JP2010061471A - Character recognition device and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To reduce worker's load of work for determining whether the results of character recognition are correct or erroneous. <P>SOLUTION: In a character recognition device, character recognition processing is performed for the scanned document image by using a character image segmenting part 12 and a character recognizing part 13 to generate character code, next a character image generating part 14 generates character image based on the character code, and then a character image comparing part 15 compares the character image generated based on the character code with the original character image segmented from the document images to output similarity. A correctness or error determining part 16 determines whether the results of recognition are correct or erroneous based on the degree of similarity. A recognition result outputting part 17 takes out the document image including character parts being typically likely to be erroneously recognized and generates the output image data highlighting the character images being likely to be erroneously recognized based on coordinate data of the character images being likely to be erroneously recognized to show it to a user. <P>COPYRIGHT: (C)2010,JPO&INPIT

Description

  The present invention relates to a character recognition device and a program.

  A character recognition (hereinafter also referred to as OCR) device is used to extract information from a paper document, which is very convenient. However, the accuracy of OCR results is not perfect. The OCR result changes due to various factors such as the image quality of the input image, character cutting conditions, image conversion in the recognition process, OCR dictionary, and the like. At present, there is no definitive means for determining whether or not the OCR result is correct, and ultimately a human (operator: correctness / incorrectness determination person) is forced to make an inquiry / comparison with the original document or original image. Therefore, it is necessary for the operator to check all the input document (image) and the OCR result, which is a heavy load on the operator.

  Patent documents related to the present invention include the following.

  Patent Document 1 displays a file containing handwritten characters that could not be recognized in the thumbnail display of all pages in a different color so that it is easy to understand where (in which folder) there are unrecognizable characters. Has proposed. However, in character recognition (OCR), even if some character code is obtained, the character code is not always correct. For this reason, when seeking accuracy, it was necessary for the operator to confirm all of the results.

  In Patent Document 2, when a character image is encoded, a set of a character code and a character image is registered in a registration dictionary under an index, and the character image to be encoded is recognized as a character code. Correspondingly, the character image registered in the registration dictionary is read out, and then matched with the character image to be encoded to determine the matching degree. If the matching degree is equal to or higher than the reference, the corresponding index is used for encoding. If the matching degree is less than the standard, the encoding target character image is registered in a registration dictionary under a new index, and is then encoded using the index. The index stream and registration dictionary are sent to the decoding side and used for decoding.

  Citation 3 and Citation 4 propose to recognize a character by referring to a character image in the list using a list in which a character image and a character code are associated with each other through character recognition processing. .

These proposals do not determine whether the character recognition result is correct or incorrect.
JP 2006-106905 A Japanese Patent Laying-Open No. 2005-301662 JP-A-5-20500 JP 2000-90200 A

  An object of the present invention is to provide a character recognition technique for reducing the amount of correctness determination work of a character recognition result by an operator.

  According to the first aspect of the present invention, the character recognition device includes: a character recognition unit that executes character recognition processing on a character image to be recognized to generate a character code; and a character from the character code generated by the character recognition unit. A character image generating means for generating an image; an image comparing means for comparing the character image to be recognized with the character image generated by the character image generating means; and the character recognition based on a comparison output of the character image comparing means. And determining means for determining the correctness of the recognition result of the means.

  According to a second aspect of the present invention, in the character recognition device according to the first aspect, the recognition target character image determined by the determination means to be erroneous in the recognition result is included in the page including the recognition target character image. The image further includes display means for displaying the image separately from the others.

  According to a third aspect of the present invention, in the character recognition device according to the first or second aspect, the image comparison unit adopts a feature amount of a type different from the feature amount employed by the character recognition unit for character recognition. The character image to be recognized is compared with the character image generated by the character image generating means.

  According to a fourth aspect of the present invention, in the character recognition device according to any one of the first to third aspects, the character image generation unit generates a character image of a different type of character font, and the image comparison unit The character image to be recognized is compared with the character image of the different type of character font.

  According to a fifth aspect of the present invention, in the character recognition device according to the fourth aspect, the type of the character font employed by the image comparison means is narrowed down by learning.

  According to a sixth aspect of the present invention, in the character recognition device according to the fifth aspect, the image comparison means narrows down by excluding the character font type having a low correlation with the character image to be recognized. .

  According to a seventh aspect of the present invention, in the character recognition device according to the fifth or sixth aspect, the type of the character font employed by the image comparing means is different for each character object type.

  According to an eighth aspect of the present invention, in the character recognition device according to any of the fifth to seventh aspects, the learning is performed on a character image of a character selected based on the appearance frequency.

  According to a ninth aspect of the present invention, the character recognition device according to any one of the fifth to seventh aspects is such that the learning is performed on a character image of a character selected based on the likelihood of the character recognition means. ing.

  According to the invention of claim 10, the computer: character recognition means for generating a character code by executing character recognition processing on the character image to be recognized; generating a character image from the character code generated by the character recognition means A character image generating means for comparing; an image comparing means for comparing the character image to be recognized with the character image generated by the character image generating means; a recognition result of the character recognizing means based on a comparison output of the character image comparing means; A character recognition program for functioning as a determination means for determining correctness is realized.

  These and other aspects of the invention are set forth in the appended claims and will be described in detail below with reference to examples.

  According to the first aspect of the present invention, the correctness / incorrectness determination work amount of the character recognition result can be reduced.

  According to the invention of claim 2, it is possible to immediately grasp the character image related to the recognition error in the document image.

  According to the third aspect of the present invention, it is possible to prevent an error from being mixed for the same reason as the character recognition in the recognition correctness determination.

  According to the invention of claim 4, it is possible to prevent the accuracy of character recognition from being determined inaccurately depending on the font type of characters in the document.

  According to the fifth aspect of the present invention, it is possible to prevent the character recognition correct / incorrect determination from being inaccurate depending on the font type of the characters in the document, and to reduce the number of character image matching steps.

  According to the invention of claim 6, it is possible to prevent the character recognition correct / incorrect determination from being inaccurate depending on the font type of the characters in the document, and to reduce the matching man-hours of the character images. , It can prevent misjudgment due to drastic narrowing of the number of font types.

  According to the seventh aspect of the present invention, it is possible to prevent the character recognition correctness / incorrectness determination from being inaccurate depending on the character font type for each area of the document.

  According to the invention of claim 8, the man-hour for optimization of the font type can be reduced.

  According to the invention of claim 9, the man-hour for optimization of the font type can be reduced.

  According to the invention of claim 10, it is possible to reduce the correctness determination work amount of the character recognition result.

  Examples of the present invention will be described below.

  First, Embodiment 1 of the present invention will be described.

  FIG. 1 shows a character recognition system 400 according to a first embodiment of the present invention as a whole. Although the character recognition system 400 is implemented using one computer system, typically a personal computer 200, the character recognition system 400 may be realized using a plurality of computer systems connected via a communication network. Good. Further, the character recognition system 400 may be realized using various information devices.

  In FIG. 1, an image reading apparatus 100 and a personal computer 200 are connected to a communication network 300. The image reading apparatus 100 optically scans a document (hereinafter also referred to as scanning) to acquire a document image, and typically includes an automatic document feeding mechanism. The character recognition system 400 is realized by installing a character recognition program in the personal computer 200 using the recording medium 201 or the communication medium. The personal computer 200 includes a known CPU, main memory, bus, input / output device, and the like, and realizes various functions of the character recognition system 400 by cooperating hardware resources and programs thereof. Various functions are shown as functional blocks in the figure.

  The character recognition system 400 includes an image input unit 10, an image storage unit 11, a character image cutout unit 12, a character recognition unit 13, a character image generation unit 14, a character image comparison unit 15, a correctness determination unit 16, and a recognition result output unit 17. The display unit 18 and the like are included.

  The image input unit 10 stores the document image received from the image reading apparatus 100 via the communication network 300 in the image storage unit 11. The document image is managed in units of pages, for example. The image input unit 10 supplies image data in page units or page part units to the character image cutout unit 12. The character image cutout unit 12 cuts out each character image from the image data and supplies the character image data and the coordinate data to the character recognition unit 13. The character recognition unit 13 generates a character code (character identifier) by analyzing a character image by an arbitrary method, and performs pattern matching with a standard template, various feature amounts, typically direction contributions, strokes Character recognition is performed using feature quantities and the like. The character recognition result (character code) is supplied to the recognition result output unit 17 together with the coordinate data of the cutout position. The character recognition result (character code) from the character recognition unit 13 is also supplied to the character image generation unit 14. The character image generation unit 14 holds font data, generates character image data from font data (contour data, pixel data, or the like) corresponding to the character code, and outputs the character image data. The character image data generated from the font data is sent to the character image comparison unit 15. The character image comparison unit 15 compares the character image cut out from the document image with the character image generated from the character code as the recognition result. The comparison method of the character image comparison unit 15 performs comparison using a method different from the character recognition method of the character recognition unit 13. The comparison method is not limited to this, but is based on a feature amount such as a pixel distribution, a positional relationship between black pixel contacts, and a line segment. For example, when character recognition is performed using a black pixel distribution (no black pixels in a predetermined area) as shown in FIG. 3 for character recognition of handwritten numerals, comparison is performed using template matching as shown in FIG. The comparison result output from the character image comparison unit 15 is output as a similarity (also referred to as a correlation or coincidence).

  The correctness / incorrectness determination unit 16 determines that the character recognition is correct when the degree of correlation or the degree of coincidence from the character image comparison unit 15 exceeds a predetermined threshold, and determines that the character recognition is incorrect otherwise.

  The recognition result output unit 17 receives a position image of a character image that is incorrectly recognized and a document image of a page unit or a part thereof, generates a document image that highlights the character image portion that is incorrectly recognized, and displays a display unit. 18 is supplied. The display unit 18 presents to the user a document image that highlights a portion of the character image that is incorrectly recognized. The recognition result output unit 17 typically holds a record including a character code, a coordinate position, and a recognition correct / incorrect flag for each character image, and based on this, a portion of the character image that has been recognized incorrectly is high. Although the written original image is generated, the present invention is not limited to this.

  FIG. 2 illustrates an operation example of the embodiment of FIG.

  Reference is mainly made to FIG. First, a scanned document image (also referred to as a document image) is subjected to character recognition processing using the character image cutout unit 12 and the character recognition unit 13 to generate a character code (S10). Next, the character image generation unit 14 generates a character image based on the character code (S11). Next, the character image comparison unit 15 compares the character image generated based on the character code with the original character image cut out from the document image, and outputs the similarity (S12). The correctness determination unit 16 determines the correctness of the recognition result based on the similarity (S13). The recognition result output unit 17 typically takes out a document image including a misrecognized character part, and highlights the misrecognized character image based on the coordinate data of the misrecognized character image. Output image data is generated and displayed to the user (S15). In this example, only the erroneous document, page, and page portion are displayed so that the manuscript image portion unnecessary for the correction operation is not displayed to reduce the amount of correction work, but the recognition error is determined. Pages that are not displayed may also be displayed.

  3 and 4 show a simplified example.

  In FIG. 3, a character image of a handwritten numeral is recognized based on whether or not there is a pixel in a predetermined area. In the case where there is no black pixel in the portion indicated by the satin in FIG. 3A, the character is recognized as “6”. FIG. 3B shows an example in which “6” is correctly recognized as “6”. However, in the case of the character image “5” shown in FIG. 3C, since there is no black pixel in the satin portion, it is erroneously recognized as “6”.

  FIG. 4 shows a method for comparing character images. In this example, template matching is performed to obtain the similarity (correlation). Since the extracted character image “6” has a large correlation with the template “6”, it is determined that the recognition result is correct. On the other hand, since the extracted character image “5” has a small correlation with the template “6”, the recognition result is determined to be an error. Then, as shown in FIG. 5, the character image portion “5” having a recognition error is highlighted with a display attribute different from that of the other portions.

  If the character code of the recognition result is accompanied by the coordinate position data, each character image of the document image can be associated with each character code of the recognition result. For example, the highlighted portion is displayed as a pointing device. Is designated in the text display of the recognition result on the basis of the coordinate data, the corresponding portion is similarly highlighted, and the operator can correct the text display.

  Also, if the character code of the recognition result is accompanied by a correct / incorrect flag, it is possible to highlight a portion having a recognition error when the recognition result is displayed, and by pointing this portion, for example, the corresponding in the original image. Parts can be highlighted as well.

  In this embodiment, a paper document is scanned using a device having an image input function such as a copying machine. A character code is obtained by performing OCR processing on the scanned document image using a device having a character recognition function. A character image is generated from the obtained character code. On the other hand, a character image is cut out from the scanned document image. The two types of character images thus obtained (a character image generated from the character code of the OCR result and a character image cut out from the scanned document image) are compared. As a result of the comparison, if the similarity between the two is high, it is determined that the character recognition result of the OCR device is correct. Conversely, if the degree of similarity is low, it is determined that there is a possibility of error in the character recognition result of the OCR device. The presentation to the operator (right / wrong judgment person) is typically only a document including a character that is determined to have an error, and at the same time, the character that has been determined to have an error is clearly indicated. Conventionally, the operator has been required to confirm the correctness of the character recognition result in all the regions in all the documents. However, in the processing of this embodiment, the operator need only confirm the specified characters.

  Next, a character recognition system 401 according to the second embodiment of the present invention will be described.

  In the second embodiment, in addition to the configuration of the character recognition system 400 of the first embodiment shown in FIG. 1, a generation control unit 19 that adjusts a character image generated by the character image generation unit 14, and a character image cut out from a document image A character image deformation unit 20 for performing various operations is provided.

  FIG. 6 shows the character recognition system 401 of the second embodiment as a whole. In this figure, portions corresponding to those in FIG.

  In this embodiment, a plurality of types of font data are prepared, and the character recognition result is accurately obtained without depending on the character font of the document by matching the cut-out character image with the character image of each character type of various font sets. It is possible to judge the correctness of At the beginning of character recognition, while matching a wide variety of character images of various font types, a font type having a high likelihood of matching in the process is narrowed down to gradually reduce the amount of matching work.

  In addition, binarization threshold processing, thinning, expansion / contraction processing, and the like are performed on the cut-out character image so that proper character images can be compared.

  Details will be described below.

  In FIG. 6, the generation control unit 19 instructs the font type of the character image generated by the character image generation unit 14. The generation control unit 19 receives correct / incorrect data indicating whether the recognition result is correct and whether the font type is correct or incorrect from the correct / incorrect determination unit 16, and adapts (learns) the generated font type. This point will be described later in detail with reference to FIGS. By adapting the font types to be generated, the number of font types used for comparing character images is narrowed down to reduce the matching processing amount.

  The character image deforming unit 20 receives correct / incorrect data indicating the correctness of the recognition result from the correct / incorrect determination unit 16 and applies deformation to the cut out character image to optimize comparison. For example, if the cut-out character image is bold, the degree of coincidence may normally be small, but by thinning this, it is possible to maintain a sufficient degree of coincidence and erroneously determine that the recognition is wrong You can avoid it. Further, the character image may be rotated in consideration of the skew of the document.

  Next, the overall font optimization will be described with reference to FIG.

  FIG. 7 shows an example in which comparison is performed using character images generated from font data of a plurality of font types. In this example, first, character images of all fonts are generated, and character images are sequentially compared. In the example shown in the figure, the character image of the clipped character image is recognized to obtain the character code “A”, the character images “A” of various fonts corresponding thereto are output and sequentially matched.

  In accordance with the progress of character recognition, the number of fonts to be used may be gradually reduced every time one character is recognized or every time a predetermined number of characters are recognized (also referred to as a recognition stage). For example, the number of fonts is reduced by half for each recognition stage until the desired number of fonts, typically 1, is reached. At that time, fonts are left in descending order of similarity.

  You may make it optimize a font only on the character with a high appearance frequency.

  FIG. 8 shows an example in which an object is determined by analyzing a layout to determine whether a character image corresponds to a body text, a bar title, a title, or the like, and font optimization is performed for each object type. In this example as well, the font may be optimized based only on characters having a high appearance frequency.

  FIG. 9 shows an example in which font optimization is performed on the basis of a character image having a high certainty (likelihood) for character recognition. In this example, a character with a high certainty factor and its coordinate position are stored, and font type optimization (narrowing) is performed based on this character.

  FIG. 10 shows a comparative example of character images. In this example, the character image “a” is cut out, and the character recognition results are “a”, “me”, and “o”, respectively. , “Me”, or “O” character images generated from character codes are compared. In this comparison, the similarity is calculated by a method different from the method used in character recognition. As described above, a typical calculation method is based on pixel distribution, contact positional relationship, line segment, and the like. By the way, even when the recognition result is correct “A”, depending on the appearance of the cut-out character image, the degree of similarity may be small, and it may be erroneously determined as erroneous recognition. For this reason, the character image that has been cut out is preprocessed by the character image deforming unit 20 so that the similarity is not accidentally reduced. Pre-processing includes, for example, binarization processing, thinning, expansion / contraction processing, rotation, and the like, but is not limited thereto.

  In addition, this invention is determined based on description of a claim, and is not limited to the specific structure of the Example, a subject, and an effect. The present invention is not limited to the above-described embodiments, and various modifications can be made without departing from the spirit of the present invention. Hereinafter, some of the technical features of the present invention will be listed. Of course, the present invention is not limited to the following technical features.

  The image feature amount used in the character image comparison processing is typically different from the feature amount used in the OCR processing, but may be the same.

  The input character images (character images cut out from the scanned images) determined to have the correct character recognition result may be sequentially superimposed, and the superimposed character images may be used for subsequent comparison / determination.

  When generating images from character codes after OCR, create character images for all fonts in the device (PC, etc.), increase the number of comparison characters, and drop them as candidates in descending order of similarity. Continue to focus on one type of font.

  When character recognition is performed, appearance character frequency information is acquired, and a font is determined using characters having a high appearance frequency. Since a character image with less noise can be obtained by overlapping characters with high appearance frequency, the accuracy of similarity calculation with a character image generated from a character code is high.

  When character recognition is performed, the font is determined using characters with high recognition confidence. If the font cannot be determined using the character code having the highest recognition certainty, the font is narrowed down using the second, third,.

  Object analysis is performed on the scanned document, and a candidate font such as a title, a chapter, or a body is switched and determined for each object type. Even with the same character code, comparison and determination are performed using a font that is presumed to be optimal for each object, such as bold font for titles and Mincho font for text.

  For character codes with a low appearance frequency, the corresponding input character images (character images cut out from scanned documents) are also few. Therefore, a plurality of cut character images are duplicated, and character recognition is performed after different image processing is performed on the duplicated character images. If the character codes output as the character recognition result are the same even after a plurality of processes (if it is robust), it is determined that the character recognition result is correct.

  For character codes with a low appearance frequency, a plurality of character images intentionally added with image distortion are created from one character code. Each created character image and input character image are referred to, and if the similarity is low in any created character image, it is determined that the character recognition result is incorrect.

  If it is determined that there is an error after image comparison / correction determination is performed once, the character image generated from the character code or the character image cut out from the input image, or both, noise removal, binarization threshold change, thinning, Different image processing methods and image processing parameters such as expansion and contraction are applied, and image comparison / correction determination is performed again. In any case, when the similarity is low, it is determined that the character recognition result is an error.

  At the time of scanning, images of multiple resolutions are prepared, such as standard resolution and high resolution. If it is determined that there is an error after image comparison / correction determination is performed once, characters are cut out from the high-resolution scan image, and image comparison / correction determination is performed again. (OCR processing can be omitted). Since a character image with less noise can be created from a font image in an arbitrary size, higher resolution is advantageous for calculating similarity.

  The above processing is performed in units of one document (one page), and when processing is started for the next (page) document, the method and parameters used in the previous document are preferentially executed.

  If there is no character that is determined to be erroneous in one document, the document is not displayed. That is, only documents with a possibility of error are displayed.

  If there is a large percentage of characters that are determined to be erroneous in one document, re-processing is attempted based only on that one document. When reprocessing is executed, character code appearance frequency information, superimposition of input character images, font designation for each object type, and the like are changed. The result of the original process, the result of the reprocess, and the result having the higher ratio of characters determined to be correct are displayed.

  If the character recognition is judged to be correct even after reprocessing, it is judged that the character recognition result is correct only for characters recognized as correct in both the original processing and reprocessing, and other characters are determined. Display as possible error.

  If the ratio of characters determined to be erroneous in a single document is large, and there are different determination conditions in the past, reprocessing is attempted using the different conditions. If the situation does not improve as a result of the reprocessing, a different parameter is searched again.

It is a block diagram explaining the structure of Example 1 of this invention. FIG. 6 is a diagram for explaining an operation example of the first embodiment. It is a figure explaining the simple example of the character recognition of Example 1. FIG. It is a figure explaining the simple example of the comparison of the character image of Example 1. FIG. It is a figure explaining the example of a display of the correct / incorrect determination result of Example 1. FIG. It is a block diagram explaining the structure of Example 2 of this invention. FIG. 10 is a diagram illustrating a font determination method according to the second embodiment. It is a figure explaining the determination method of the other font in Example 2. FIG. It is a figure explaining the determination method of the other font in Example 2. FIG. It is a figure explaining the comparison of the character image in Example 2, and the pre-processing of a character image.

Explanation of symbols

DESCRIPTION OF SYMBOLS 10 Image input part 11 Image memory | storage part 12 Character image cutting part 13 Character recognition part 14 Character image generation part 15 Character image comparison part 16 Correct / error determination part 17 Recognition result output part 18 Display part 19 Generation control part 20 Character image deformation | transformation part 100 Image reading apparatus 200 Personal computer 201 Recording medium 300 Communication network 400 Character recognition system 401 Character recognition system

Claims (10)

Character recognition means for generating a character code by executing character recognition processing on a character image to be recognized;
Character image generation means for generating a character image from the character code generated by the character recognition means;
Image comparing means for comparing the character image to be recognized and the character image generated by the character image generating means;
A character recognition apparatus comprising: a determination unit that determines whether the recognition result of the character recognition unit is correct based on a comparison output of the character image comparison unit.   2. The display unit according to claim 1, further comprising: a display unit configured to display the recognition target character image, which is determined to have an error in the recognition result, as distinguished from others in the page image including the recognition target character image. Character recognition device.   The image comparison means employs a feature quantity of a type different from the feature quantity used by the character recognition means for character recognition, and compares the character image to be recognized with the character image generated by the character image generation means. Item 3. The character recognition device according to Item 1 or 2.   The character image generation unit generates a character image of a different type of character font, and the image comparison unit compares the character image to be recognized with the character image of the different type of character font. The character recognition apparatus in any one.   The character recognition device according to claim 4, wherein the type of the character font employed by the image comparison means is narrowed down by learning.   The character recognition apparatus according to claim 5, wherein the image comparison unit performs narrowing down by excluding the character font type having a low correlation with the character image to be recognized.   The character recognition apparatus according to claim 5 or 6, wherein a type of the character font employed by the image comparison unit is different for each character object type.   The character recognition device according to claim 5, wherein the learning is performed on a character image of a character selected based on the appearance frequency.   The character recognition device according to claim 5, wherein the learning is performed on a character image of a character selected based on the likelihood of the character recognition means. Computer
Character recognition means for generating a character code by executing character recognition processing on a character image to be recognized;
Character image generation means for generating a character image from the character code generated by the character recognition means;
Image comparison means for comparing the character image to be recognized and the character image generated by the character image generation means;
A character recognition program for functioning as a determination unit that determines whether the recognition result of the character recognition unit is correct based on the comparison output of the character image comparison unit.