction means, and similarity of at least one category output from the recognition means Degree and the similarity degree corrected by the similarity degree correction means to calculate a corrected similarity degree, and a correction category for determining a category to which each character pattern belongs from the corrected similarity degree output from the adder. 16. The optical character reading device according to claim 15, further comprising a determining unit.

With this configuration, the subtle difference in the similarity output from the recognition means can be reflected in the final category determination, and the detection result of the structural feature can be quantitatively used as the correction value of the similarity. Since it can be reflected, it has an effect of improving recognition performance.

According to a seventeenth aspect of the present invention, in the second handwriting processing means, at least one of the size of the circumscribed rectangle of the character pattern output from the character cutting means and the recognition means is output. Based on the structural features output from the category and the structural feature extraction means,
Similarity correction means for correcting the similarity output from the recognition means, and correction similarity by adding the similarity of at least one category output from the recognition means and the similarity corrected by the similarity correction means An adder for calculating degrees,
16. The optical character reading device according to claim 15, further comprising a correction category determination unit that determines a category to which each character pattern belongs from the correction similarity output from the adder.

With this configuration, the subtle difference in the similarity output from the recognition means can be reflected in the final category determination, and the detection result of the structural feature can be quantitatively used as the correction value of the similarity. Since it can be reflected, it has an effect of improving recognition performance.

In the eighteenth aspect of the present invention, the basic feature extracting means includes contour point extracting means for extracting contour points of a character pattern in a cutout image output from the character cutting means, and the contouring point extracting means. A contour direction calculation means for calculating a contour direction component in pixel units from the adjacency relation of individual contour points outputted from the point extraction means, and a circumscribed character pattern based on the position of the contour point outputted from the contour point extraction means. The background distance calculation means for obtaining the background distance from each side of the rectangle to the character pattern and the cutout image output from the character cutout means are scanned for each line in the horizontal and vertical directions, and white pixels continuous for a certain number of pixels (background ) Continuous length / position information and background information calculation means for calculating the background order which is the generation order of the background along the scanning direction. The optical character reading device according to any one of claim 6, 8, 13, and 15.

With this configuration, the contour direction component is calculated in the contour direction calculation unit, the background distance from each side of the circumscribed rectangle to the character pattern is calculated in the background distance calculation unit, and the background information calculation is performed in the background information calculation unit. By extracting the contour points of the character pattern by the contour point extraction unit as common preprocessing, the efficiency of basic feature extraction is improved. Also,
By extracting contour points, contour direction components, background distances, and background information as basic features, not only statistical features but also structural feature quantities that reflect the character structure can be diverted and the efficiency of feature extraction processing can be improved. Has the effect of.

According to a nineteenth aspect of the present invention, the statistical feature extracting means includes a bidirectional block dividing means for dividing the clipped image output from the character clipping means into a plurality of blocks in both horizontal and vertical directions. From the contour direction totalizing means and the contour direction totaling means for totaling the number of each of the contour direction components output from the contour direction calculating means of the basic feature extracting means in block units divided by the bidirectional block dividing means The contour direction normalizing means for obtaining the contour statistical feature amount for each contour direction obtained by dividing the number of the outputted contour direction components for each contour direction by the block size, and the output from the character cutting means. Output from the character cutting means and a horizontal block dividing means for dividing the cut-out image into a plurality of blocks in the horizontal direction. A vertical block dividing unit that divides the clipped image into a plurality of blocks in the vertical direction, and a horizontal block dividing unit and a vertical block dividing unit based on the contour direction component output from the basic feature extracting unit. The number of contour direction components for each contour direction of the pixels that form the first character pattern that scans in a direction parallel to the block division boundary line from the side of the circumscribed rectangle perpendicular to the block division boundary line in the unit of the block. An outline direction totaling means for totaling the number of outline direction components, and an outline for each outline direction obtained by dividing the number of outline direction components for each outline direction for each block output from the outline direction totalizing means by the size of the block. Based on the background distance output from the outline direction normalizing means for obtaining the statistical feature amount and the background distance calculating means of the basic feature extracting means, The background distance totaling means for totaling the background distances in the block dividing direction in units of blocks divided by the horizontal block dividing unit and the vertical block dividing unit, and the aggregate background distance in block units output from the background distance aggregating unit are the block sizes. 16. The optical character reading device according to claim 6, further comprising a background distance normalization unit that obtains a background distance feature amount obtained by dividing by the distance.

With this configuration, the bidirectional block dividing means, the horizontal block dividing portion, and the vertical block dividing means divide the character pattern into a plurality of blocks,
By adding and totaling the feature amounts of each block by the contour direction totaling means, the outline direction totaling means, and the background distance totaling means, the statistical characteristics of the blocks can be extracted from the basic characteristics extracted by the basic characteristic extracting means, Further, by normalizing the statistical characteristics of the block by the contour direction normalizing means, the outer contour direction normalizing means, and the background distance normalizing means,
This has the effect of extracting features that are less likely to be affected by changes in the recognition target.

According to a twentieth aspect of the present invention, the structural feature extracting means extracts a local background distance feature based on the background distance output from the background distance calculating means of the basic feature extracting means. Local stroke feature extracting means for extracting a local stroke direction change based on the contour direction component output from the contour direction calculating means of the basic feature extracting means; and the character cutting means. A first extracting a loop of a character pattern based on the output cut-out character image and the contour direction component output from the contour direction calculating means of the basic feature extracting means.
Loop extraction means and the background information calculation means of the basic feature extraction means, and the background information calculated by the background information, and a block of adjacent background pixels for each background degree is obtained as a background block, and its size information and position information are calculated. The background block extracting means, and the label number is obtained as a background label lump of adjacent white pixels based on the background information calculated by the background information calculating means of the basic feature extracting means, and its size, position, its Background label extracting means for calculating label information indicating which side of the circumscribed rectangle the background label touches, and end points of strokes forming a character pattern based on the background information calculated by the background information calculating means of the basic feature extracting means. End point / intersection point extraction means for extracting the intersection point and the background information calculated by the background information calculation means of the basic feature extraction means and the end point / intersection point. 16. The optical character reading device according to claim 6, further comprising: a straight line extracting unit that extracts a straight line based on the end points and the intersections extracted by the extracting unit. is there.

With this configuration, based on the basic features extracted by the basic feature extraction unit, the local background feature extraction unit extracts partial unevenness information of the background portion, which is effective for discrimination between specific categories, The local stroke shape extraction unit extracts the difference in stroke bending direction and linearity between similar categories having high universality, and the first loop extraction unit extracts the number of stroke loops, which is a characteristic of high universality. The block extraction unit extracts the structure of the character background portion, the background label extraction unit extracts the size of the character background portion,
By extracting the positional relationship with the circumscribed rectangle, extracting the information of the end points / intersections in the end point / intersection point extraction section, and extracting the straight line in the straight line extraction section Features can be extracted efficiently, and by using these structural features, it is possible to improve the recognition performance between similar categories, which is difficult to identify with an identification method that uses only statistical features. .

In the twenty-first aspect of the present invention, the structural feature extraction unit extracts a local background distance feature based on the background distance output from the background distance calculation unit of the basic feature extraction unit. Local background feature extraction means, local stroke shape extraction means for extracting local stroke direction changes based on the contour direction component output from the contour direction calculation means of the basic feature extraction means, and the basic feature extraction means From the background information calculated by the background information calculating unit, a background block extracting unit that obtains a block of adjacent background pixels for each background degree as a background block, and calculates size information and position information thereof, and the basic feature extracting unit. From the background information calculated by the background information calculation means, a block of adjacent white pixels is obtained as a background label, a label number is given, and its size and position are calculated. A background label extracting means for calculating label information indicating which side of the circumscribing rectangle the background label touches, and a stroke forming a character pattern based on the background information calculated by the background information calculating means of the basic feature extracting means. End point / intersection point extracting means for extracting the end point and the intersection point, and a straight line is extracted based on the background information calculated by the background information calculating means of the basic feature extracting means and the end point / intersection point extracted by the end point / intersection point extracting means. Straight line extracting means,
7. A second loop extracting means for extracting a loop of a character pattern based on label information calculated by the background label extracting means.
5. The optical character reading device according to any one of 5 above.

With this configuration, based on the basic features extracted by the basic feature extraction unit, the local background feature extraction unit extracts the partial unevenness information of the background portion which is effective for discrimination between specific categories, and The stroke shape extraction unit extracts the difference in stroke bending direction and linearity between similar categories with high universality, the background block extraction unit extracts the structure of the character background portion, and the background label extraction unit extracts the size of the character background portion. Now, the positional relationship with the circumscribed rectangle is extracted, the end point / intersection point extraction unit extracts end point / intersection point information, the straight line extraction unit extracts a straight line, and the second loop extraction unit has a highly universal characteristic. By extracting the number of stroke loops, it is possible to efficiently extract structural features that more explicitly represent the structural features of a character. Ri The identification method using only statistical characteristic quantity has the effect that it is possible to improve the recognition performance between hard similar categories identified. In addition, by extracting the number of loops based only on the label information, there is an effect that the efficiency of processing can be improved.

According to a twenty-second aspect of the present invention, the end point / intersection point extracting means is configured so that the end point / intersection point of the character pattern is based on the background information of each line in the vertical and horizontal directions output from the background information calculating means. By comparing the respective positional relationships of the extracted end points and intersection point candidates with the end point / intersection point candidate extraction means for extracting the candidates, inappropriate end points and intersection point candidates are removed to extract appropriate end points and intersection points. 21. An end point / intersection point determining means is provided.
Or the optical character reading device described in 21.

With this configuration, in the end point / intersection point extraction means, the background information extracted in advance by the background information of the basic feature extraction means is used to compare the background information of the adjacent lines and to perform a simple process of checking the difference. You can extract end points and intersection candidates by using the end point / intersection determination section to remove end points or intersection candidates that are too close to other end points or intersection candidates based on the positional relationship between the end points and intersection candidates. This has the effect of reducing erroneous extraction of end points or intersections due to noise, and more accurately extracting end point and intersection information that is effective in separating similar categories such as category 2 and category 3.

According to the twenty-third aspect of the present invention, the straight line extracting means sets a threshold value at which a black pixel is predetermined based on the background information of each line in the vertical and horizontal directions output from the background information calculating means. A black pixel continuous area extracting means for extracting a continuous area longer than that, a straight line candidate extracting means for extracting vertical and horizontal straight line candidates based on the black pixel continuous area output from the black pixel continuous area extracting means, The non-straight line candidates are removed from the straight line candidates output from the straight line candidate extraction means based on the position information of the end points and the intersection points output from the end point / intersection point extraction means, and a neighboring line is selected from the remaining straight line candidates. 22. The optical character reading device according to claim 20, further comprising: a straight line determination unit that determines the type according to the connection state of.

With this configuration, in the black pixel continuous region extracting means and the straight line candidate extracting means, the background information extracted in advance by the background information calculating means of the basic feature extracting means is used, and the degree of continuity of black pixels and a plurality of lines are extracted. Straight line candidates can be extracted by a simple process that only checks continuity, and erroneous straight line candidates that occur due to line merging or branching can be removed using the position information of end points and intersections in the straight line determination unit. However, by classifying the remaining straight line candidates according to the connection state with neighboring lines, it is possible to obtain a highly accurate straight line that is effective for separating a category such as the category \ that is characteristic of a short horizontal straight line. It has the effect that it can be extracted.

According to a twenty-fourth aspect of the present invention, the size of the circumscribing rectangle of the character pattern output from the character cutting means, at least one similarity degree output from the recognition means, and the background information calculation means. 7. An error determination means for determining a cutout error on the basis of at least one of the background complexity levels calculated based on the background information, and for giving a recutout instruction to the character cutout means. The optical character reader according to any one of claims 8 , 8, 13 and 15 .

With this configuration, when the circumscribed rectangle is significantly different from the circumscribed rectangle of another character, is out of the prescribed circumscribed rectangle, or when the similarity is extremely lower than the similarity of other characters, , If the difference between the similarity of the first rank and the similarity of the second rank or less is small, it is judged that there is a cutout error and the cutout is performed again to reduce the cutout error,
This has the effect of reducing erroneous recognition due to clipping errors.

According to a twenty-fifth aspect of the present invention, the category similarity calculated by the recognizing means is 2 in descending order.
One category similarity is selected, and the value of the ratio of the second similarity and the maximum similarity is larger than a preset threshold value a, or the maximum similarity is smaller than a preset threshold value b, or A reject signal is output when the background complexity calculated based on the background information calculated by the background information calculating unit is larger than a preset threshold value c, and otherwise a category of maximum similarity is output as a recognition result. The optical character reading device according to any one of claims 6, 8, 13, 15, and 24, characterized in that the recognition result calculating means (1) is provided.

With this configuration, the magnitude of the maximum similarity and the background complexity are also used for the rejection determination, so that the preprocessing error character which is difficult to reject only by the ratio of the similarities can be efficiently rejected.

According to a twenty-sixth aspect of the present invention, for a plurality of character patterns, the ratio of the second highest similarity to the highest similarity, the highest similarity, and the background information calculating means calculate The background complexity based on the background information and the recognition result output by the recognition unit are calculated, and at least one of the similarity ratio, the maximum similarity and the background complexity is input, and the recognition result is an incorrect answer. 7. A second recognition result calculation means configured by a hierarchical neural network that is preliminarily learned so as to output a reject signal in the case of being correct and not to output the reject signal in the case of a correct answer. The optical character reading device according to any one of 8, 13, 15, and 24.

With this configuration, the nonlinear boundary between the reject area and the non-reject area in the space centered on at least one of the similarity ratio, the maximum similarity and the background complexity is learned in advance by the hierarchical neural network. Since this is done, there is an effect that more accurate reject determination processing becomes possible.

[0064]

[0065]

[0066]

[0067]

[0068]

(Embodiment 1) Hereinafter, a first embodiment of the present invention will be described with reference to the drawings. FIG. 1 is a block diagram of an optical character reading device according to a first embodiment of the present invention. In FIG. 1, reference numeral 1 denotes a character cutout unit, which inputs character string data including a character frame line, separates the character frame line and the character string pattern, and cuts out individual character patterns from the separated character string pattern. The size of the circumscribing rectangle with respect to the character pattern is output via the path 11, and the cut-out character image is output via the path 12. Reference numeral 2 denotes a basic feature extraction unit (see FIG. 2), which is based on the cut-out image of each character output from the first character cut-out unit 1 via the path 12 and determines the contour direction component of the character pattern and the circumscribed rectangle Background distance from side to character pattern,
The background information which is a continuation of white pixels other than the character portion for each horizontal and vertical line is obtained, and the contour direction component is output and output via the path 13, the background distance via the path 14, and the background information via the path 15. Reference numeral 3 denotes a statistical feature extraction unit (see FIG. 3), which is a cutout image of a character pattern output from the first character cutout unit 1 via a route 12 and output from the basic feature extraction unit 2 via a route 13. The statistical feature amount is obtained based on the contour direction component and the background distance output via the path 14. A recognizing unit 4 obtains the similarity for each category of the character pattern based on the statistical feature amount output from the statistical feature extracting unit 3, and selects one of a plurality of category similarities and at least one in descending order of similarity. Output category. Reference numeral 5 denotes a first structural feature extraction unit (see FIG. 4), which is a contour direction component output from the basic feature extraction unit 2 via the route 13, a background distance output via the route 14, and a route 15. Then, the structural feature is obtained based on the background information output. 6
Is a character type determination unit that determines whether the character type is handwritten or printed based on the size of the circumscribed rectangle output from the character cutout unit 1 via the path 12. 7 is the first structure identification unit 7
(See FIG. 5), the size of the circumscribed rectangle of the character pattern output from the character cutout unit 1, the category output from the recognition unit 4, the structural feature output from the first structural feature extraction unit 5, and the character Based on the character type determination result output from the type determination unit 6, the final category to which the character pattern belongs is determined.

FIG. 2 is a block diagram specifically showing the configuration of the basic feature extraction unit 2. Reference numeral 21 is a contour point extraction unit that extracts the contour points of the character pattern from the size of the circumscribing rectangle output from the character cutout unit 1 via the path 11 and the cutout image output from the character cutout unit 1 via the path 12. . A contour direction calculation unit 22 obtains a contour direction component of each contour point based on the adjacency relationship between the contour points of the character pattern output from the contour point extraction unit 21. Reference numeral 23 denotes a background distance calculation unit that calculates the background distance from each side of the circumscribed rectangle of the character pattern to the character pattern based on the position of the contour point of the character pattern output from the contour point extraction unit 21. Reference numeral 24 denotes a background information calculation unit, which is based on the positions of the contour points of the character pattern output from the contour point extraction unit 21 and has a length (consecutive background) of consecutive white pixels continuous for a certain number of pixels in each line in the horizontal and vertical directions. -Calculate the position information and the background order which is the generation order of the background along the scanning direction.

FIG. 3 is a block diagram specifically showing the configuration of the statistical feature extraction unit 3. A bidirectional block dividing unit 31 divides the entire cutout region in both horizontal and vertical directions based on the cutout image output from the character cutout unit 1 via the path 11. A horizontal block dividing unit 32 divides the cutout region in the horizontal direction based on the cutout image output from the character cutout unit 1 via the path 11. A vertical block dividing unit 33 divides the cutout region in the vertical direction based on the cutout image output from the character cutout unit 1 via the path 11. A contour direction totaling unit 34 totalizes the number of contour direction components of each pixel output from the basic feature extracting unit 2 via the path 13 for each block divided by the bidirectional block dividing unit 31 and for each contour direction. Reference numeral 35 denotes a contour direction normalization unit, which is provided for each of the blocks divided by the bidirectional block division unit 31,
In step 4, the contour direction components tabulated in the block are divided by the size of the block and normalized to obtain a statistical contour direction statistic. Reference numeral 36 denotes an outline direction totaling unit, which is for each block divided by the horizontal block dividing unit 32 and the vertical block dividing unit 3.
For each of the blocks divided in 3, parallel to the block division boundary line in the contour direction component output from the basic feature extraction unit 2 via the path 13 from the side of the circumscribed rectangle perpendicular to the block division boundary line. Then, the number of contour direction components for each contour direction component of the first character pattern pixel is counted. Reference numeral 37 denotes an outline direction normalization unit which, for both the blocks divided by the horizontal direction block division unit 32 and the blocks divided by the vertical direction block division unit 33, outline direction components (outside direction Component) is divided by the size of the block and normalized to obtain the statistical contour direction statistic. Reference numeral 38 denotes a background distance totaling unit, which is for each block divided by the horizontal block dividing unit 32.
For each side of the circumscribed rectangle perpendicular to the block division boundary line and each side of the circumscribed rectangle perpendicular to the block division boundary line of each block divided by the vertical direction block division unit 33, the path from the basic feature extraction unit 2 to the route 14 Among the background distances output via, the background distances measured from each side of the circumscribed rectangle in the direction parallel to the block division boundary line are selected and totaled. Reference numeral 39 denotes a background distance normalization unit, which divides the background distance calculated by the background distance calculation unit 38 by the length of the side of the circumscribing rectangle perpendicular to the block boundary line and normalizes the background distance to obtain a statistical background distance feature.

FIG. 4 is a block diagram specifically showing the configuration of the first structural feature extraction unit 5. A local background feature extraction unit 41 extracts a local change in background distance and a difference in distance for identifying similar categories based on the background distance output from the basic feature extraction unit 2 via the route 13. . Four
Reference numeral 2 is a local stroke shape extraction unit that extracts a local change in the inclination of the stroke for identifying a similar category based on the contour direction component output from the basic feature extraction unit 2 via the path 14. Reference numeral 43 denotes a loop extraction unit, which is the first
The loop generated by the stroke of the character pattern is extracted based on the cutout image output from the character cutout unit 1 and the contour direction component output from the basic feature extraction unit 2 via the path 14. Reference numeral 44 denotes an end point / intersection point extraction unit that, based on the background information output from the basic feature extraction unit 2 via the path 15, extracts a character pattern from the difference in the background information of adjacent lines in both the vertical and horizontal directions. The end points and intersections of the constituent strokes are extracted. Reference numeral 45 denotes a straight line extraction unit, which extracts straight line candidates forming a character pattern based on the background information output from the basic feature extraction unit 2 via the route 15 and outputs the end points and intersection points output from the end point / intersection point extraction unit 44. The type of straight line candidate is determined using the position information of. Reference numeral 46 is a background block extraction unit that, based on the background information calculated by the background information calculation unit 24 of the basic feature extraction unit 2 and output via the path 15, forms a block of adjacent background pixels for each background order. And the size information and position information are calculated. Reference numeral 47 is a background label extraction unit, which obtains a cluster of adjacent white pixels as a background label based on the background information calculated by the background information calculation unit 24 of the basic feature extraction unit 2 and output via the path 15. Label number And the label information indicating the size, position, and which side of the circumscribed rectangle the background label contacts.

FIG. 5 is a block diagram specifically showing the structure of the first structure identifying section 7. Reference numeral 51 denotes a first type character processing unit (see FIG. 6) which, when the character type determination unit 6 determines that the character pattern is a type character, determines the size of the circumscribed rectangle with respect to the character pattern output from the first character cutout unit. Recognition unit 4
The final category is determined by selecting the structural feature extracted by the first structural feature extraction unit 5 based on the category determined in. 52 is a first handwriting processing unit (see FIG. 7),
When the character type determination unit 6 determines that the character pattern is handwritten, the first structural type is determined based on the size of the circumscribed rectangle for the character pattern output from the first character cutout unit and the category determined by the recognition unit 4. The final category is determined by selecting the structural features extracted by the feature extraction unit 5.

FIG. 6 is a block diagram specifically showing the configuration of the first character processing unit 51. Reference numeral 61 denotes a first circumscribing rectangle category determining unit, which outputs the category from the recognizing unit 4 when it is highly possible to determine one category based on the size of the circumscribing rectangle output from the character cutting unit 1. If it is included in the category, the category is determined. Reference numeral 62 is a first recognition-dependent feature selection unit,
Of the plurality of structural features output from the first structural feature extracting unit 5 based on the category output from the recognizing unit 4 for the character whose category cannot be determined by the circumscribing rectangle category determining unit 61. Select valid structural features. 63
Is a first selected structural feature identifying unit, which makes a final category determination using the structural feature selected by the first recognition-dependent feature selecting unit 62.

FIG. 7 is a block diagram specifically showing the configuration of the first handwriting processing section 52. Reference numeral 71 denotes a second circumscribing rectangle category determining unit, which outputs the category from the recognizing unit 4 when there is a high possibility that it can be determined as one category based on the size of the circumscribing rectangle output from the character cutting unit 1. If it is included in the category, the category is determined. Reference numeral 72 denotes a second recognition-dependent feature selection unit, which extracts the first structural feature on the basis of the category output from the recognition unit 4 for the character whose category cannot be determined by the second circumscribing rectangle category determination unit 71. An effective structural feature is selected from the plurality of structural features output from the unit 5. 7
Reference numeral 3 denotes a second selected structural feature identifying unit, which makes a final category determination using the structural feature selected by the recognition-dependent feature selecting unit 72.

FIG. 8 is a block diagram specifically showing the structure of the end point / intersection point extraction section 44. Reference numeral 81 denotes an end point / intersection point candidate extraction unit that extracts end points and intersection points of the character pattern based on the white pixel continuous area information of each line in the vertical and horizontal directions output from the background information calculation unit 23. Reference numeral 82 denotes an end point / intersection point determination unit, which compares the individual positional relationships of the end point and the intersection point candidate output from the end point / intersection point candidate extraction unit 81. , Remove end points and intersections.

FIG. 9 is a block diagram specifically showing the structure of the straight line extracting section 45. 91 is a black pixel continuous area extraction unit,
Based on the white pixel continuous area information of each line in the vertical and horizontal directions output from the background information calculation unit 23, an area in which black pixels continue for longer than a predetermined threshold value is obtained, and the start point and the end point of the area are stored. . A straight line candidate extraction unit 92 extracts vertical and horizontal straight line candidates based on the black pixel continuous region output from the black pixel continuous region extraction unit 91.
A straight line determination unit 93 collates the straight line candidates output from the straight line candidate extraction unit 92 with the position information of the end points and the intersection points output from the end point / intersection point extraction unit 44, and removes the non-straight line candidates. Then, the type of straight line is determined from the remaining straight line candidates according to the connection state with the neighboring line.

The operation of the optical character reading device configured as described above will be described below. Below,
Although the case where the recognition target is mainly a number has been described, the actual recognition target may be a character other than a number. First, in the character cutout unit 1, as shown in FIG. 10, in a multivalued original image composed of a plurality of frame lines and character character string data, the frame lines and the character strings are separated and the character strings are extracted. The width and height of the clipped image and the circumscribing rectangle enclosed by the circumscribing rectangle of each character are calculated, and the width and height of the circumscribing rectangle are output via route 11 and binarized to black and white. The cut-out image is output via the path 12. As shown in FIG. 11, the contour point extraction unit 21 of the basic feature extraction unit 2 extracts contour points from the cut-out character image output from the character cutout unit 1 to obtain a contour image. As a specific contour point extraction method, when all pixels adjacent in the vertical and horizontal diagonal directions to an arbitrary black pixel are black, the arbitrary black pixel is replaced with white, and the finally remaining black pixel is There is a method of setting contour points.

In the contour direction calculation unit 22, the contour direction is calculated as shown in FIG.
The direction of the contour point of the central pixel is determined by matching with the mask pattern as shown in FIG. 13 with respect to the contour image extracted by the contour point extraction unit 21 by encoding as shown in FIG. For example, in the 3 × 3 mask of FIGS. 13 (a) and 13 (b), the pixels shown in gray may be white pixels or black pixels, and the other pixels shown in white or black always match the same color. In the matching method that is necessary, if the same pattern as the 3 × 3 mask of FIG. 13 (a) exists in any 3 × 3 region of the contour image, the central pixel of the 3 × 3 region is coded into 7. , And if it is the same as in FIG. The number of coding in the contour direction may be other than 8,
It may be 4 or 16. The size of the pattern matching mask for determining the contour direction may be other than 3 × 3, and may be 2 × 2 or 5 × 5 depending on the number of coding in the contour direction.

In the background distance calculation unit 23, the contour point extraction unit 2
With respect to the contour image extracted in step 1, the distance to the contour pixel first found by scanning in the right direction for each row of the image, with the point on the left side of the circumscribed rectangle as the origin, as shown in FIG. , As shown in FIG. 14 (b), with the point on the right side of the circumscribed rectangle as the origin, the distance to the first contour pixel found by scanning left in each row of the image, and as shown in FIG. 14 (c). The origin to the point on the upper side of the circumscribed rectangle, and the distance to the first contour pixel found by scanning downward in each column of the image, and the point on the lower side of the circumscribed rectangle as shown in Fig. 14 (d). As the origin, all the distances to the first contour pixel found by scanning upward in each column of the image are obtained.

In the background information calculation unit 24, the contour point extraction unit 2
As shown in FIG. 15, the contour image extracted in 1 is horizontally scanned from the left side of the circumscribed rectangle to detect a character background in which white pixels are continuous. In the case of line a, the left edge of the circumscribed rectangle and the first contour point, the first contour point and the second contour point,
The continuation of the white pixel sandwiched between the second contour point and the right edge of the circumscribed rectangle is used as the 0th background and the 1st background, respectively, and the respective continuous lengths (x, y) and the start and end points as position information are calculated. As in the case of the line b, the region sandwiched between the circumscribed rectangle end and the contour point is regarded as the background even if the number of continuous white pixels is 0. In this case, the background order is 2 (the background number is 3). Similarly, scanning from the right edge of the circumscribing rectangle in the horizontal direction,
The same applies to scanning from the upper and lower ends of the circumscribed rectangle in the vertical direction. Here, when detecting the background sandwiched between the contour points, the background noise may be removed by providing a restriction that only when a certain number of white pixels continue, it is regarded as the background and the background order is counted.

Next, the bidirectional block division unit 31 of the statistical distance feature extraction unit 3 divides the cutout image input from the character cutout unit 1 into a plurality of regions. Specifically, FIG.
As shown in, the center of gravity is divided into a plurality of regions in the horizontal / vertical direction so that the number of black pixels contained in each region becomes equal. It should be noted that the area division method may be other than the centroid division, and may be an uneven division method or an equal division method other than the centroid division. The horizontal block division unit 32 and the vertical block division unit 33 also perform division in the same manner. Note that the horizontal division position in the bidirectional block division unit 31 and the division position in the horizontal block division unit 32 are made the same, and the vertical division position in the bidirectional block division unit 31 and the division position in the vertical block division unit 33 are made the same. It doesn't matter.

The contour direction totaling unit 34 calculates the number of the contour direction components output from the basic feature extracting unit 2 via the path 13 for each block divided by the bidirectional block dividing unit 31 and for each contour direction. For example, in the contour direction image of FIG. 17 (a), if a thick line is used as a block boundary and each number in the block border represents the sign of the contour direction component of each pixel, the number is aggregated for each block and each contour direction. The results are shown in FIGS. 17 (b) to 17 (e).

In the contour direction normalizing section 35, the number of contour direction components for each block and for each contour direction tabulated by the contour direction tabulating section 34 is determined by the size of the block for contour direction tabulation output from the bidirectional block dividing section. Divide and normalize,
The contour direction statistical feature amount is obtained. As the size of the block, there is either the height or the width, whichever is longer, or the area obtained from the height and the width. Note that the number of adjacent contour direction components may be weighted and integrated before normalization.

In the outer contour direction totaling unit 36, the contour direction component output from the basic feature extracting unit 2 via the route 13 is divided into blocks by the horizontal direction block dividing unit 32 and the vertical direction block dividing unit 33. The number of contour direction components is calculated for each contour direction component (outer contour direction component) of the first character pattern pixel by scanning in parallel with the division boundary line. For example, in FIG. 18, which is the contour direction image of a certain block divided by the horizontal direction block division unit 32, each number represents the sign of the contour direction component of each pixel, and the symbol − represents a pixel that is not a contour point, When the scanning direction is from the left end to the right direction as indicated by the arrow, the first character pattern pixel scanned in parallel with the block division boundary line is a shaded pixel, and the outline direction component of the whole shaded pixel is The number of (outer contour direction components) is totaled for each contour direction (0 to 3 in the example). In the block divided by the horizontal block dividing unit, scanning is performed in two directions from the left end to the right and from the right end to the left, and in the block divided by the vertical block dividing unit, from the upper end to the downward direction,
Scan in two directions from the bottom to the top.

The contour direction normalizing unit 37 outputs the number of contour direction components for each block and contour direction calculated by the contour direction calculating unit 36 from the horizontal block dividing unit 32 or the vertical block dividing unit 33. The outline direction statistical feature amount is obtained by dividing by the size of the block of the direction aggregation target and normalizing.

In the background distance summing unit 38, the background distance output from the basic feature extracting unit 2 via the route 14
In the blocks divided by the horizontal block dividing unit 32,
The total of the background distance from the left end to the right direction and the total of the background distance from the right end to the left direction are calculated, and in the block divided by the vertical block dividing unit 33, the total of the background distance from the upper end to the downward direction, Calculate the sum of background distances from bottom to top.

In the background distance normalizing section 39, the total of the background distances in the four directions output from the background distance summing section 38 is outputted from the horizontal block dividing section 32 or the vertical block dividing section 33 in the outer direction. The background distance statistical feature amount is obtained by dividing by the size of the block to be aggregated and normalized.

Next, the recognition unit 4 obtains the category to which the character pattern belongs based on the contour direction statistical feature amount, the contour direction statistical feature amount, and the background distance statistical feature amount obtained by the statistical feature extraction unit 3. FIG. 19 shows a case where the recognition unit 4 is configured using a hierarchical neural network after learning is completed. Contour direction statistical features in the first layer neurons,
By inputting the outline direction statistical feature amount and the background distance statistical feature amount and sending a signal to the final layer via the intermediate layer, an output signal corresponding to the similarity of each category is generated from each final layer neuron. The output value and its category are output from the recognizing unit 4 from the larger output signal. In the example of FIG. 19, assuming that the number of output categories is 2, the output of the final layer neuron at the left end is the maximum and the output of the final layer neuron at the right end is the second, so the recognition unit 4 outputs those output values and categories ( Categories 0, 2) are output. It should be noted that the recognition unit 4 may be configured by a hierarchical neural network other than that shown in FIG. 19 or a recognition means that can output the identification category using a statistical feature amount other than the hierarchical neural network.

Next, in the local feature extracting unit 41 of the first structural feature extracting unit 5, based on the background distance output from the basic feature extracting unit 2 via the route 13, the local features for separating similar categories are extracted. For specific characteristics. In FIG. 20 (a), the number of times that the ratio of the background distance viewed from the left end of the circumscribed rectangle to the circumscribed rectangle width in the shaded area (center area) becomes larger than the first threshold value is counted, and finally the number of times is When it exceeds the threshold of 2, it is determined that the center left background is convex. In the similar categories, category 0, category 6, category 8, and category 9, it is not usually determined that the center left background is convex, so the structural feature that the center left background is convex is effective for separating category 3. is there. By using the same extraction method, in FIG. 20 (b), the convexity of the upper right background, which is a powerful extraction material of category 6 (and category 5), is determined, and in FIG. It shows that the convexity of the lower left background, which is a powerful extraction material of category 5), is determined. Note that in the actual case, the extraction area / direction of the convex background may be one that analyzes an error of the recognition unit 4 and reduces the error. In addition to the above method, for example, there is a method of determining the curvilinearity (linearity) of the outer stroke from the magnitude of the change in the background distance in the vicinity. For example, FIG. The upper stroke is determined as a straight line because
In FIG. 20 (e), since the magnitude of the change in the background distance is considerably different depending on the place, the upper stroke is determined as a curve, so that the structural feature effective for separating the categories 5 and 6 can be extracted. Further, in order for the first structure identification unit 7 to perform the effective category determination, the extracted structural feature may be divided into handwriting, printing, and handwriting / printing.

The local stroke shape extraction unit 42 traces the contour line in the local region from the contour direction component output from the basic feature extraction unit 2 via the path 14, and draws the contour direction on the traced contour line. By examining the distribution of the components, we obtain local structural features for separating similar categories. In FIG. 21, the contour line is traced from the contour direction component on the contour line in the central shaded area, and it is checked whether the shape of the contour line is convex upward or downward.
Represents the extraction of structural features effective for the separation of In the example of FIG. 21, the shape of the upper and lower contour lines is convex upward in category 6 of FIG. 21 (a), and the shape of the upper contour line is convex downward in category 8 of FIG. 21 (b). , The shape of the lower contour line is determined to be convex upward. As an example of a concrete method of determining the contour line shape, when the reference numerals of FIG. 12 are used as the reference numerals indicating the contour direction components, 5, 6, 7 are on the left side of the local region, 0, 1, 7 are on the center side, and on the right side of the local area. When there are many contour direction component codes of 1, 2, and 3, it is determined that the shape of the contour line is convex, and 1, 2, and 3 at the left side, 0 at the center,
If there are many 1, 7, and 5, 6, 7 on the right side, there is a method of determining that the shape of the contour line is convex downward.

In the first loop extraction unit 43, a stroke of a character pattern is generated based on the cutout image output from the character cutout unit 1 and the contour direction component output from the basic feature extraction unit 2 via the route 14. Extract the loop. A specific extraction method is to trace a contour line from a pixel having a contour direction component and finally return to the tracking start point as a contour line loop, and then refer to the cutout image to refer to the contour line loop. If a white pixel exists inside, it is extracted as a stroke loop. In the example of FIG. 22, if the numeral is the contour direction component code shown in FIG. 12, the code − is a white pixel, and the code + is a black pixel, the shaded code − in FIG. 22 (a) is extracted as a stroke loop. And FIG.
In (b), since there is no sign − inside the contour loop, there is no stroke loop. Here, noise may be removed by providing a limitation that extraction is performed as a loop only when there are white pixels above a certain level.

In the end point / intersection point extraction unit 81 of the end point / intersection point extraction unit 44, adjacent lines are extracted based on the background information of the vertical and horizontal lines output from the basic feature extraction unit 2 via the path 15. The number of white pixel continuous areas and the positional relationship are compared, and the generation point as shown in FIG. 24 (a) or the vanishing point as shown in FIG.
And a branch point as shown in FIG. 24 (d) are extracted, and the endpoint / intersection determination unit 82 removes inappropriate endpoints and intersection candidates from the positional relationship of the endpoints and intersection candidates, Get the final endpoints and intersections. A specific method for extracting the end points and the intersection point candidates in the end point / intersection point candidate extraction unit 81 will be described below. As shown in FIG. 24 (a), when black pixels are generated in the white line in the front line and black pixels in the back line, it is defined as “occurrence”, and as shown in FIG. If all are white pixels in the following line, it is considered to be “disappeared”. Further, as shown in FIG. 24 (c), when two black pixel continuous areas on a certain line (front line) are merged into one on the rear line, it is regarded as “merge”, and conversely, FIG. When one black pixel continuous area is branched into two as in d), it is determined as “branch”. The points classified as “merging”, “branching”, “occurrence”, and “disappearance” are set as end points and intersection candidates, and are output separately for vertical line scanning and horizontal line scanning.

In the end point / intersection point determination unit 82 of the same end point / intersection point extraction unit 44, the positional relationship between the end points and the intersection point candidates output from the end point / intersection candidate extraction unit 81 obtained by scanning in the same direction is examined, If the distances between the end point candidate and the end point candidate and between the end point candidate and the intersection point candidate are short, the end point and the intersection point candidate are removed. For example, for a character pattern with a large circumscribed rectangle (eg, a height and width of 50 pixels or more), an end point candidate of the same type as another end point candidate (“occurrence”) as shown in FIG.
If so, the distance between "occurrence" and "disappear" is very close, and there is an intersection candidate between the two points (if the endpoint is "occurrence", "merge", if "disappear" is "branch". ) Or there is an intersection candidate on a line near and immediately adjacent to an end point candidate as shown in FIG. 25 (b), and another kind of end point (“occurrence”, “disappear”, If there is no "occurrence" if "disappeared", remove the end points and intersections. The remaining end point candidates are determined as end points, and the intersection candidate is determined as an intersection point.

In the straight line extraction unit 45, the black pixel continuous region extraction unit 91 detects the black pixels based on the background information of each line in the vertical and horizontal directions output from the basic feature extraction unit 2 via the path 15. In the case where the pixels are continuously present for a predetermined threshold value (black pixel line continuous threshold value) or more, the coordinates of the start point and the end point are recorded. For example, assuming that the continuous threshold value on the black pixel line is 10, FIG.
In (a), the second to fourth lines have a threshold value or more.

In the straight line candidate extraction unit 92, from the coordinates of the start point and the end point of the black pixel continuous area output from the black pixel continuous area extraction unit 91, the number of lines of the continuous black pixel area is equal to or more than a predetermined threshold value (continuous line). A straight line is determined when the lines are continuous with each other (threshold value). For example, if the continuous threshold value on the black pixel line is 10 and the continuous line number threshold value is 3, it is determined that the line in FIG. 26 (a) is a straight line (in the horizontal direction), and the line in FIG. 26 (b) is not determined to be a straight line. By introducing a threshold value for the number of continuous lines and making a judgment, it is possible to consider the thickness of a handwritten line and remove the influence of a straight line other than the original line due to a frame or noise as shown in FIG. .

The background block extraction unit 46 detects a background block by combining adjacent backgrounds of the same degree from the background information for each line calculated by the background information calculation unit 24. When scanning from the left end of the circumscribed rectangle, as shown in FIG. 23, the first block and the second block are detected for the 0th-order background, and the first block is detected for the primary background. At this time, the number of pixels forming the background block is calculated as the size information of the background block, and the upper left and lower right coordinates of the rectangle circumscribing the background block are calculated as the position information.

The background label extracting section 47 detects the background label by combining adjacent white pixels from the background information for each line calculated by the background information calculating section 24. When scanning from the left end of the circumscribed rectangle, as shown in FIG.
Two background labels are detected and a label number is given. At this time, the number of pixels forming the background label as the label information, the upper left and lower right coordinates of the rectangle circumscribing the background label as the position information, and the presence or absence of contact with each side of the character circumscribed rectangle are calculated.

Next, in the character type judging unit 6, the character cutting unit 1
From the width and height of the circumscribing rectangle of the character pattern output from the path 11 to 11, it is determined whether the character type is handwritten or printed by the size of the height variation. Specifically, in the case of a transfer slip, the characters entered in the same column are basically the same font, so check the height variation of the entire characters entered in the same column, and if the height variation is large. Is determined to be a handwritten character, and if the variation is small, it is determined to be a printed character. In other cases, the variation of the heights of all the characters existing in the area corresponding to the same column of the slip is checked and the same determination is performed. An example of an area corresponding to the same field on the slip is one line (or one column) of a document or postcard address field.
There is.

Next, in the first circumscribing rectangle category determining section 61 of the first type processing section 51 of the first structure identifying section 7, one category depending on the size of the circumscribing rectangle output from the character cutting section 1 is used. If there is a high possibility that
When the category is included in the categories output from the recognition unit 4, the category is determined. As a specific example, category 1 as shown in FIG.
The width of the circumscribed rectangle is extremely narrower than that of category 5 in FIG. 27 (b), and the numbers in other categories are also the same as those in FIG. 27 (b). There is a method to forcibly correct narrow numbers to category 1.

In the first recognition-dependent feature selection section 62, the first
Of the plurality of structural features output from the first structural feature extraction unit 5 based on the category output from the recognition unit 4 for the character that cannot be determined by the circumscribed rectangle category determination unit 61 of Select a feature. As a concrete example,
The category 6 on the verge of closing the background in the upper right part of FIG. 28 has a strong tendency to be erroneously recognized as the category 8 by the recognition method using the statistical features, and thus the recognition unit 4 determines that the category 8 is the first place. In this case, the structural feature extracted by the local stroke shape extraction unit 42 and the loop extraction unit 43 is selected, and the structural feature shown in FIG. 21 is selected from the plurality of structural features extracted by the local stroke feature 42. To do. In addition, when the recognition unit 4 outputs the categories 3 and 9 in order from the top of the recognition, the structural features output from the local background feature extraction unit 41 are shown in FIG. 20 (a) (b) (c). Select.

The first selected structural feature identifying section 63 makes a final category determination based on the structural feature selected by the first recognition-dependent feature selecting section. As a concrete example,
For the characters in FIG. 28, the reliability of the judgment is enhanced by using the selected local stroke feature and loop feature of FIG. When the determination results are the same in both determination methods, the category is determined, and when the determination results are different, the category determined by the recognition unit 4 is determined. If one loop feature is used and the number of loops is 0, then category 3
Estimated. If the number of loops is 1, it is estimated to be category 9 if the loop position is above the center, and category 6 if it is below the center. If the number of loops is 2, it is estimated to be category 8. 2 Using the local stroke feature, if both the upper and lower strokes in the center are convex upward, it is estimated as category 6. If the upper stroke is convex downward and the lower stroke is convex upward, it is presumed to be category 3 or category 8. If both upper and lower strokes are convex downward, it is presumed to be category 9.

With respect to the characters shown in FIGS. 20 (a) (b) (c), the first feature of FIG. 20 (a) is that the center left background is convex.
It is checked whether or not it has been extracted by the structural feature extraction unit 5 of. If it has been extracted, the category 3 is finally decided, and if not, it is next checked whether or not the convex portion of the lower left background of FIG. 20 (c) is extracted. If it has been extracted, the final decision is made to category 9, and if not, it is then determined in FIG.
The same survey is conducted for 0 (b). If the features of FIG. 20 (b) have been extracted, the category 6 is finally determined, and if not, the category 8 is finally determined. Further, in the case of category 5 and category 6 in order from the highest recognition level, it is checked whether the convexity of the lower left background of FIG. 20 (c) or the linearity of the upper stroke of FIG. A specific category. For other categories as well, the features extracted by the first structural feature extraction unit 5 are selected based on the categories output from the recognition unit 4 in the same manner to determine the final category.

Depending on the category output from the recognition unit 4, the first character processing unit 51 may not perform the processing, and the category of the recognition result output from the recognition unit 4 may be used as the final category. When the recognition unit 4 outputs a recognition pending signal and the first type processing unit 51 finds that the structural feature does not apply to any of the structural features, the final category is not determined and the reject (read reject) is performed. ).

Next, in the second circumscribing rectangle category determining section 71 of the first handwriting processing section 52 of the first structure identifying section 7, one category depending on the size of the circumscribing rectangle output from the character cutting section 1 is used. If there is a high possibility that the category can be determined, the category is determined as the category if the category is included in the categories output from the recognition unit 4. The second recognition-dependent feature selecting section 72 of the first handwriting processing section 52.
Then, when the character type determination unit 6 determines that the character pattern is handwritten, among the plurality of structural features output from the first structural feature extraction unit 5 based on the category output from the recognition unit 4. Select valid structural features. As a specific example, handwritten category 3 has a strong tendency to be erroneously recognized as category 8 in the recognition method using the statistical feature, so that the first candidate (category of maximum similarity) is category 8 in recognition unit 4, When it is determined that the second candidate (the category having the second degree of similarity) is the category 3, the background block extracting unit 4
6 and the structural features extracted by the loop extraction unit 43 are selected. Further, when the recognition result is obviously inconsistent based on the number of loops, the size, and the position, the structural feature extracted by the first loop extracting unit 43 is selected. In the second selection structural feature identifying section 73, the second recognition-dependent feature selecting section 7
A final category decision is made based on the structural features selected in 2.

As a concrete example, handwritten characters of category 2 and category 3 represented in FIGS. 29 (a) and 29 (b) are likely to be confused with each other, and the recognition result of the recognition unit 4 is 1
Often, the rank is category 2 and the second rank is category 3, or vice versa. In the case where the recognition result is such and the handwriting determination is performed by the character type determination unit 6, determination is performed based on the difference in the number and position of horizontal generation points (end points) output from the end point / intersection point extraction unit 44. For example, if there are two occurrence points, one in each of the upper and lower areas, then category 2 is determined, and if there are one in each of the upper, middle, and lower areas, and there are three occurrence points, then category 3 is determined. Also, FIG.
There are many handwritten categories in the shape of (a) ¥, handwritten categories of the shape in FIG. 30 (b), and handwritten categories of the shape in FIG. 30 (c), and these shapes are similar. In the recognition unit 4, the category \ in FIG. 30 (a) is the category 7
It is often erroneously recognized as a category 9 or category 9. Such an error can be separated from the category 7 and the category 9 by extracting the two horizontal lines existing in the central portion of the category \ with the straight line extraction unit 45. Also, the category 6 on the verge of closing the background in the upper right part of FIG.
Since the recognition method using the statistical features has a strong tendency to be erroneously recognized as the category 8, if the recognition unit 4 determines that the first place is the category 8, the local stroke shape extraction unit 4
2 and the structural feature extracted by the loop extraction unit 43 are selected, and the structural feature of FIG. 21 is selected from the plurality of structural features extracted by the local stroke feature 42. Further, when the recognition unit 4 outputs the categories 3 and 9 in order from the higher-ranked recognition, the structural features output from the local background feature extraction unit 41 are shown in FIG. 20 (a) (b) (c). Select.

Further, when the recognition result of category 3 is category 8 as shown in FIG. 31 (a),
When the number of loops, the size of the horizontal 0th order background block, and the maximum horizontal 0th order background length are focused on from the selected structural feature amount, and the following conditions (1), (2), (3), and (4) are satisfied Is
The recognition result is corrected to the second candidate category 3 to be the final recognition result. (1) The first candidate of the recognition result is 8 and the second candidate is 3 (2) The number of loops is 0 or 1 (3) Horizontal 0th-order background block area / circumscribed rectangular area>
Threshold 1 (4) Maximum horizontal 0th background length / circumscribing rectangle width> Threshold 2

If the recognition result of category 7 is category 9 as shown in FIG. 31 (b),
Paying attention to the number of loops and the number of vertical primary background blocks from the selected structural feature amount, the following (1) (2) (3)
When the condition of is satisfied, the recognition result is corrected to the category 7 of the second candidate and is set as the final recognition result. (1) The first candidate of the recognition result is 9, and the second candidate is 7 (2) The number of loops is 0 (3) The number of vertical primary background blocks is 1

As another example, if a category such as categories 1 and 7 in which no loop clearly exists is recognized as a category in which a loop exists in normal writing, the category of the second candidate. In addition, even if the size of the loop is very large, if it is recognized in categories 6 and 9, the recognition result is corrected to category 0.

For other categories, an appropriate feature amount is selected from the features extracted by the first structural feature extracting unit 5 based on the category output from the recognizing unit 4, and the second feature is selected.
The final structural feature discriminator 73 determines the final category. Depending on the category output from the recognition unit 4, the first handwriting processing unit 52 may not perform the process, and the category of the recognition result output from the recognition unit 4 may be used as the final category. If the recognition pending signal is output from the recognition unit 4 and it is determined by the first handwriting processing unit 52 that it does not apply to any structural feature, the final category is not determined and reject (read reject) is performed. )
It doesn't matter.

As described above, according to the first embodiment, the basic feature extraction unit 2 extracts the basic features that can be used as both the statistical features and the structural features, thereby making it possible to make the configuration common. In addition, the first structural feature extraction unit 5 mainly extracts structural features that are effective for separating similar categories that are difficult to be identified by the recognition unit 4, thus making it possible to improve the efficiency of the process and improve the statistical features. Based on the recognition result output from the recognition unit 4 that performs identification using the first structure identification unit 7
It is possible to further improve the processing efficiency and improve the recognition performance of similar categories by selecting the structural features required in and determining the final category.

(Second Embodiment) FIG. 33 shows a second embodiment of the present invention.
3 is a block diagram of an optical character reading device according to an embodiment of FIG. In FIG. 33, a character cutout unit 1, a basic feature extraction unit 2, a statistical feature extraction unit 3, a recognition unit 4, a character type determination unit 6, a first structure identification unit 7, a route 11, a route 12, a route 13, a route. 14, the route 15 is the above-mentioned first
The operation is the same as that in the above embodiment. The configuration of this embodiment different from that of FIG. 1 is that a second structural feature extraction unit 8 is provided in place of the first structural feature extraction unit 5.

FIG. 34 is a block diagram specifically showing the configuration of the second structural feature extraction unit 8. In FIG. 34, a local background feature extraction unit 41 and a local stroke shape extraction unit 4
2, background block extraction unit 46, background label extraction unit 47,
It is similar to that in the first embodiment described above,
The operation is similar. The configuration of this embodiment different from that of FIG. 4 is that a second loop extracting unit 48 is provided in place of the first loop extracting unit 43.

The operation of the optical character reading device configured as described above will be described below. The operation other than the operation of the second loop extraction unit 48 is the same as that of the first embodiment, and thus the description thereof is omitted. The second loop extraction unit 48 extracts the loop generated by the stroke of the character pattern based on the label information output from the background label extraction unit 47.
As a specific extraction method, a background label that does not touch any side of the character circumscribed rectangle is extracted as a stroke loop.
In the example of FIG. 35, the numbers are the label numbers of the background labels,
When the symbol + is a black pixel, the shaded background label of label number 3 in FIG. 35 is extracted as a stroke loop. Here, noise may be removed by providing a limitation that extraction is performed as a loop only when there are white pixels above a certain level.

As described above, according to the second embodiment, the basic feature extraction section 2 extracts a basic feature that can be used as both a statistical feature and a structural feature, thereby enabling a common configuration. In addition, the second structural feature extraction unit 8 extracts structural features that are effective for separating similar categories that are difficult to be identified by the recognition unit 4 to enable efficient processing, and statistical features. Based on the recognition result output from the recognition unit 4 that performs identification using the first structure identification unit 7
It is possible to further improve the processing efficiency and improve the recognition performance of similar categories by selecting the structural features required in and determining the final category. Further, by extracting the number of loops based only on the label information, the efficiency of processing can be improved.

(Third Embodiment) FIG. 36 is a block diagram of an optical character reading device according to a third embodiment of the present invention. In FIG. 36, a character cutout unit 1, a basic feature extraction unit 2, a statistical feature extraction unit 3, a first structural feature extraction unit 5, a character type determination unit 6, a route 11, a route 1
2, the route 13 and the route 14 are similar to those in the above-described first embodiment, and the operation is also similar. The structure of this embodiment different from that of FIG. Fix it,
This is the point that the second structure identification unit 9 that determines the final category from the magnitude of the modified similarity is provided.

FIG. 37 is a block diagram specifically showing the structure of the second structure identifying section 9. Reference numeral 371 denotes a second type processing unit, which is the first type when the type determining unit 6 determines that the type is a type.
The similarity correction value is obtained based on the structural feature output from the structural feature extraction unit 5 and at least one category output from the recognition unit 4. Reference numeral 372 denotes a second handwriting processing unit, and when the character type determination unit 6 determines that the handwriting is performed, the structural feature and recognition unit 4 output from the first structural feature extraction unit 5 is output.
The similarity correction value is obtained based on at least one category output from. An adder 373 adds the similarity of a plurality of categories from the highest similarity output from the recognition unit 4 and the similarity correction value output from the second type processing unit 371 or the second handwriting processing unit 372. Then, the modified similarity is obtained. A correction determination unit 374 outputs the category number with the maximum correction similarity based on the magnitude of the correction similarity output from the adder 373.

The operation of the optical character reading device configured as described above will be described below. The recognition unit 4 obtains the similarity for each category based on the contour direction statistical feature, the contour direction statistical feature, and the background distance statistical feature obtained by the statistical feature extraction unit 3, and determines the similarity of a plurality of categories in order from the highest similarity. Output the degree and at least one category number. Next, the second type processing unit 37 of the second structure identification unit 9
In the case of No. 1, if the character type determination unit 6 determines that the type is
The similarity correction value is obtained based on the structural feature output from the structural feature extraction unit 5 and at least one category output from the recognition unit 4. A specific method of calculating the similarity correction value will be described with reference to FIG. First, all the similarity correction values are set to 0. Next, when two categories, category 3 and category 9, are output from the recognition unit 4 in order from the top of the recognition, among the structural features output from the first structural feature extraction unit 5, FIG. ) (b) (c) is selected, and first, FIG.
It is checked whether or not the feature in which the center left background of (a) is convex is extracted by the first structural feature extraction unit 5. If it has been extracted, the category 3 similarity correction value is increased. If not extracted, the similarity correction value of category 3 is lowered, and it is checked whether or not the convex on the lower left background of FIG. 20 (c) is extracted. If it has been extracted, the similarity correction value of category 5 and category 9 is increased. If not extracted, the similarity correction value of category 5 and category 9 is lowered. Moreover, it is checked whether or not the convex portion of the upper right background of FIG. 20 (b) is extracted. If it has been extracted, the similarity correction value of category 5 and category 9 is increased, and if it has not been extracted, the similarity correction value of category 5 and category 9 is decreased. The upper and lower widths of the similarity correction value in each case do not have to be constant, and the importance of the structural feature, the distribution of the similarity output from the recognition unit 4, and the first structural feature extraction unit 5 It may be changed according to the certainty factor of the extraction of the structural feature of. Further, the category for which the similarity correction value is obtained may be limited to only the categories corresponding to the plurality of similarities output from the recognition unit 4.

In the second handwriting processing section 372 of the second structure identifying section 9, when the character type judging section 6 judges that the character is handwritten, the structural characteristics output from the first structural characteristic extracting section 5 are The similarity correction value is calculated based on at least one category output from the recognition unit 4. A specific method of calculating the similarity correction value is the same as that of the second type processing unit 371. In the adder 373, the plurality of similarities output from the recognition unit 4 and the second type processing unit 371 or the second handwriting processing unit 372 are used.
The corrected similarity is calculated by adding the similarity correction values output from The modification determining unit 374 compares the modification similarities output from the adder 373 for each category and outputs the category having the maximum modification similarity, and determines the final category. If the maximum correction similarity output from the correction determining unit 374 is low, or if the difference between the maximum correction similarity and the second or lower correction similarity is small, final category determination is not performed, It may be rejected (read refusal).

As described above, according to the third embodiment, the similarity output from the recognition section 4 and the second type processing section 37.
By determining the final category in the correction determination unit 374 based on the corrected similarity calculated by adding the similarity correction value output from the first or second handwriting processing unit 372, the similarity output from the recognition unit 4 is determined. The subtle difference in degree can be reflected in the final category determination, and the reliability of structural feature detection in the second print processing unit 371 and the second handwriting processing unit 372 can be reflected as a similarity correction value. As a result, the recognition performance is improved.

(Fourth Embodiment) FIG. 38 shows a third embodiment of the present invention.
3 is a block diagram of an optical character reading device according to an embodiment of FIG. In FIG. 38, a basic feature extraction unit 2, a statistical feature extraction unit 3, a recognition unit 4, a first structural feature extraction unit 5, a character type determination unit 6, a first structure identification unit 7, and a route 1
1, the route 12, the route 13, the route 14, and the route 15 are the same as those in the above-described first embodiment, and the operations are also the same. The first structure identification unit 7 may be replaced with the second structure recognition unit 9.

The configuration of this embodiment different from that of FIG. 1 is that instead of the character slicing section 1, individual character patterns are cut out from the photoelectrically converted character string data, and the size of the circumscribed rectangle with respect to each character pattern is set. A second character cutout unit 381 is provided that outputs a cutout image and, when a recutout instruction signal is output from a first cutout error determination unit 382, which will be described later, recuts the character of the number for which the cutout error has been determined. And the size of the circumscribing rectangle of the individual character patterns output from the second character cutout unit 381, the cutout error is determined, and in the case of the cutout error, the recutting instruction signal and the cutout error character number are set to the second value. The point is that a first cutout error determination unit 382 for outputting to the character cutout unit 381 is provided.

The operation of the optical character reading device configured as described above will be described below. In the second character cutout unit 381, the character cutout unit 1 is normally used.
Similarly, the original image including a plurality of characters is cut out for each character, and the size of the circumscribed rectangle of each character and the cutout image are output. If the first cutout error determination unit 382 detects a cutout error, a different operation is performed, and the operation will be described later.

Next, the first cut-out error judgment section 382 judges a cut-out error based on the size of the circumscribed rectangle of each character output from the second character cut-out section 381. A specific operation will be described with reference to FIG.
39 (a) is the original image, FIGS. 39 (b) to 39 (e) are the sizes of the circumscribed rectangles of the clipped characters, and FIGS. 39 (f) to 39 (i) are the clipped images. When the original image of FIG. 39 (a) is first cut out by the second character cutout unit 381 and the character frame remains unseparated as shown in FIG. 39 (f), the character frame is removed as shown in FIG. 39 (g). When part of the protruding characters is deleted, when the character frame is cut out as one character as shown in FIG. 39 (h), two adjacent characters are cut out as one character as shown in FIG. 39 (i). Sometimes. Since the size of the circumscribing rectangle of these clipped character images is often different from the size of the circumscribing rectangle for one character as originally shown in FIG. 39 (b) to (e), When the size of the circumscribed rectangle of the character is input to the first cutout error determination unit 382 and a character far from the height or width of the circumscribed rectangle of the character in the same field is found, or when it is determined in advance. If it is out of the range of the height or width of the circumscribed rectangle, a recutting instruction signal, an error character number, and an error content are output to the second character cutout unit 381 as a cutout error. In the second character cutout unit 381, when the recutout instruction signal is input from the first cutout error determination unit 382, the cutout area of the simultaneously input error character number is recut out according to the error content. For example, in FIG. 39 (b), since the error content is that the width of the circumscribing rectangle is large, the clipping position is re-examined based on the horizontal / vertical black pixel histogram as shown in FIG. 39 (c) to (e), the cutout position is reexamined in the same manner.

As described above, according to the fourth embodiment, based on the size of the circumscribing rectangle of the character pattern output from the second character cutout unit 381, the first cutout error determination unit 382 fails in the cutout. The character is detected and a re-cutout instruction signal is sent to the second character cutout unit 381,
By analyzing the cause of the cutout failure in the second character cutout unit 381 and executing the recutout process, it is possible to reduce the cutout error, and the error in the recognition unit 4 and the first structure identification unit 7 due to the cutout error is reduced. Perception can be reduced.

(Fifth Embodiment) FIG. 41 shows the fifth embodiment of the present invention.
3 is a block diagram of an optical character reading device according to an embodiment of FIG. In FIG. 41, the second character cutout unit 381, the basic feature extraction unit 2, the statistical feature extraction unit 3, the recognition unit 4, the structure first artificial feature extraction unit 5, the character type determination unit 6, and the recognition first The structure identifying unit 7, the route 11, the route 12, the route 13, the route 14, and the route 15 are the same as those in the above-described fourth embodiment, and the operations are also the same. In addition,
The first structure identification unit 7 may be replaced with the second structure recognition unit 9.

The configuration of the embodiment of the present invention different from that of FIG. 38 is that instead of the first cutout error judgment unit 382, the size of the circumscribed rectangle of each character pattern output from the second character cutout unit 381 and the recognition unit. A second cutout error determination unit that determines a cutout error from the similarities of the plurality of categories output from FIG. 4, and outputs a recutout instruction signal and a cutout error character number to the second character cutout unit 381 in the case of a cutout error. That is, 411 is provided.

The operation of the optical character reading device configured as described above will be described below. The second cut-out error determination unit 411 uses the same method as the first cut-out error determination unit 382 to determine a cut-out error based on the size of the circumscribing rectangle of each character pattern. The difference is that the cutout error is determined by using the similarity of the categories. In particular,
If the first similarity is extremely low compared to other characters, or if the difference between the first similarity and the second or lower similarity is small, a clipping error may occur. It is determined that the property is high, and the re-cutout instruction signal and the cutout error character number are output to the second character cutout unit 381.

As described above, according to the fifth embodiment, the size of the circumscribing rectangle of the character pattern output from the second character cutout unit 381 is also determined by the output of the recognition unit 4 to the second cutout error determination unit. By adding to the determination criterion of the cutout error in 411, the cutout error can be determined based on the certainty of the recognition result that can be determined from the similarity in the recognition unit 4 even when the cutout error cannot be clearly determined from the size of the circumscribed rectangle. .

(Embodiment 6) FIG. 42 shows a sixth embodiment of the present invention.
3 is a block diagram of an optical character reading device according to an embodiment of FIG. In FIG. 42, a character cutout unit 1, a basic feature extraction unit 2, a statistical feature extraction unit 3, a recognition unit 4, a first structural feature extraction unit 5, a character type determination unit 6, a first structure identification unit 7, Route 11, Route 12, Route 13, Route 1
The route 4 and the route 15 are the same as those in the above-described first embodiment, and the operations are also the same. The first structure identification unit 7 may be replaced with the second structure recognition unit 9.

The configuration of this embodiment different from that of FIG. 1 is that the ratio of the second largest similarity degree to the largest similarity degree output from the recognition unit 4, the magnitude of the largest similarity degree, and the first structure. The point is that a first recognition result calculation unit 421 for determining the presence / absence of a reject is provided based on the background complexity calculated based on the background information calculated by the background information calculation unit 24 of the dynamic feature extraction unit 5.

The operation of the optical character reading device configured as described above will be described below. In the same manner as in the first embodiment, the recognition unit 4 outputs a plurality of category similarities to the input character string data, and at least one category from the larger one. Two category similarities are selected from the larger one of the plurality of category similarities output from the recognition unit 4, and the value of the ratio of the second similarity and the maximum similarity is set to the preset threshold value a. Is larger than the value (that is, the second similarity is close to the maximum similarity), or the value of the maximum similarity is smaller than the value of the preset threshold value b,
If the background complexity, which is the sum of the maximum values of the background orders in the vertical and horizontal directions, is larger than the value of the preset threshold value c, a reject signal is output; otherwise, the recognition unit 4 outputs the reject signal. The calculated category similarity and category are output to the first structure identification unit 7. The first structure identification unit 7 determines the final recognition result by using the structural feature as in the first embodiment. Note that the background complexity is not limited to the sum of the maximum values of the background orders in the vertical and horizontal directions, and may be based on the background information such as the sum of the number of lines of the background orders having a certain value or more.

In this embodiment, the case where the maximum similarity value and the second ratio between the similarity and the maximum similarity are used will be described. Here, the rejection determination threshold values a and b are, as shown in FIG. 43, the ratio of the second highest similarity degree (second highest similarity degree) to the maximum similarity degree for a plurality of character data in advance, It is empirically determined by obtaining the maximum similarity value and plotting it on the graph for each correct and incorrect answer of the recognition result. In the case of FIG. 43, only incorrect characters are distributed in the region where the maximum similarity is small, regardless of the value of the ratio of the similarities. For example, when a character pattern (147 characters) having a maximum similarity value of 500 or less is examined, most (144 characters) are preprocessing errors. FIG. 44 lists an example of characters whose maximum similarity value is 500 or less. From FIG. 44,
Preprocessing error characters such as noisy characters, remaining frame lines, clipping of amount correction part, concatenated character separation error, clipping of unrecognized characters, etc. are added by adding the maximum similarity to the criteria for reject judgment. You can see that you can reject.

As described above, according to the sixth embodiment, not only the value of the ratio of the similarity but also the magnitude of the maximum similarity is used together in the rejection determination to efficiently reject the preprocessing error. It becomes possible to do it. Further, if the background complexity is used, it is possible to perform the rejection more efficiently.

(Embodiment 7) FIG. 45 shows a seventh embodiment of the present invention.
3 is a block diagram of an optical character reading device according to an embodiment of FIG. In FIG. 45, a character cutout unit 1, a basic feature extraction unit 2, a statistical feature extraction unit 3, a recognition unit 4, a first structural feature extraction unit 5, a character type determination unit 6, a first structure identification unit 7, Route 11, Route 12, Route 13, Route 1
The route 4 and the route 15 are the same as those in the above-described sixth embodiment, and the operations are also the same. The first structure identification unit 7 may be replaced with the second structure recognition unit 9.

The configuration of this embodiment different from that of FIG. 42 is that instead of the first recognition result calculation unit 421, the ratio of the second largest similarity degree to the maximum similarity degree output from the recognition unit 4,
A second recognition result calculation unit 451 that determines the presence or absence of a rejection based on the maximum similarity and the background complexity based on the background information calculated by the background information calculation unit 24 of the first structural feature extraction unit 5. That is the point.

FIG. 46 is a block diagram specifically showing the structure of the second recognition result calculation unit 451. A similarity ratio calculation unit 461 calculates two ratios of the second category similarity (the second highest similarity) and the maximum similarity by selecting two categories having the highest category similarity output from the recognition unit 4. Is. 46
Reference numeral 6 denotes a background complexity calculation unit that calculates the background complexity based on the background information output from the background information calculation unit 24. Four
Reference numeral 62 denotes a reject determination unit that inputs the calculated similarity ratio, the maximum similarity and the background complexity, and outputs whether the character pattern is a correct answer (non-reject) or an incorrect answer (reject) according to these values. Is. Here, an example will be described in which the reject determination unit is configured by a hierarchical neural network in which a plurality of data has been learned in advance. Reference numeral 463 is a correct answer category table that stores correct answer categories of character data used for learning of the reject determination unit. 464, during learning, the presence / absence of learning is determined based on the maximum similarity category output from the recognition unit 4, the correct answer category written in the correct answer category table, and the output of the reject determination unit, and during recognition, the output of the reject determination unit Is an output determination unit that determines rejects. Reference numeral 465 is a connection weight update unit that updates the connection weight of the hierarchical neural network that constitutes the reject determination unit when the output determination unit 464 determines that learning is required during learning.

The operation of the optical character reading device configured as described above will be described below. In the same manner as in the first embodiment, the recognition unit 4 outputs a plurality of category similarities to the input character string data, and at least one category from the larger one. Second recognition result calculation unit 451
Includes a learning mode for learning the relationship between the similarity, the background complexity, and the presence / absence of a rejection for a plurality of character data prepared in advance, and a recognition mode for performing a rejection determination based on the learning result.

First, the learning mode will be described. The similarity ratio calculation unit 461 calculates the ratio between the second-rank similarity and the maximum similarity from the category similarity output from the recognition unit 4 and outputs the ratio to the rejection determination unit 462. Background complexity calculator 4
66 calculates the sum of the maximum values of the background orders in the vertical and horizontal directions as the background complexity based on the background information output from the background information calculation unit 24, and outputs the background complexity to the reject determination unit 462. Note that the background complexity is not limited to the sum of the maximum values of the background orders in the vertical and horizontal directions, and may be based on the background information such as the sum of the number of lines of the background orders having a certain value or more. The hierarchical neural network forming the reject determination unit 462 inputs the similarity ratio, the maximum similarity ratio, and the background complexity, and outputs an output signal from the output layer according to the connection weight at that time. Here, the number of neurons in the output layer is 2, one output represents reject, and the other reject represents non-reject. Output determination unit 4
64 selects the larger one of the outputs of the reject determination unit and, at the same time, compares the maximum similarity category of the currently input pattern with the correct category output from the correct category table 463. When the maximum category matches the correct category (correct answer), the reject determination unit 4
If the output of the neuron representing the reject of the output of 62 is large or the maximum similarity category does not match the correct category (incorrect answer), and the output of the neuron representing the non-reject of the output of the reject determination unit 462 is large, learning is performed. The signal is output to the connection weight updating unit. When the learning signal is output from the output determining unit 464, the coupling weight updating unit uses the algorithm such as the steepest descent method to reject the reject determining unit 46.
The connection weights of the hierarchical neural network forming 2 are updated. Whether the learning signal is not output from the output determining unit 464 for all the learning data
The process is repeated until the magnitude relationship between the two output values of the reject determination unit does not change. With this learning mode, it is possible to perform reject determination according to the similarity ratio, the maximum similarity and the background complexity.

Next, the recognition mode will be described. When the reject determination unit 462 outputs an output according to the similarity ratio, the magnitude of the maximum similarity, and the background complexity based on the combined weight value after learning, the output determination unit 464 determines that the output value is larger. Is selected, the reject signal is output when the output indicating the reject is large, and the maximum similarity category is output as the recognition result when the output indicating the non-reject is large. The boundary between the correct answer character (unrejected character) area and the incorrect answer character (character to be rejected) area in the space of the similarity ratio, the maximum similarity and the background complexity is very complicated. For example, it is considered that the linear separation is impossible even in the two-dimensional space of the similarity ratio and the maximum similarity as shown in FIG. 43. However, according to the present embodiment, the rejection is performed by the non-linear separable hierarchical neural network. By configuring the determination unit 462 and learning this boundary, more accurate and effective reject determination can be performed. The reject determination unit 462 may be configured with a neural network model other than the hierarchical neural network. Further, in the present embodiment, three of the similarity ratio, the maximum similarity, and the background complexity are input to the reject determination unit 462, but any one or two of the three may be input. I don't mind.

(Embodiment 8) Next, in a document image including frame ruled lines and characters such as a form, the structure of the frame ruled lines is recognized, the form is identified, and a specific character area filled in the form is specified. A form document recognition method and a form document recognition device for recognizing the characters described therein, and a form automatic processing system for converting the recognized characters into an identification code for processing will be described.

First, regarding the conventional form document recognition device,
This will be described with reference to FIG. First, the image input means 501
The document image written on the form is converted into a binary image by. Then, the character position detecting means 502 extracts a desired character area from the image signal. As a method of detecting the character position, there is a method of identifying the type of the form based on the difference in the local frame line structure and reading the previously designated area, as disclosed in Japanese Patent Laid-Open No. 2-217977. Next, the character slicing means 503 cuts out the character by character, and the character recognizing means 504 recognizes the character. Next, in the word recognition means 505, a word dictionary 506 in which words are stored in advance for the character string recognized by the character recognition means 503.
A word is read out from and the final recognition result is obtained by comparing and matching the read word with the recognized character string. For example, when this character string is an address, the address name to be read is registered in the address dictionary in advance, and the character recognition result is compared with the address name read from the address dictionary to recognize the address name. For the method, refer to Japanese Patent Laid-Open No. 2-1
It is disclosed in Japanese Patent No. 81287.

However, in such a conventional form document recognition apparatus, since all the words to be read must be registered in the word dictionary, the number of cases to be compared and compared is very large and the processing speed is very slow. There was also a problem.

Therefore, in order to solve such a conventional problem, the eighth embodiment aims to provide a form document recognizing method and device and a form automatic processing system which have a high processing speed and a high accuracy. By registering the frame structure feature of each form and the candidate character string described in the form having the frame structure feature in the frame structure reference table and checking the frame structure feature by referring to this table, the form type can be determined. Roughly classify and detect the character reading target area in the form, then cut out the character area, recognize the cut out character, and collate the recognition result with the candidate character string registered in the frame structure reference table for final recognition As a result, compared to the conventional method using the word dictionary, it is possible to recognize the form document more quickly and with higher accuracy.

The eighth embodiment of the present invention will be described below with reference to FIGS. 47 to 50. 47 is a block diagram of a form document recognition device according to the eighth embodiment of the present invention. In FIG. 1, reference numeral 101 denotes an image input means, which reads a form including frame ruled lines and characters and converts it into a binary image. An image memory 102 stores the binary image. Reference numeral 103 denotes a frame ruled line extracting means, which is an image memory 1
The frame ruled lines are extracted from the binary image stored in 02. Reference numeral 104 denotes a frame structure extracting means, which extracts frame structure features necessary for identifying the form of the form from the frame ruled lines extracted by the frame ruled line extracting means 103. Reference numeral 105 denotes a frame structure reference table in which the frame structure features of individual forms and candidate character strings described in the form having the frame structure features are registered in advance. Reference numeral 106 denotes a frame structure collating unit, which refers to the frame structure reference table 105 to collate the frame structure features extracted by the frame structure feature extracting unit 104 to identify the type of the input form, and to read the character reading target area. Is to detect. Reference numeral 107 denotes a character cutout unit corresponding to the character cutout unit 1 in Embodiments 1 to 7 described above.
It cuts out the characters one by one. Reference numeral 108 denotes a character recognition means constituted by means after the character cutout unit 1 in the first to seventh embodiments, and the character cutout means 107 is provided.
It recognizes the characters cut out by. A post-processing unit 109 collates the recognition result recognized by the recognition unit 108 with a candidate character string corresponding to the input form registered in the frame structure reference table 105 to obtain a final recognition result.

With respect to the form document recognition device configured as described above, the operation will be described below in the case where the form shown in FIG. 48 is input and the local government name described on the form is used as the character string to be recognized. I will explain. FIG. 49 is a diagram showing an example of the frame structure reference table 105. It is assumed that there are four types of forms to be input, groups A to D. In the frame structure reference table 105, the structure of the outline frame of the form and the internal structure of each outline frame, the character reading target area, and the name of the local government to be described are registered in advance.

First, the recognition operation will be described. Form 2
When 01 is input, the image input unit 101 binarizes the document image including the frame ruled line and the character string, and stores the binarized image in the image memory 102. Next, the frame ruled line extracting means 103 extracts only the frame ruled lines from the binary image stored in the image memory 102.

Next, in the frame structure feature extraction means 104, the structure of the outer frame necessary for identifying the form of the form from the frame ruled lines extracted by the frame ruled line extraction means 103, and the internal structure of each outer frame. Frame structure features such as. The extracted frame structure feature is analyzed by the frame structure matching unit 106.
The input form 301 is determined to belong to the group A by collating with the frame structure feature registered in advance in the frame structure reference table 105. Next, the character cutout unit 107 cuts out a character in character units from the character reading target area 202 of the input form 201 corresponding to the group A of the frame structure reference table 105. Next, the character recognition means 108 performs character recognition.

Next, the operation of the post-processing means 109 will be described. It is assumed that “Atsugi City, Kanagawa Prefecture” is obtained as the recognition result by the character recognition means 108. In the post-processing means 109, the frame structure reference table 105 is searched, and the local public organization names “Atsugi City, Kanagawa Prefecture”, “Fukuoka City”, and “Minato Ward, Tokyo” registered as candidate character strings described in Group A.
Matches "Suita City, Osaka Prefecture" and "Hayato Town, Kagoshima Prefecture" with the recognition result, "Atsugi City, Kanagawa Prefecture", and outputs "Atsugi City, Kanagawa Prefecture" as the final recognition result. Also, a character cutting error,
Even if the recognition result of the character recognizing means 108 is incorrect due to the influence of noise, for example, “Kanagawa prefecture” and “Atsumoto city, Yuzugawa prefecture” are obtained as the recognition result, the candidate character string is selected. The most probable local government name was selected,
The final recognition result is "Atsugi City, Kanagawa Prefecture".

In the eighth embodiment, a character string recognized as a local public entity is converted into an identification code, for example, a local public entity code, and further converted into a bar code or the like, and the bar code is replaced with the identification code. Or the identification code is printed on the form, the process of the form automatic processing system can be made very efficient. An outline of the form automatic processing system in this case is illustrated in FIG. In FIG. 50, 401 is a form document recognition means, 402 is an identification code conversion means, 403 is an identification code conversion table, 404 is an identification code printing means, and 405 is a processing means.

In this form automatic processing system, the form is
When input to the form document recognition means 401, which is the same form document recognition device as in the first embodiment, the local public organization name written on the form is recognized. Next, the identification code conversion unit 402 converts the recognition result of the form document recognition unit 401 into an identification code by referring to the identification code conversion table. Here, the identification code may uniquely correspond to the local public entity. For example, there is a local government code.
If the recognition result is "Atsugi City, Kanagawa Prefecture", the corresponding local government code is "142123". The identification code is not limited to the local government code, but may be uniquely determined. Next, the identification code printing means 404 prints the converted identification code at a predetermined position on the input form. Instead of the identification code itself, it may be converted into a barcode and the barcode may be printed. Next, in the processing means 405, the printed identification code is read, and the form is journaled based on the identification code.

As described above, according to the eighth embodiment,
By registering the frame structure features of individual forms and candidate character strings described in the form having the frame structure features in the frame structure reference table in advance, and referring to this table to collate the frame structure features, the type of the form Are roughly classified to detect the character reading target area in the form, then the character area is cut out, the cut out character is recognized, and the recognition result is compared with the candidate character string registered in the frame structure reference table to make the final This is the recognition result, and it is possible to recognize the form document more quickly and more accurately than the conventional method using the word dictionary.

[0153]

As described above, according to the present invention,
It has the following effects. According to the optical character reading method of the first aspect of the present invention, the basic features that can be used for both the statistical features and the structural features are extracted to make the processing common, and the statistical features are used. The feature recognition unit enables efficient feature extraction by mainly extracting the structural features that separate similar categories that are difficult to identify. By determining the category, it is possible to further improve the processing efficiency and improve the recognition performance of similar categories.

According to the invention described in claim 2 of the present invention,
Based on structural features, we extracted basic features that can be used for both statistical and structural features to standardize processing, and for similar categories that are difficult to identify by the recognition unit that uses statistical features. By determining the final category based on the corrected similarity after adding the similarity correction value,
It is possible to reflect the subtle difference in the similarity output from the recognition unit using statistical features in the final category determination, and the reliability of structural feature detection as a similarity correction value. Therefore, the recognition performance can be improved.

According to the invention of claim 3 of the present invention,
By performing specialized processing according to the character type, it is possible to use without confusing the structural features peculiar to type and handwriting.
The recognition performance between similar categories is improved, and the processing efficiency can be improved by separating and extracting structural features according to the character type in advance.

According to the invention of claim 4 of the present invention,
Based on the size and the similarity of the circumscribed rectangle of the character pattern, if the circumscribed rectangle is significantly different from the circumscribed rectangle of other characters, or is out of the prescribed circumscribed rectangle, or the similarity is the similarity of other characters. When it is extremely low compared to, or when the difference between the similarity of the first rank and the similarity of the second rank or lower is small,
By detecting the cut-out failure character, analyzing the cause of the failure, and executing the cut-out again, it is possible to reduce the cut-out error and reduce the misrecognition due to the cut-out error.

According to the invention of claim 5 of the present invention,
By performing the rejection determination based on the ratio of the similarities, the magnitude of the maximum similarity, and the background complexity, it is possible to efficiently reject the preprocessing error character that is difficult to reject.

According to the optical character reader of the sixth aspect of the present invention, basic features that can be used as both statistical features and structural features are extracted to achieve common processing, and statistical The feature-based recognition unit enables efficient feature extraction by extracting structural features that separate similar categories that are difficult to identify, and selects the necessary structural features based on the recognition results. By determining the final category, it is possible to further improve the processing efficiency and improve the recognition performance of similar categories.

According to the invention of claim 7 of the present invention,
Based on the recognition results using statistical features, by selecting only the structural features necessary for identifying similar categories and determining the final category, it is possible to further improve the processing efficiency and improve the recognition performance of similar categories. It will be possible.

According to the invention of claim 8 of the present invention,
By performing specialized processing according to the character type, it is possible to use without confusing the structural features peculiar to type and handwriting.
The recognition performance between similar categories is improved, and the processing efficiency can be improved by separating and extracting structural features according to the character type in advance.

According to the invention of claim 9 of the present invention, based on the recognition result using the statistical feature for the type,
By selecting only the structural features necessary for identifying similar categories and determining the final category, it is possible to further improve the processing efficiency and improve the recognition performance of similar categories.

According to the invention of claim 10 of the present invention,
With regard to print characters, the categories that can be distinguished only by the size of the circumscribed rectangle are first judged to efficiently separate the categories, and then the structural features required for similar category separation are selected according to the result of the recognition unit. By performing the final category judgment based on the above, it is possible to efficiently and effectively separate similar categories.

According to the eleventh aspect of the present invention, only the structural features necessary for identifying similar categories are selected based on the recognition result using the statistical features of handwritten characters, and the final category is selected. By determining, it is possible to further improve the processing efficiency and improve the recognition performance of similar categories.

According to the twelfth aspect of the present invention, with respect to handwritten characters, the categories that can be discriminated by only the size of the circumscribing rectangle are first discriminated, so that the categories can be efficiently separated and then recognized. By selecting the structural features necessary for the similar category separation according to the result of the section and making the final category judgment, the efficient and effective similar category separation becomes possible.

According to the thirteenth aspect of the present invention, the basic features that can be used for both the statistical features and the structural features are extracted, the processing is made common, and the statistical features are used. For similar categories that are difficult to identify in the recognizer, the final category is determined based on the modified similarity obtained by adding the similarity correction value based on the structural feature. The subtle difference in the output similarity can be reflected in the final category determination, and the reliability of structural feature detection can be reflected as the similarity correction value, which can improve recognition performance. Become.

According to the fourteenth aspect of the present invention, the final category is determined based on the corrected similarity obtained by adding the similarity correction values. The subtle difference in the output similarity can be reflected in the final category determination, and the reliability of structural feature detection can be reflected as the similarity correction value, which can improve recognition performance. Become.

According to the fifteenth aspect of the present invention, by performing the specialized processing according to the character type, the structural characteristics peculiar to the print / handwriting can be used without confusion, so that it is possible to use the same between the similar categories. The recognition performance is improved, and the processing efficiency can be improved by separating and extracting the structural features according to the character type in advance.

According to the sixteenth aspect of the present invention, the final category is determined for the print type based on the corrected similarity obtained by adding the similarity correction values.
It is possible to reflect the subtle difference in the similarity output from the recognition unit using statistical features in the final category determination, and the reliability of structural feature detection as a similarity correction value. Therefore, the recognition performance can be improved.

According to the seventeenth aspect of the present invention, statistical characteristics are utilized for handwritten characters by determining the final category based on the corrected similarity obtained by adding the similarity correction value. The subtle difference in the similarity output from the recognition unit can be reflected in the final category determination, and the reliability of structural feature detection can be reflected as the similarity correction value. It is possible to improve.

According to the eighteenth aspect of the present invention, as common preprocessing for calculating the contour direction component, calculating the background distance from each side of the circumscribed rectangle to the character pattern, and calculating the background information. By extracting the contour points of the character pattern, the efficiency of basic feature extraction is improved. In addition, by extracting contour points, contour direction components, background distances, and background information as basic features, it is possible to use not only statistical features but also structural feature quantities that reflect the character structure, improving the efficiency of feature extraction processing. Is possible.

According to the nineteenth aspect of the present invention, the statistical characteristics of the block can be extracted from the basic characteristics, and the statistical characteristics extracted in block units are normalized by the size of the block. It is possible to extract features that are not easily affected by changes in the recognition target.

According to the twentieth aspect of the present invention, based on the basic features, the partial unevenness information of the background portion, which is effective for the identification between specific categories, and the bending direction of the stroke between similar categories And linearity difference, number of stroke loops, character background structure, character background size, positional relationship with circumscribed rectangle, end point / intersection information, and straight line extraction to extract structural features of characters It is possible to efficiently extract the structural features that express more explicitly, and by using these structural features, the recognition performance between similar categories that is difficult to identify with the identification method that uses only statistical features It is possible to improve.

According to the twenty-first aspect of the present invention, based on the basic feature, the partial unevenness information of the background portion, which is effective for the identification between specific categories, and the bending direction of the stroke between similar categories And linearity difference, number of stroke loops, character background structure, character background size, positional relationship with circumscribed rectangle, end point / intersection information, and straight line extraction to extract structural features of characters It is possible to efficiently extract the structural features that express more explicitly, and by using these structural features, the recognition performance between similar categories that is difficult to identify with the identification method that uses only statistical features It is possible to improve. Further, by extracting the number of loops based only on the label information, it is possible to improve the processing efficiency.

According to the invention of claim 22 of the present invention,
By using the background information extracted in advance as the basic feature, it is possible to extract the candidates for the end points and intersections by a simple process of comparing the background information of adjacent lines and checking the differences. By removing endpoints or intersections that are too close to other endpoints or intersection candidates based on the positional relationship of the candidates, erroneous extraction of endpoints or intersections due to noise is reduced, and handwritten categories 2 and 3 It becomes possible to extract the information of the end points and the intersections, which are effective in separating similar categories, with higher accuracy.

According to the invention of claim 23 of the present invention,
Using the background information extracted in advance as a basic feature, it is possible to extract straight lines by a simple process of checking the degree of continuity of black pixels and the degree of continuity to multiple lines, and the position information of end points and intersections. Using, to remove erroneous line candidates that occur due to line merging and branching, and classify the remaining line candidates according to the connection state with neighboring lines It becomes possible to extract a straight line effective for separating a category such as a handwritten category \ with higher accuracy.

According to the invention of claim 24 of the present invention,
Based on the size and the similarity of the circumscribed rectangle of the character pattern, if the circumscribed rectangle is significantly different from the circumscribed rectangle of other characters, or is out of the prescribed circumscribed rectangle, or the similarity is the similarity of other characters. When it is extremely low compared to, or when the difference between the similarity of the first rank and the similarity of the second rank or lower is small,
By detecting the cut-out failure character, analyzing the cause of the failure, and executing the cut-out again, it is possible to reduce the cut-out error and reduce the misrecognition due to the cut-out error.

According to the twenty-fifth aspect of the present invention,
By performing the rejection determination based on the ratio of the similarities, the magnitude of the maximum similarity, and the background complexity, it is possible to efficiently reject the preprocessing error character that is difficult to reject.

According to the twenty-sixth aspect of the present invention, the nonlinearity of the reject area and the non-reject area in the space around at least one of the similarity ratio, the maximum similarity and the background complexity is used. Since the boundary is preliminarily learned by the hierarchical neural network, more accurate reject determination processing can be performed.

[0179]

[0180]

[0181]

[0182]

[0183]

[Brief description of drawings]

FIG. 1 is a block diagram showing a configuration of an optical character reading device according to a first embodiment of the present invention.

FIG. 2 is a block diagram showing a configuration of a basic feature extraction unit according to the first embodiment.

FIG. 3 is a block diagram showing a configuration of a statistical feature extraction unit according to the first embodiment.

FIG. 4 is a block diagram showing the configuration of a first structural feature extraction unit in the first embodiment.

FIG. 5 is a block diagram showing a configuration of a first structure identification unit according to the first embodiment.

FIG. 6 is a block diagram showing the configuration of a first type processing unit according to the first embodiment.

FIG. 7 is a block diagram showing the configuration of a first handwriting processing unit according to the first embodiment.

FIG. 8 is a block diagram showing a configuration of an end point / intersection point extraction unit according to the first embodiment.

FIG. 9 is a block diagram showing a configuration of a straight line extraction unit according to the first embodiment.

FIG. 10 is a schematic diagram for explaining a character cutting method according to the first embodiment.

FIG. 11 is a schematic diagram for explaining a contour point extraction method according to the first embodiment.

FIG. 12 is a schematic diagram for explaining correspondence between a contour direction and encoding in the first embodiment.

FIG. 13 is a schematic diagram for explaining the contour direction encoding method according to the first embodiment.

FIG. 14 is a schematic diagram for explaining a background distance calculation method in the first embodiment.

FIG. 15 is a schematic diagram for explaining a background information calculation method according to the first embodiment.

FIG. 16 is a schematic diagram for explaining a block division method according to the first embodiment.

FIG. 17 is a schematic diagram for explaining the contour direction totaling method in the first embodiment.

FIG. 18 is a schematic diagram for explaining a method for finding outer points in the first embodiment.

FIG. 19 is a schematic diagram of a hierarchical neural network that is an example of the configuration of a recognition unit according to the first embodiment.

FIG. 20 is a schematic diagram for explaining a local background feature extraction method according to the first embodiment.

FIG. 21 is a schematic diagram for explaining a local stroke shape extraction method according to the first embodiment.

FIG. 22 is a schematic diagram for explaining a loop extraction method according to the first embodiment.

FIG. 23 is a schematic diagram for explaining a background block extraction method according to the first embodiment.

FIG. 24 is a schematic diagram for explaining an end point / intersection point extraction method by comparing the states of individual lines according to the first embodiment.

FIG. 25 is a schematic diagram for explaining conditions for removing inappropriate endpoints or intersections in the first embodiment.

FIG. 26 is a schematic diagram for explaining a horizontal straight line extraction criterion according to the first embodiment.

FIG. 27 is a schematic diagram for explaining a category separation method based on the size of a circumscribing rectangle in the first embodiment.

FIG. 28 is a schematic diagram showing an example of category 6 on the verge of closing the upper right portion in the first embodiment.

FIG. 29 is a schematic diagram for explaining a category separation method according to the difference in the number of end points viewed from the left side according to the first embodiment.

FIG. 30 is a schematic diagram for explaining a similar category separation method by straight line extraction in the first embodiment.

FIG. 31 is a schematic diagram for explaining a recognition result correction method based on background block information according to the first embodiment.

FIG. 32 is a schematic diagram for explaining the background label extraction method in the first embodiment.

FIG. 33 is a block diagram showing a configuration of an optical character reading device according to a second embodiment of the present invention.

FIG. 34 is a block diagram showing the configuration of a second structural feature extraction unit according to the second embodiment.

FIG. 35 is a schematic diagram for explaining a loop extraction method according to the second embodiment.

FIG. 36 is a block diagram showing a configuration of an optical character reading device according to a third embodiment of the present invention.

FIG. 37 is a block diagram showing the configuration of a second structure identification unit according to the third embodiment.

FIG. 38 is a block diagram showing a configuration of an optical character reading device according to a fourth embodiment of the present invention.

FIG. 39 is a schematic diagram showing a change in the size of a circumscribed rectangle due to a character cutout failure in the fourth embodiment.

FIG. 40 is a schematic diagram for explaining a character cutout correction method using a histogram according to the fourth embodiment.

FIG. 41 is a block diagram showing a configuration of an optical character reading device according to a fifth embodiment of the present invention.

FIG. 42 is a block diagram showing a configuration of an optical character reading device according to a sixth embodiment of the present invention.

FIG. 43 is a characteristic diagram showing an example of distribution of correct and incorrect characters in a space of similarity ratio and maximum similarity.

FIG. 44 is a schematic diagram showing a character image example of a character pattern having a small maximum similarity (<500).

FIG. 45 is a block diagram showing the configuration of a re-optical character reading device according to a seventh embodiment of the present invention.

FIG. 46 is a block diagram showing the configuration of a second recognition result calculation unit in the seventh embodiment.

FIG. 47 is a block diagram showing the configuration of the form document recognition device in the eighth embodiment.

FIG. 48 is a schematic diagram showing an example of an input form according to the eighth embodiment.

FIG. 49 is a schematic diagram showing an example of a frame structure reference table according to the eighth embodiment.

FIG. 50 is a block diagram showing the configuration of a form automatic processing system according to an eighth embodiment.

FIG. 51 is a block diagram showing the configuration of a conventional form document recognition device.

[Explanation of symbols]

1 character cutout part 2 Basic feature extraction unit 3 Statistical feature extraction unit 4 recognition section 5 First structural feature extraction unit 6 character type determination unit 7 First structure identification part 8 Second structural feature extraction unit 9 Second structure identification part 11 Path (which outputs the size of the circumscribed rectangle) 12 Path (output clipped image) 13 Path (which outputs the contour direction component) 14 (output background distance) route 15 (output background information) route 21 contour point extraction unit 22 Contour direction calculation unit 23 Background distance calculator 24 Background information calculator 31 Bi-directional block division 32 Horizontal block division 33 Vertical block division 34 Contour direction counting section 35 Contour direction normalization unit 36 Outward direction totaling section 37 Outer direction normalization unit 38 Background distance counting section 39 Background distance normalization unit 41 Local Background Feature Extraction Unit 42 Local stroke shape extraction unit 43 First Loop Extraction Unit 44 End point / intersection point extraction section 45 Straight line extraction unit 46 background block extractor 47 background label extraction unit 48 Second loop extractor 51 First type processing unit 52 First Handwriting Processing Unit 61 First circumscribing rectangle category determination unit 62 First recognition-dependent feature selection unit 63 First Selection Structural Feature Identification Unit 71 Second circumscribed rectangle category determination unit 72 Second recognition-dependent feature selection unit 73 Second selection structural feature identification unit 81 End point / intersection candidate extraction unit 82 End point / intersection point determination part 91 Black Pixel Continuous Area Extraction Unit 92 Straight line candidate extraction unit 93 Straight line determination unit 101 image input means 102 image memory 103 frame ruled line extraction means 104 frame structure extraction means 105 Frame structure reference table 106 frame structure matching means 107 character cutting means 108 character recognition means 109 post-processing means 371 Second type processing unit 372 Second handwriting processor 373 adder 374 Corrected category determination unit 381 Second character cutout part 382 First clipping error determination unit 411 Second cut-out error determination unit 421 First recognition result calculation unit 451 Second recognition result calculation unit 461 similarity ratio calculation unit 462 Rejection judgment unit 463 Correct answer category table 464 Output determination unit 465 Coupling load update unit 466 Background Complexity Calculation Unit

Front page continuation (31) Priority claim number Japanese Patent Application No. 10-62991 (32) Priority date March 13, 1998 (Mar. 13, 1998) (33) Priority claim country Japan (JP) (56) References JP-A-6-602222 (JP, A) JP-A-2-217977 (JP, A) JP-A-1-199285 (JP, A) JP-A-7-152856 (JP, A) JP-A-6- -114337 (JP, A) JP-A-7-306913 (JP, A) JP-A-7-129720 (JP, A) JP-A-62-257582 (JP, A) Multiple by structured neural network (NARA) Font Kanji recognition, IEICE Technical Report, Japan, September 20, 1991, Vol. 91 No. 231, pp. 9-16 (58) Fields investigated (Int.Cl. ⁷ , DB name) G06K 9/00-9/82

Claims (26)

(57) [Claims] 1. For character string data including a photoelectrically converted character frame line, the character frame line and the character string pattern are separated, and individual character patterns are cut out from the separated character string pattern to obtain individual character patterns. For the size of the circumscribed rectangle and the clipped character image surrounded by the circumscribed rectangle, and the contour information and the background structure information for each clipped individual clipped character image are extracted as basic features, and based on the extracted basic features. And extract at least one type of statistical feature, extract at least one type of structural feature based on the extracted individual cut-out character image and the extracted basic feature, and extract from the extracted statistical feature. The degree of similarity to each category of the cut-out character image is obtained, and the category with the highest degree of similarity is determined as the category to which each character pattern belongs. And a step of determining a category to which the final character pattern belongs based on the category to which the determined character pattern belongs, the degree of similarity to the category, and the extracted structural feature. Optical character reading method. 2. For character string data including photoelectrically converted character frame lines, a character frame line and a character string pattern are separated, and individual character patterns are cut out from the separated character string pattern, and individual character patterns are extracted. For the size of the circumscribed rectangle and the clipped character image surrounded by the circumscribed rectangle, and the contour information and the background structure information for each clipped individual clipped character image are extracted as basic features, and based on the extracted basic features. And extract at least one type of statistical feature, extract at least one type of structural feature based on the extracted individual cut-out character image and the extracted basic feature, and extract from the extracted statistical feature. The degree of similarity to each category of the cut-out character image is obtained, and the category with the highest degree of similarity is determined as the category to which each character pattern belongs. , The similarity is corrected based on at least one of the size of the circumscribed rectangle for each of the clipped individual character patterns, the category to which the determined character pattern belongs, and the extracted structural feature. An optical character reading method, comprising: each step of determining a category to which a final character pattern belongs based on the degree of similarity. 3. A character type determination of printed characters or handwriting is performed based on the size of a circumscribed rectangle with respect to the cut out character pattern, and the result of the character type determination and the size of the circumscribed rectangle with respect to the cut out character pattern and the And a step of determining a final category of the character pattern based on at least one of the category to which the determined character pattern belongs, the similarity to the category, and the extracted structural feature. The optical character reading method according to claim 1 or 2. 4. A clipping error is determined based on at least one of the size of a circumscribing rectangle for the clipped character pattern, the similarity to the determined category, and the extracted structural feature. The optical character reading method according to claim 1, further comprising steps for re-cutting out a character pattern when the judgment is made. 5. The determined character based on the maximum similarity of the cut-out character pattern, the value of the ratio between the second similarity and the maximum similarity, and the complexity of the background of the cut-out character pattern. Rejection processing is performed when the reliability of the category to which the pattern belongs is low, and in other cases, each step is provided that sets the category to which the determined character pattern belongs as a final recognition result. Item 5. The optical character reading method according to any one of Items 1 to 4. 6. For character string data including a photoelectrically converted character frame line, a character frame line and a character string pattern are separated, and individual character patterns are cut out from the separated character string pattern to obtain individual character patterns. A character slicing means for outputting a size of a circumscribing rectangle and a slicing character image surrounded by the circumscribing rectangle, and a basic feature for obtaining contour information and background structure information as a basic feature for each clipped character image clipped by the character slicing device. Extraction means, statistical feature extraction means for obtaining at least one type of statistical features based on the basic features output from the basic feature extraction means, individual cut-out character images cut out by the character cut-out means, and A structural feature extraction means for obtaining at least one type of structural feature based on the basic feature output from the basic feature extraction means; From the statistical features output from the statistical feature extraction means, the degree of similarity of the extracted character image to each category is obtained, and the degree of similarity of at least one category selected in descending order of similarity and at least one category selected in the same order And a category to which the final character pattern belongs based on the category output from the recognizer, the degree of similarity to the category, and the structural feature output from the structural feature extractor. An optical character reading device, comprising: 7. The first structural identification means selects an effective structural feature from at least one structural feature output from the structural feature extraction means based on the category output from the recognition means. 7. The optical character reader according to claim 6, further comprising recognition-dependent feature selection means for selecting and selected structural feature identifying means for determining a category based on the selected structural feature. 8. A character type determination means for determining a type or a handwritten character based on the size of a circumscribed rectangle output from the character cutout means, wherein the first structure identification means is a character type determined by the character type determination means. First character processing means for determining the category to which each character pattern belongs to the character determined to be, and category to which each character pattern belongs to the character determined to be handwritten by the character type determination means. The optical character reading device according to claim 6 or 7, further comprising: first handwriting processing means for determining. 9. The at least one structure output from the structural feature extraction means based on the size of the circumscribing rectangle output from the character clipping means and the category output from the recognition means, by the first type processing means. Dependent feature selecting means for selecting an effective structural feature from the physical features, and selected structural feature identifying means for performing category determination based on the structural feature selected by the recognition dependent feature selecting means. 9. The optical character reading device according to claim 8, wherein: 10. The circumscribing method in which the first type processing means makes a final category judgment for some categories based on the size of the circumscribing rectangle output from the character cutting means and the category output from the recognizing means. A structural category effective from at least one structural feature output from the structural feature extraction section based on the category output from the recognition section for the character that cannot be determined by the rectangular category determination section and the circumscribing rectangular category determination section. 9. The recognition-dependent feature selecting means for selecting a feature, and the selected structural feature identifying means for performing a category determination based on the structural feature selected by the recognition-dependent feature selecting means. Optical character reader. 11. The at least one structure output from the structural feature extraction means based on the size of the circumscribing rectangle output from the character clipping means and the category output from the recognition means, by the first handwriting processing means. Dependent feature selecting means for selecting an effective structural feature from the physical features, and selected structural feature identifying means for performing category determination based on the structural feature selected by the recognition dependent feature selecting means. 9. The optical character reading device according to claim 8, wherein 12. The circumscribing wherein the first handwriting processing means makes a final category judgment for some categories based on the size of the circumscribing rectangle output from the character clipping means and the category output from the recognizing means. A structural category effective from at least one structural feature output from the structural feature extraction section based on the category output from the recognition section for the character that cannot be determined by the rectangular category determination section and the circumscribing rectangular category determination section. 9. The recognition-dependent feature selecting means for selecting a feature, and the selected structural feature identifying means for performing a category determination based on the structural feature selected by the recognition-dependent feature selecting means. Optical character reader. 13. A character frame line and a character string pattern are separated from character string data including a photoelectrically converted character frame line,
Individual character patterns are cut out from the separated character string patterns, and the size of the circumscribing rectangle for each character pattern and the character cutting means that outputs the cut out character image surrounded by the circumscribing rectangle, and the individual character cutting means Basic feature extraction means for obtaining contour information and background structure information as basic features for the cut-out character image, and statistical feature extraction means for obtaining at least one type of statistical feature based on the basic features output from the basic feature extraction means. A structural feature extracting means for obtaining at least one type of structural feature based on each of the cut-out character images cut out by the character cutting means and the basic feature output from the basic feature extracting means; From the statistical features output from the feature extraction means, the similarity for each category of the extracted character image is obtained. , A recognition means for outputting the similarity of at least one category selected in descending order of similarity and at least one category selected in the same order, the size of the circumscribed rectangle of the character pattern output from the character cutting means, and The similarity output from the recognition unit is corrected based on at least one of the category output from the recognition unit and the structural feature output from the structural feature extraction unit, and based on the corrected similarity. An optical character reading device comprising: a second structure identifying means for finally determining a category to which a character pattern belongs. 14. The second structure identifying means outputs from the structural feature extracting means and at least one category output from the recognizing means, the size of the circumscribed rectangle of the character pattern output from the character cutting means, and the output from the structural feature extracting means. Based on the structural characteristics thus obtained, the similarity correction means for correcting the similarity output from the recognition means, the similarity of at least one category output from the recognition means, and the similarity correction means for correction. And a modified category determination means for determining a category to which each character pattern belongs from the modified similarity output from the adder. The optical character reading device according to claim 13. 15. A character type determination means for determining a print type or handwriting based on the size of a circumscribing rectangle output from the character cutout means, wherein the second structure identification means is a character type determined by the character type determination means. Second type processing means for determining a category to which each character pattern belongs to the character determined to be, and a category to which each character pattern belongs to the character determined to be handwritten by the character type determination means. 15. The optical character reading device according to claim 13 or 14, further comprising a second handwriting processing unit for determining. 16. The second type processing means outputs from the structural feature extraction means and at least one category output from the recognition means, the size of the circumscribed rectangle of the character pattern output from the character cutout means, and the output from the structural feature extraction means. Based on the structural characteristics thus obtained, the similarity correction means for correcting the similarity output from the recognition means, the similarity of at least one category output from the recognition means, and the similarity correction means for correction. And a modified category determination means for determining a category to which each character pattern belongs from the modified similarity output from the adder. 16. The optical character reading device according to claim 15. 17. The second handwriting processing means outputs the size of the circumscribed rectangle of the character pattern output from the character cutting means, at least one category output from the recognition means, and the structural feature extraction means. Based on the structural characteristics thus obtained, the similarity correction means for correcting the similarity output from the recognition means, the similarity of at least one category output from the recognition means, and the similarity correction means for correction. And a modified category determination means for determining a category to which each character pattern belongs from the modified similarity output from the adder. 16. The optical character reading device according to claim 15. 18. The contour point extracting means for extracting the contour points of the character pattern in the clipped image output from the character clipping means by the basic feature extracting means, and the individual contours output from the contour point extracting means. Contour direction calculation means for calculating a contour direction component in pixel units from the adjacency relationship of points;
Based on the position of the contour point output from the contour point extraction means, the background distance calculation means for obtaining the background distance from each side of the circumscribing rectangle of the character pattern to the character pattern, and the cutout image output from the character cutout means. Background information is calculated by scanning each line in the horizontal and vertical directions, and calculating the background length, which is the order in which the length and position of white pixels (background) that are consecutive over a certain number of pixels and the background in the scanning direction Means, and
16. The optical character reader according to any one of 13 and 15. 19. The bidirectional block dividing means for dividing the clipped image output from the character clipping means into a plurality of blocks in both the horizontal and vertical directions, and the block divided by the bidirectional block dividing means. For each unit, a contour direction totaling unit that totals the number of contour direction components output from the contour direction calculating unit of the basic feature extraction unit, and a block unit for each contour direction output from the contour direction totaling unit. Contour direction normalization means for obtaining the contour statistical feature value for each contour direction obtained by dividing the number of contour direction components by the size of the block, and the cutout image output from the character cutout means is divided into a plurality of blocks in the horizontal direction. Horizontal block dividing means for dividing into a plurality of pieces and a plurality of cut-out images output from the character cutting means in the vertical direction. Vertical block dividing means for dividing into horizontal blocks, and a block division boundary for each block divided by the horizontal block dividing portion and the vertical block dividing means based on the contour direction component output from the basic feature extracting means. The number of outline direction components, which is the number of outline direction components for each outline direction of the pixels that form the first character pattern, is scanned by scanning from the side of the circumscribed rectangle perpendicular to the line in the direction parallel to the block division boundary line. A contour direction normalization unit and a contour direction normal for obtaining a contour statistical feature amount for each contour direction obtained by dividing the number of contour direction components for each contour direction in block units output from the contour direction counting unit Based on the background distance output from the background distance calculation means of the basic feature extraction means,
The background distance summing means for summing the background distances in the block division direction in block units divided by the horizontal block dividing portion and the vertical block dividing portion, and the aggregate background distance in block units output from the background distance summing means 16. The optical character reading device according to claim 6, further comprising: a background distance normalizing unit that obtains a background distance feature amount obtained by dividing by a size. 20. A local background feature extracting means for extracting local background distance features based on the background distance output from the background distance calculating means of the basic feature extracting means, and the structural feature extracting means, Local stroke shape extraction means for extracting a local change in stroke direction based on the contour direction component output from the contour direction calculation means of the feature extraction means; a cutout character image output from the character cutout means and the basic feature Based on the first loop extracting means for extracting a loop of the character pattern based on the contour direction component output from the contour direction calculating means of the extracting means, and the background information calculated by the background information calculating means of the basic feature extracting means. A background block that calculates a block of adjacent background pixels for each background order as a background block and calculates its size information and position information Output means, and a label number is obtained as a background label a block of adjacent white pixels based on the background information calculated by the background information calculation means of the basic feature extraction means, the size, position, the background label is Background label extraction means for calculating label information indicating which side of the circumscribed rectangle is in contact, and end points and intersections of strokes forming a character pattern based on the background information calculated by the background information calculation means of the basic feature extraction means. End point to extract ...
An intersection point extracting means; and a straight line extracting means for extracting a straight line based on the background information calculated by the background information calculating means of the basic feature extracting means and the end points and the intersection points extracted by the end point / intersection point extracting means. Claims 6, 8, 1
16. The optical character reader according to any one of 3 and 15. 21. A local background feature extracting means for extracting a local background distance feature based on the background distance output from the background distance calculating means of the basic feature extracting means, and the basic feature extracting means, Local stroke shape extraction means for extracting a local stroke direction change based on the contour direction component output from the contour direction calculation means of the feature extraction means, and background information calculated by the background information calculation means of the basic feature extraction means. From the background block, a background block extracting means for calculating the size information and position information of the background block, which is a block of adjacent background pixels for each background degree, and background information calculated by the background information calculating means of the basic feature extracting means. From this, a block of adjacent white pixels is obtained as a background label, a label number is given, and its size, position, and background label are circumscribed rectangles. Background label extracting means for calculating label information indicating whether or not it is in contact with the edge of the stroke, and end points for extracting end points and intersection points of strokes forming a character pattern based on the background information calculated by the background information calculating means of the basic feature extracting means. An intersection point extracting means, a straight line extracting means for extracting a straight line based on the background information calculated by the background information calculating means of the basic feature extracting means, and the end point and the intersection point extracted by the end point / intersection point extracting means; and the background. The second loop extraction means for extracting a loop of a character pattern based on the label information calculated by the label extraction means, and the optical device according to any one of claims 6, 8, 13, and 15. Character reader. 22. The end point / intersection point candidate extraction means for extracting end points / intersection point candidates of a character pattern based on the background information of vertical and horizontal lines output from the background information calculation means. Means and end point / intersection point determination means for removing inappropriate end points and intersection point candidates by comparing the respective positional relationships of the extracted end points and intersection point candidates to extract appropriate end points and intersection points. 22. The optical character reading device according to claim 20, wherein: 23. The straight line extracting means extracts a region in which black pixels are continuous for longer than a predetermined threshold value based on the background information of each line in the vertical and horizontal directions output from the background information calculating means. Pixel continuous area extracting means, straight line candidate extracting means for extracting vertical and horizontal straight line candidates based on the black pixel continuous area extracting means output from the end point / intersection point extracting means Non-straight lines are removed from the straight line candidates output from the straight line candidate extraction means based on the position information of the end points and the intersections, and the type is determined from the remaining straight line candidates according to the connection state with a neighboring line. The optical character reading device according to claim 20 or 21, further comprising: a straight line determining means. 24. The size of a circumscribing rectangle of a character pattern output from the character cutting means, at least one degree of similarity output from the recognizing means, and background complexity based on background information calculated by the background information calculating means. claims to determine the error cut based on at least one of, characterized by comprising error determination means for giving an instruction of re-cut to the character segmentation means 6,8,13,1 5
An optical character reader according to any one of 1. 25. Two category similarities are selected from the larger category similarity calculated by the recognizing means, and 2 are selected.
The value of the ratio of the second similarity to the maximum similarity is larger than a preset threshold value a, or the maximum similarity is smaller than a preset threshold value b, or the background calculated by the background information calculating means. A first recognition result calculation means is provided which outputs a reject signal when the background complexity based on the information is larger than a preset threshold value c, and otherwise outputs a category of maximum similarity as a recognition result. The optical character reading device according to any one of claims 6, 8, 13, 15, and 24, wherein: 26. For a plurality of character patterns, the ratio of the second highest similarity to the highest similarity, the highest similarity, the background complexity based on the background information calculated by the background information calculating means, and A recognition result output by the recognition means is calculated, and at least one of the similarity ratio, the maximum similarity and the background complexity is input, and if the recognition result is incorrect, a reject signal is output and the correct answer is output. In the case of, the second recognition result calculation means constituted by a hierarchical neural network learned in advance so as not to output the reject signal is provided.
4. The optical character reading device according to any one of 4 above.