JP2009223612A - Image recognition device and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To prevent the degradation of recognition rate in recognition of a handwritten character entered to a fixed form sheet, compared with the case of considering all possibilities for the presence of an image around an entry column. <P>SOLUTION: A device includes: a recognition dictionary 120 which stores one or more class candidates corresponding to each category; an application class determination part 112 which selects, from the one or more class candidates corresponding to the category stored in the recognition dictionary 120, one or more classes specified based on a positional relationship between an entry column and an image around the entry column as application classes; a pattern recognition part 116 which recognizes a class to which an entered image belongs from the application classes by executing pattern recognition processing to the entered image to the entry column; and a category determination part 118 which specifies a category corresponding to the class recognized by the pattern recognition part 116 based on the candidate storage means. <P>COPYRIGHT: (C)2010,JPO&INPIT

Description

  The present invention relates to an image recognition apparatus and a program.

  One application field of handwritten character recognition is the recognition of handwritten entries on standard forms (also called forms) such as forms, answer sheets, and questionnaire forms. In recognition of handwritten entry on a standard form, for example, a difference between an image of a standard form after handwriting and an image of a non-filled form is obtained, and pattern recognition is performed on the obtained difference image. In such recognition, if the handwritten entry overlaps the image that originally exists on the standard paper, such as an entry frame, a photograph, or a graphics figure, a complete image of the handwritten entry cannot be obtained. There was a possibility that the accuracy of would deteriorate.

  On the other hand, in the apparatus disclosed in Patent Document 1, a large number of frame contact characters in which an entry frame and characters are superimposed are created while changing the position, size, or inclination of the character with respect to the entry frame. Contact characters are registered in the knowledge table and used for pattern recognition.

Japanese Patent Laid-Open No. 10-154204 (in particular, paragraphs 211 to 228, see FIGS. 41 to 44)

  If there is a possibility that various images such as photographs and figures exist at various positions around the entry field, it is conceivable to perform pattern recognition considering all these possibilities. However, there is a concern that as the number of patterns to be considered in pattern recognition increases, misrecognition increases and the recognition rate decreases.

  An object of the present invention is to suppress a decrease in the recognition rate as compared with a case where all the possibilities regarding the presence of an image around the entry field are considered.

  The present invention corresponds to candidate storage means for storing one or more class candidates corresponding to the category for each of the plurality of categories, and corresponding to the category stored in the candidate storage means for each of the plurality of categories. Selecting means for selecting one or more classes specified based on the positional relationship between the entry field and the image elements around the entry field from among one or more class candidates as a class to be used for determination of the category; The pattern recognition process is performed on the image entered in the entry field to recognize the class to which the entry image in the entry field belongs from the classes selected by the selection unit for each of the plurality of categories. And a category corresponding to the class recognized by the pattern recognition means based on the candidate storage means. Identified, an image recognition device and an output means for outputting the identified category as image belongs categories of entries for the answer column.

  In one aspect, the image recognition apparatus includes: an extraction unit that extracts an image of an entry for the entry field from an input image; and an end point detection unit that detects an end point of the entry image extracted by the extraction unit. Further, the selection means specifies one or more image elements that are in contact with any of the end points detected by the end point detection means, and performs the determination based on a positional relationship between the specified one or more image elements and the entry field. Select the class to use.

  In another aspect, the image recognition apparatus includes class candidates stored in the candidate storage unit for each entry field of a standard document including one or more image elements and one or more entry fields, and for each category. Entry field information storage means for storing one or more classes specified based on the positional relationship between the image elements around the entry field and the entry field as a class used for determination of the entry field, and an input target Extraction means for extracting an image of entry for each entry field from an image, and the selection means is configured to determine the entry field for each entry image for each entry field extracted by the extraction means. The class to be used is obtained from the entry field information storage means.

  Hereinafter, an embodiment of the present invention (hereinafter referred to as an embodiment) will be described with reference to the drawings.

  The image recognition apparatus according to the present embodiment is an apparatus for recognizing the entry contents of characters and symbols (hereinafter collectively referred to as “characters”) entered by a user on a fixed form (form) such as a form or answer sheet. is there.

  In this embodiment, the positional relationship between the entry field and the image objects around the entry field (for example, a print image that exists before entry, such as a photograph, a figure, an entry frame, and a character string). Based on this, the patterns to be taken into account when recognizing the contents entered in the entry field are limited. The number of patterns that should be considered that there is a possibility that an image object exists at any position with respect to the entry field is enormous, but if the position of the image object is limited, the number of patterns to be considered can be reduced. it can.

  In general, in pattern recognition, for example, a recognition dictionary that holds feature information of each character to be identified is prepared, and by referring to the recognition dictionary, it is determined which character the character image to be identified corresponds to. Here, in this embodiment, not only feature information corresponding to characters in a complete state but also feature information of characters partially hidden by surrounding image objects is registered in the recognition dictionary. Here, even if the character is the same, if the relative positional relationship with the surrounding image object is different, the portion hidden by the image object is different, so that the character is relative to the surrounding image object for each character. Feature information when hidden in various positional relationships is registered in the recognition dictionary. The positional relationship here includes, for example, a relative direction and / or distance. At the time of pattern recognition, among the various feature information registered in the recognition dictionary, information corresponding to the positional relationship between the character to be recognized and the surrounding image objects is selected and used.

  In the following, the type of character is referred to as “category”, and various variations of the same type of character but differently hidden by surrounding image objects are referred to as “class”. For example, if the circle is one “category”, the shape of the circle that is partially hidden by surrounding image objects is one “class”. In this case, there are various “categories” in the same circle depending on the shape and size of the surrounding image object, or the positional relationship of the image object with respect to the circle, for example, where the image object is located up, down, left and right "Class" is possible.

  Considering all such various classes for a category, the calculation of pattern recognition is costly. However, if there is an image object that partially hides the character around the character to be recognized, and what the positional relationship of the object with respect to the character can be determined by some means, You can narrow down the classes to be considered in the category. The present embodiment is based on such a concept.

  An example of the configuration of the image recognition apparatus of this embodiment is shown in FIG. In this example, an already-filled document obtained as a result of a user filling in a fixed form is read by an image reading device such as a scanner, thereby generating an electronically filled image 10. The completed image 10 is input to the entry extraction unit 102 included in the image recognition apparatus.

  The entry extraction unit 102 extracts an image of the entry contents of the user from the entered entry image 10. Extraction of the contents of entry can be performed by subtracting the original image from the completed image 10, for example. As another example, an image of entry content (hereinafter referred to as “entry content image”) can be extracted by extracting a pixel group of the color of the writing instrument used for entry from the completed image 10. This method can be applied, for example, when writing is performed using a writing instrument of a color that is not included in the image on the standard paper. There is also a method of subtracting the original image from the completed image 10 and extracting the pixel group of the writing instrument color of interest from the subtraction result image. This method can be used when, for example, an answer written by a subject with a pencil and a scoring result such as a round mark or cross mark written by a grader with a red pen are separated and extracted. When a plurality of entry content images are included in the completed image 10, the entry extraction unit 102 extracts information on position coordinates of the image in the completed image 10 in association with each entry content image. May be. The position coordinates of the entry content image are, for example, predetermined reference points (for example, diagonal lines of the rectangle) in the circumscribed rectangle (eg, the vertical and horizontal sides are parallel to the x and y axis directions, respectively) of the entry content image. It may be represented by a set of coordinates of the intersections of each other or the point in the upper right corner of the rectangle).

  The recognition processing unit 110 performs pattern recognition processing on the entry content image extracted by the entry extraction unit 102 to determine a character indicated by the entry content image. The recognition dictionary 120 is used for this determination.

  The recognition processing unit 110 includes an application class determination unit 112, a determination information storage unit 114, a pattern recognition unit 116, and a category determination unit 118.

  The application class determination unit 112 determines a class to be applied in pattern recognition of the extracted entry content image. In this embodiment, with reference to the determination information storage unit 114, a class to be applied to pattern recognition of the entry content image (hereinafter referred to as “application class”) is specified.

  For example, in the case of a standard document such as a questionnaire, the position of each entry column in the document is fixed, and what image object is present at which position around each entry column is known. The entry content image entered in the entry field is not hidden unless there is an image object in the vicinity of the entry field, for example. Therefore, in this case, there is no need to consider the possibility that the entry content image is a partially hidden character. Further, if there is a large, high-density image object near the lower part of the entry field, it is necessary to consider that the lower part of the entry content image for that field is partially hidden.

  For example, the standard document illustrated in FIG. 2 includes a group of image objects such as text articles 11 and photographic images 12-1 and 12-2, graphics figures 13, and entry fields 14-1 and 14-2. Are arranged at predetermined positions. For example, since the photographic image 12-1 exists near the lower part of the entry column 14-1, the possibility that the lower part of the entry content image in the entry column 14-1 is hidden by the photographic image 12-1 over a wide range is considered. There is a need to. However, since the left and right and above the entry field 14-1 are parenthesis line drawings and text articles 11 indicating the entry field, even if the entry content image is hidden by them, the entry content image is only hidden by the thin line. As the shape feature changes, the possibility of being hidden over a wide range is low. If it is only partially hidden by the thin line width, the portion hidden by known image processing such as missing line segment interpolation can be almost restored, so that it can be handled in the same way as when it is not hidden. Accordingly, in the recognition of the contents of entry in the entry field 14-1, for each category such as “circle” and “cross”, the characters of the category are not hidden at all (that is, the user correctly fits in the entry field). Each class corresponding to the case where the character is entered) and the lower part is hidden may be considered. For the entry field 14-2, there is only a parenthesis line representing the entry field in the vicinity (assuming that the upper photographic image 12-1 is too far from the field), so that characters are not hidden at all. Only relevant classes need to be considered.

  As described above, from the layout of the standard document or the information on the image, it can be known in advance which part should be considered for the recognition of the entry in each entry field. Therefore, in this case, the determination information storage unit 114 may store the determination information indicating for each entry column which class to apply from among the many classes in each category.

  As another example, for each combination of positions and types of peripheral image objects with respect to the entry content image, information indicating an application class corresponding to the combination may be stored in the determination information storage unit 114. In this example, the application class determination unit 112 checks whether there is an image object near the periphery of the entry content image extracted from the completed image 10, for example, determines the position and type of the image object, and determines the determination. An application class corresponding to the result is obtained from information in the determination information storage unit 114. This example can also be applied when there is no prior knowledge about the image objects near the periphery of the entry content image.

  The pattern recognition unit 116 determines a class corresponding to the entry content image from the application classes determined by the application class determination unit 112 by performing a known pattern recognition process. In this determination, the recognition dictionary 120 is referred to.

  FIG. 3 shows an example of data contents of the recognition dictionary 120. This example is an example of a recognition dictionary for discriminating round marks, triangle marks, and cross marks for scoring a test. In this example, for each of the categories of circle, triangle, and cross, the class ID (identification information) and class determination information of each class included in the category are included. The class determination information is information that serves as a determination criterion when determining whether or not an entry content image belongs to the class in pattern recognition. For example, in the case of a circle, a feature amount representing the feature of a circle obtained by learning a sample of a circle written by various people is an example. As for the class in the case where the lower part of the circle is hidden, a feature amount learned from a large number of samples of the circles whose lower part is hidden may be used as the class determination information. Such a feature is called a “prototype”. In addition to the prototype, there is also known a method that considers the variance of the learning sample group. In this case, information on the variance is also included in the class determination information. The use of a prototype as class determination information is merely an example, and other determination information used in a known pattern recognition method may be used. For example, a set of sample images corresponding to a class may be used as class determination information for the class.

  In order to create class determination information for a class corresponding to a partially hidden state, each sample image of a character that is not hidden is partially hidden in a state corresponding to that class, Just learn. In addition to the sample image class that is only partially hidden, an image in which the end points of the sample image resulting from the partial occlusion are connected is created, and the end points are connected by learning them. Another class of determination information corresponding to the state may be generated.

  Using such a recognition dictionary 120, the pattern recognition unit 116 identifies one or more classes closest to the target entry content image from one or more application classes.

  The category determination unit 118 determines which category the class specified by the pattern recognition unit 116 belongs to by referring to the recognition dictionary 120. For example, when the pattern recognition unit 116 determines that the entry content image is closest to the class C03, the category determination unit 118 obtains the category C (circle) to which the class C03 belongs as a recognition result.

  The outline of the present embodiment has been described above. Next, a more specific system example will be described.

  First, an example of a preliminary preparation type will be described as a first example. This example is used for recognizing entry for a document that knows in advance how the entry may be hidden, such as a standard document. In this example, an application class for each entry field is obtained in advance using information known in advance, and information on the application class is used at the time of actual pattern recognition. This example will be described with reference to FIGS.

  FIG. 4 is a diagram showing an overall view of the preparatory system. As shown, the system includes a preparation processing system 130 and a recognition system 150. The preparation processing system 130 is a system that executes preparation processing for constructing the above-described determination information storage unit 114. In this example, the determination information storage unit 114 includes an entry column position storage unit 142 and an entry column / class correspondence storage unit 144 (both will be described in detail later). The preparation processing system 130 analyzes the image of the standard document before entry, identifies the application class for each entry field in the fixed form document, and registers it in the entry field / class correspondence storage unit 144. The entry column position storage unit 142 stores the position of each entry column on the standard document. Information on the position of each entry field may be obtained in advance by the user and registered in the entry field position storage unit 142.

  The recognition system 150 refers to the determination information storage unit 114 constructed by the preparation processing system 130 and performs pattern recognition of the entered content in the completed image.

  Hereinafter, detailed examples of the preparation processing system 130 and the recognition system 150 will be described in order.

  FIG. 5 shows an example of a detailed configuration of the preparation processing system 130. The preparation processing system 130 reads the original image 30 of the standard document and examines image objects existing in the vicinity of each entry field in the original image 30 to determine a class to be applied to recognition of the entry field. judge. The original image 30 is an image of a manuscript of a standard document that is not filled in at all.

  The original image 30 is input to the image input unit 132 in this mechanism. The image input unit 132 may be a scanner that reads the original image 30 printed on paper, or may be an interface that inputs electronic data of the original image 30. The object extraction unit 134 extracts image objects such as character strings, photographs, and line drawings existing on the input original image 30 by a known technique. At this time, the object extraction unit 134 also obtains the position of the image object on the original image 30. The position of the image object is, for example, an intersection of diagonal lines of a rectangle circumscribing the image object (vertical and horizontal sides parallel to the vertical and horizontal sides of the original image 30), or a set of coordinates of two points on the diagonal of the rectangle. It can be expressed as

  In the entry column position storage unit 142, as illustrated in FIG. 6, for each entry column on the original image 30, identification information (entry column ID) of the entry column and a position on the original image 30 are registered. Yes. In this example, the position of the entry column is represented by the coordinates of two points, the upper right corner and the lower left corner of the rectangular range of the entry column.

  The duplication determination unit 136 determines, for each entry field registered in the entry field position storage unit 142, image objects that may overlap with characters in the entry field from among the image objects extracted by the object extraction unit 134. To do. In this process, for example, the area of the entry field indicated by the position information of the entry field is enlarged at a predetermined magnification, and it is determined whether or not the enlarged entry field area and each image object overlap each other. In this example, by expanding the area of the entry field, it corresponds to the case where the user has entered so as to slightly protrude from the entry field. In this determination, for each of the entry fields, the type of image object (text, line drawing, dark photographic image, thin photographic image, etc.) that overlaps the entry field and the entry field of the image object. The positional relationship can be specified.

  In the selection rule storage unit 139, for each combination of the type of the image object and the positional relationship between the entry field and the image object, an application class when the combination is applicable is stored. FIG. 7 schematically shows an example of information stored in the selection rule storage unit 139. Strictly speaking, the leftmost column in the table shown in FIG. 7 shows a specific example of overlap, and is not stored in the selection rule storage unit 139. In this column, the range of the entry column is indicated by a broken-line rectangle for convenience. What is stored in the selection rule storage unit 139 is information on columns of “duplicate object” and “class applied to recognition” (applied class).

  In the “duplicate object” column, a combination of the type and position (positional relationship) of the duplicate object is registered. For example, the example of the first line from the top is an example in the case where, for example, the overlapping image object is a parenthesis indicating an entry field. In this example, since the type of the overlapping object is a parenthesis, it is “thin line segment”, and the position is on the left and right of the entry column. The second line corresponds to the case where a dark photographic image object exists in the upper right of the entry field. The class ID of the applicable class is registered in the “class provided for recognition” column. In the figure, for convenience of explanation, character shapes corresponding to classes are shown instead of class IDs.

  For example, in the example of the first line from the top, since the overlapping object is only a thin line segment, for each category such as a circle mark and a cross mark, class determination information of a basic class that is not hidden by the image object is registered. Yes. On the other hand, in the example of the second row, since there is a dark photographic image area at the upper left of the entry column, it is necessary to consider the possibility of hiding the upper left of the characters entered in the entry column. Therefore, in the application class in the case of the second line, for each category, in addition to the basic class, a class corresponding to the case where the upper left of the characters in the category is hidden and missing is registered. The example on the fourth line is an example in which a thick rectangular frame surrounding the entry field exists. The rectangular frame is regarded as a thick line segment on the top, bottom, left and right of the entry field. In this case, it may be recognized as a series of character images in which the entry content image and a thick line segment as an image object are connected. For example, when a round mark crosses a part of a side of a rectangular frame, for example, a half-moon-shaped image in which a line constituting the round mark and a side of the rectangular frame are integrated is recognized as an entry content image. There is a case. Therefore, a class corresponding to such a half-moon image is added to the application class.

  If there is a photo image object that is thin enough to extract the entry content image even if the entry content image overlaps the object, the entry content image is partially missing due to duplication. Need not be considered.

  Whether the object is a dark photographic image or a thin photographic image can be determined by a known technique (such as calculating an average density).

  The class correspondence information registration unit 138 determines the application class of each entry field based on the determination result of the duplication determination unit 136 and the information stored in the selection rule storage unit 139, and the correspondence between the determined entry field and the application class Is registered in the entry field / class correspondence storage unit 144.

  For example, when the duplication determination unit 136 determines that a dark photographic image object overlaps in the upper right of a certain entry field, the class correspondence information registration unit 138 uses the entry field / class correspondence storage unit 144 as the ID of the entry field. Correspondingly, the ID of each class registered in the column of “applicable class” on the second line in the example of FIG. 7 is registered. FIG. 8 shows an example of data stored in the entry field / class correspondence storage unit 144. “C01”, “T02”, etc. are class IDs.

  By performing the processing as described above for all entry fields on the original image 30, the information on the applicable classes corresponding to each entry field is accumulated in the entry field / class correspondence storage unit 144.

  The preparation processing system 130 has been described above. Next, a detailed example of the recognition system 150 will be described with reference to FIGS. 9 and 10.

  The recognition system 150 illustrated in FIG. 9 includes an image input unit 152, a positional distortion correction unit 154, a difference extraction unit 156, a preprocessing unit 160, a recognition processing unit 110, and a recognition dictionary 120.

  The completed image 10 to be recognized is input to the image input unit 152. The completed image is an image after the user has entered the original image 30. The image input unit 152 may be a scanner that reads a paper document, may be an interface for inputting image data, and may also be used as the image input unit 132 of the preparation processing system 130. The positional distortion correction unit 154 performs well-known positional deviation correction and distortion correction (for alignment with the original image 30) on the input completed image 10. The difference extraction unit 156 subtracts the original image from the completed image that has been subjected to position correction and distortion correction, thereby obtaining a differential image that shows only the user's written content. The preprocessing unit 160 performs preprocessing on the difference image obtained by the difference extraction unit 156 to obtain an image suitable for subsequent pattern recognition. In the illustrated example, the preprocessing unit 160 includes a predetermined color pixel extraction unit 162, a missing line segment interpolation unit 164, and a contact separation unit 166.

  The predetermined color pixel extraction unit 162 extracts only pixels of a predetermined color from the difference image. This assumes, for example, recognition of a scoring mark on a test answer. For example, in a typical case where a test taker writes an answer with a black pen on an answer sheet and a grader writes a scoring mark such as a circle or cross with a red pen on the answer sheet, a predetermined color pixel extraction is performed. The unit 162 extracts a pixel having a color belonging to the red range from the difference image.

  The missing line segment interpolation unit 164 performs missing line segment interpolation on the image composed of pixels of the predetermined color extracted by the predetermined color pixel extraction unit 162. The missing line segment interpolation is a well-known process for storing a missing line segment. As a result, even if a part of the line segment is lost due to reading noise or crossing with a thin line, the part is interpolated.

  The contact separation unit 166 performs processing for separating the images that are to be separated into the plurality of objects in the image after the missing line segment interpolation. Since a known method (for example, see Japanese Patent Application Laid-Open No. 2007-241357) can be used for the contact separation process, description thereof is omitted.

  With the functional modules as described above, the preprocessing unit 160 obtains images of individual scoring marks in the filled-in image 10 (the above-mentioned filled content images), and each of these images is filled in the filled-in image 10 of the image. Output in association with the information indicating the position above. The recognition processing unit 110 determines which category each entry content image belongs to by referring to the recognition dictionary 120.

  An example of a detailed configuration of the recognition processing unit 110 is shown in FIG. In the example of FIG. 10, the normalization unit 172 normalizes each entry content image input from the preprocessing unit 160. Normalization is a well-known process for adjusting the size, position, inclination, line width, etc. of an image to a predetermined standard. Further, the feature quantity extraction unit 174 obtains the feature quantity of each of the entry content images after normalization. This is a process of compressing information by converting an entry content image into a set of parameters (features) representing the features of the image, not the image itself. A conventionally well-known one may be used as an image feature amount or a method for obtaining the feature amount.

  In parallel with the processing for the entry content image, the entry field specifying unit 176 inputs the entry content image into which entry field based on the position information of the entry content image input from the preprocessing unit 160. To identify For this specification, information in the entry column position storage unit 142 is referred to. For example, the entry field closest to the position of the entry content image may be determined as the entry field corresponding to the image, and the ID of the entry field may be obtained. The application class determination unit 112 searches the entry column / class correspondence storage unit 144 for the ID of each application class corresponding to the entry column ID specified by the entry column specifying unit 176 for each entry content image. The ID of each application class obtained for each entry content image is input to the pattern recognition unit 116.

  The pattern recognition unit 116 obtains which class the image corresponds to for each entry content image by pattern recognition processing. More specifically, for each entry content image, class determination information corresponding to the ID of each application class corresponding to the entry content image obtained by the application class determination unit 112 is obtained from the recognition dictionary 120, and each of the application classes is determined. Based on the class determination information and the feature amount of the entry content image obtained by the feature amount extraction unit 174, known pattern recognition is performed. As a result, one or more classes closest to the entry content image are specified from the application classes.

  The category determination unit 118 obtains to which category the identified class belongs from the recognition dictionary 120 (see FIG. 3). The obtained category (in the example of FIG. 3, any one of circle, triangle, and cross) becomes the recognition result of the entry content image by the recognition system 150.

  In addition, although the description up to the recognition of each entry content image has been described above, post-processing such as counting the recognition results (for example, counting the number of circles and adding points) may be continued. .

  The advance preparation type system has been described above. This system is used when it is known in advance what kind of image object is in the vicinity of the entry content image. On the other hand, an example of a system without prior preparation that can be applied even in the absence of such prior knowledge will be described with reference to FIGS. 11 and 12.

  In this system, a mechanism such as the preparation processing system 130 illustrated in FIG. 4 is not necessary, and only a recognition system is required. The overall configuration of the recognition system in this case may be the same as that shown in FIG. 9, for example, but the configuration of the recognition processing unit 110 therein is different from that of the preparatory system. An example of the configuration of the recognition processing unit 110 in the system without prior preparation is shown in FIG.

  Of the internal configuration of the recognition processing unit 110 illustrated in FIG. 11, the remaining functional modules excluding the end point extraction unit 180, the contact determination unit 182, the applied class determination unit 186, and the contact end point / class correspondence storage unit 146 are as shown in FIG. These are the same as the modules with the same reference signs in the recognition processing unit 110 of the preparatory system exemplified in FIG.

  Generally speaking, in a system without advance preparation, processing similar to that performed by the preparation processing system 130 in the advance preparation type system is executed when recognizing a written content image (not in advance). However, in the case of pre-processing, since there is no actual entry content image, the duplication between the entry field and the image object was examined, whereas in this example, the duplication between the actual entry content image and the image object was examined. it can. Checking the entry field requires that all possible hidden states due to possible image objects be inspected as applicable classes so that there is no recognition omission even if the actual entry image is slightly shifted from that field. It was. On the other hand, in this example, only the class corresponding to the actual overlapping state between the entry content image and the surrounding image object has to be inspected as the application class, so that a further narrowing down can be performed.

  In this example, the end point extraction unit 180 extracts end points from individual entry content images given from the preprocessing unit 160. Since the method of extracting the end points from the image is well known, the description is omitted.

  The contact determination unit 182 identifies an end point in contact with another image object among the end points of the entry content image extracted by the end point extraction unit 180. Whether or not the detected end point is in contact with another image object can be determined by comparing the position of the end point with the position of each image object obtained from the original image 30.

  The application class determination unit 186 determines a class to be considered when recognizing the entry content image, that is, an application class, based on the determination result of the contact determination unit 182. In this determination, information stored in the contact end point / class correspondence storage unit 146 is referred to.

  The contact end point / class correspondence storage unit 146 stores application classes in association with the positions of the end points in contact with other image objects among the end points of the entry content image. FIG. 12 schematically shows an example of data stored in the storage unit 146. Strictly speaking, in the table shown in FIG. 12, the leftmost column shows an example of a specific entry content image corresponding to the case indicated by each row, and is stored in the contact end point / class correspondence storage unit 146. Not. What is stored in the contact end point / class correspondence storage unit 146 is information in the columns of “contact between end point and other figure” and “class applied to recognition” (applied class). In the “contact between end point and other figure” column, the position of the end point in contact with another image object among the end points in the entry content image is registered. The position of the end point in contact is the position of the end point in the entry content image. For example, the end point is on the upper, lower, left, or right side of the entry content image.

  For example, the leftmost column of the table shown in FIG. 12 shows an example of the entry content image given from the preprocessing unit 160. The black line image inside the broken rectangle is the entry content image. In the example shown in the top line, the entire handwritten circle is cut out as an entry content image. In this image, there are two end points below, but neither of them is in contact with another image object (which may be obtained from the original image 30). Here, if the entry content image is partially hidden by another image object, the entry content image should have an end point in contact with the object. Therefore, in other words, the entry content image in the top row in FIG. 12 is not hidden by other image objects. In such a case, as shown in the rightmost column of the table of FIG. 12, only the class of the basic form of characters of each category (a state in which there is no omission) may be set as the application class.

  On the other hand, in the example of the second row in FIG. 12, the entry content image given from the preprocessing unit 160 has four end points. Here, there is an image object above the circumscribed rectangle (broken line rectangle in the figure) of the entry content image (see the second line in the center column of the table of FIG. 12), and the positions of the two upper end points are the object. If it is found that the area is in contact with the area, the two end points are extracted by the end point extraction unit 180. In this case, it can be seen that these two end points are the upper end points in the entry content image, and that, from another viewpoint, the image object overlapping the entry content image is above the circumscribed rectangle of the image. Thus, it can be seen that the upper part of the entry content image is hidden. In this case, for each category, in addition to the basic class, a class corresponding to the state where the upper part is hidden is added to the application class. Will be. For example, for the category of circles, a class such as a circular arc with a missing upper part or a semicircular shape with a missing upper part (considering the case where an end point connection is made) is added.

  The application class determined by the application class determination unit 186 is supplied to the pattern recognition unit 116.

  The pattern recognition unit 116 and the category determination unit 118 may perform the same processing as the corresponding element in the example of FIG.

  The embodiments of the present invention have been described above.

  In the case of pattern recognition of an entry content image, the possibility of misrecognition increases if all the cases of partial image loss that may occur due to other image objects are considered. On the other hand, in the present embodiment, the cases to be considered are narrowed down based on the positional relationship between the entry content image (or the entry field in which it is entered) and the surrounding image object, so that there is a possibility of erroneous recognition. Can be reduced.

  Each system described above is realized, for example, by causing a general-purpose computer to execute a program representing the processing of each functional module described above. Here, for example, as shown in FIG. 13, the computer includes, as hardware, a microprocessor such as a CPU 1000, a memory (primary storage) such as a random access memory (RAM) 1002 and a read only memory (ROM) 1004, an HDD ( HDD controller 1008 that controls (hard disk drive) 1006, various I / O (input / output) interfaces 1010, network interface 1012 that performs control for connection to a network such as a local area network, etc. are connected via bus 1014, for example. Circuit configuration. In addition, for example, a disk drive 1016 for reading from and / or writing to a portable disk recording medium such as a CD or a DVD, a portable memory of various standards, such as a flash memory via the I / O interface 1010 to the bus 1014. A memory reader / writer 1018 for reading from and / or writing to the nonvolatile recording medium may be connected. A program in which the processing contents of each functional module exemplified above are described is stored in a fixed storage device such as a hard disk drive via a recording medium such as a CD or DVD, or via a communication means such as a network, and stored in a computer. Installed. The program stored in the fixed storage device is read into the RAM 1002 and executed by a microprocessor such as the CPU 1000, whereby the functional module group exemplified above is realized. Some or all of these functional module groups are used as hardware circuits such as dedicated LSI (Large Scale Integration), ASIC (Application Specific Integrated Circuit), or FPGA (Field Programmable Gate Array). It may be configured.

It is a figure which shows an example of the image recognition apparatus of embodiment. It is a figure which shows the example of the completed image of a fixed form document, and the example of the image of the result of extracting the scoring symbol from there. It is a figure which shows an example of the data content of a recognition dictionary typically. It is a figure which shows the whole image of a preparation type system. It is a figure which shows the example of a detailed structure of the preparation processing system among the systems of a prior preparation type. It is a figure which shows the example of the data which an entry column position memory | storage part memorize | stores. It is a figure which shows the example of the selection rule which a selection rule memory | storage part memorize | stores. It is a figure which shows the example of the data which an entry column and a class corresponding | compatible memory | storage part memorize | store. It is a figure which shows an example of the whole structure of an example of a recognition system. It is a figure which shows the example of an internal structure of the recognition process part of a recognition system. It is a figure which shows the example of a structure of the recognition process part in the system without prior preparation. It is a figure which shows an example of the data which a contact endpoint and class corresponding | compatible memory | storage part memorize | stores. It is a figure which shows an example of the hardware constitutions of a computer.

Explanation of symbols

  DESCRIPTION OF SYMBOLS 10 completed image, 102 entry extraction part, 110 recognition process part, 112 application class determination part, 114 determination information storage part, 116 pattern recognition part, 118 category determination part, 120 recognition dictionary

Claims (4)

Candidate storage means for storing one or more class candidates corresponding to the category for each of the plurality of categories;
Each of the plurality of categories is specified based on the positional relationship between the entry field and the image elements around the entry field from among one or more class candidates corresponding to the category stored in the candidate storage unit. A selection means for selecting one or more classes as classes to be used in the determination for the category;
By executing pattern recognition processing on the image entered in the entry field, the class to which the entry image in the entry field belongs is recognized from the classes selected by the selection means for each of the plurality of categories. Pattern recognition means;
An output means for identifying a category corresponding to the class recognized by the pattern recognition means based on the candidate storage means, and outputting the identified category as a category to which an image of entry for the entry field belongs;
An image recognition apparatus comprising: Extraction means for extracting an image of entry for the entry field from the input image;
Endpoint detection means for detecting the endpoint of the image of the entry extracted by the extraction means;
Further comprising
The selection means identifies one or more image elements that are in contact with any of the end points detected by the end point detection means, and a class to be used for the determination based on the positional relationship between the specified one or more image elements and the entry field Select
The image recognition apparatus according to claim 1. Of the class candidates stored in the candidate storage unit for each entry field of the standard document including one or more image elements and one or more entry fields and for each category, the image elements around the entry field and the Entry field information storage means for storing one or more classes identified based on the positional relationship with the entry field as a class used for determination of the entry field;
Extraction means for extracting an image of entry for each entry field from the input target image;
Further comprising
The selection means obtains, from the entry field information storage means, a class to be used for determination of the entry field for each entry image for each entry field extracted by the extraction unit.
The image recognition apparatus according to claim 1. Computer
Candidate storage means for storing one or more class candidates corresponding to the category for each of the plurality of categories;
Each of the plurality of categories is specified based on the positional relationship between the entry field and the image elements around the entry field from among one or more class candidates corresponding to the category stored in the candidate storage unit. A selection means for selecting one or more classes as classes to be used for determination of the category;
By executing pattern recognition processing on the image entered in the entry field, the class to which the entry image in the entry field belongs is recognized from the classes selected by the selection means for each of the plurality of categories. Pattern recognition means,
An output means for identifying a category corresponding to the class recognized by the pattern recognition means based on the candidate storage means, and outputting the identified category as a category to which an entry image for the entry field belongs;
Program to function as.

Images