JP2014115781A - Character recognition device and method, and character recognition program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To perform efficient rejection for reducing an erroneous reading rate while suppressing the drop width of a correct reading rate, and to reduce a calculation amount for rejection determination.SOLUTION: Rejection determinations 109 and 113 based on various indexes are serially combined when rejection values are highly independent of each other, and parallely combined when the independence is low. The rejection index of a high rejection rate and the rejection index of calculation cost are arranged before processing. A character identification part 106 recognizes a character in the image of each character unit cut out by a character cutting-out part 105. A plurality of rejection value calculations 107, 108, and 110 to 112 are arranged with the rejection value calculations 107 and 108 of high rejection performance set first in order. In the rejection determination 109, when rejection is determined based on rejection values calculated in the preceding rejection value calculations 107 and 108, whether a recognition result is rejected by omitting subsequent rejection value calculation processing.

Description

The present invention relates to a character recognition device and method, and a character recognition program, and more particularly, to an optical character recognition device and method and a character recognition program having a rejection determination method in which a plurality of rejection values are combined. In addition, the present embodiment relates to a rejection technique among character recognition techniques.

The technical field relates to an optical character recognition (OCR) apparatus. The OCR device reads a paper document with a scanner or the like, recognizes characters or symbols in the image, and encodes them into Unicode or the like, thereby digitizing them. OCR devices are digitized such as accounting slips, paid notices, salary reports, order forms, general payments, medical remuneration statements, answer sheets, etc. in general companies, local governments, financial institutions, medical institutions, educational institutions, etc. Used for. For general users, it is used for character recognition by a mobile phone, character recognition in a general document such as a memo, and the like.
A flow of document digitization processing by the OCR apparatus will be described in a simplified manner.
FIG. 6 is a diagram for explaining the flow of document digitization by the character recognition device. First, preprocessing such as imaging, binarization, and noise processing of a document by a scanner or the like is performed. Thereby, for example, a binary document image such as reference numeral 601 in FIG. 6 is obtained. Next, a character string image is obtained, for example, as indicated by reference numeral 602 in FIG. 6 by layout analysis and character string extraction such as the position of a chart and the paragraph structure of a document by the OCR device. After that, the OCR device cuts out character-by-character images by cutting out characters from the character string image, and then recognizes the characters in the individual images. Processing from document imaging to character string extraction is described in, for example, Patent Document 1 and Patent Document 2. Moreover, the process until it recognizes each character from a character string image is described in the patent document 3, the nonpatent literature 1, and the nonpatent literature 2, for example.

The present technology relates to a technology for recognizing individual character images. Hereinafter, a technique for recognizing characters drawn in individual character images will be briefly described.
First, feature extraction processing for converting a character image into a vector value is performed. If the number of dimensions of the vector value is N, one character image is expressed as an N-dimensional vector by the feature extraction process. N-dimensional vectors extracted from character images of the same character type are distributed at close positions in the N-dimensional space.
FIG. 9 is a schematic diagram showing the situation. Circles, triangles, and squares represent vector values extracted from character images corresponding to character type A, character type B, and character type C, respectively.
Next, with reference to a character identification dictionary created in advance, a character drawn in the character image is identified based on a vector value extracted from the character image.
Here, first, the character identification dictionary will be described. The character identification dictionary stores, for example, an identification function fk (x) that takes an N-dimensional vector as an argument and takes a real value as a value for each character type k to be identified. The identification function fk (x) has a large value for an N-dimensional vector x generated from a character image in which the character type k is drawn, and an N-dimensional value generated from a character image in which other character types are drawn. The vector x is generated by learning in advance so as to take a small value. The value of the discriminant function fk (x) is called the similarity, likelihood, etc., of the vector x to the character type k. For example, in the case of recognition for numbers, there are ten identification functions f0 (x), f1 (x),..., F9 (x) corresponding to 10 character types of 0-9.
In character identification, the value of the identification function fk (x) of each character type is calculated using the N-dimensional vector x extracted from the character image. Since the value of the discriminant function fk (x) is the similarity to the character type k, the character type k having the largest value of fk (x) is the first candidate for the recognition result. Similarly, the character type k for the discriminant function having the second largest value is the second candidate for the recognition result. In this way, recognition results are obtained up to the nth candidate.

FIG. 7 is a diagram for explaining the result of character identification. For example, the recognition of the character image cut out by the character cutout (reference number 603) in FIG. 6 is as shown in FIG. As described above, a recognition result is obtained as indicated by reference numeral 604 in FIG. 6 and converted into a code such as a character code that can be handled by the computer.
The character identification described above is a process of calculating a similarity between a character image and each recognition target character type, and obtaining a candidate character based on the calculated similarity. In order to increase the usefulness of the OCR device, the accuracy of this character identification is important. However, if the recognition result is suspicious, it is also important to reject the recognition result to inform it.
FIG. 12 is a diagram for illustrating examples of non-characters and ambiguous characters. Examples of objects to be rejected include non-characters as shown in a character example 1201 in FIG. 12 and ambiguous characters as shown in a character example 1202. Non-characters include, for example, a part of characters due to a mistake in character extraction, an image in which a plurality of characters are combined, or a mixture of disturbance factors such as dirt. An ambiguous character is, for example, a character that cannot be distinguished from 7 and 9 as in the leftmost image of the character example 1202.
If the rejection process is elaborate, there are several advantages. One is that if the result is saved with erroneously recognized characters, the entire recognized result must be manually rechecked to remain incorrect or to correct this. On the other hand, if the recognition result is suspicious and can be notified to the user, the user only has to correct that portion. Also, if the rejection can be performed with high accuracy, it is determined that there is a possibility that the previous processing may have failed, such as preprocessing, character line extraction, character extraction, etc. You can try the process again by changing the processing method and processing conditions. Thereby, recognition accuracy can be raised.

Hereinafter, the rate of correctly recognizing characters in a character image is referred to as a correct reading rate, the rate of erroneously recognizing is referred to as a misreading rate, and the rate of rejecting recognition results is referred to as a rejection rate. The sum of the correct reading rate, the misreading rate, and the rejection rate is 1. In general, if the rejection is made too strong, not only the misreads are rejected, but also some of the correct readings are rejected, so both the correct reading rate and the misreading rate are lowered. For this reason, it is desirable that the rejection does not decrease the correct reading rate as much as possible and reduces the misreading rate.
The method of rejection will be explained. Let an N-dimensional vector extracted from the input image be x. Also, the identification function corresponding to the first candidate character k1 is assumed to be fk1. At this time, fk1 (x) is the similarity to the character type k1. If r1 (x) = − fk1 (x), r1 (x) can be regarded as a dissimilarity with respect to the character type k1. Therefore, the threshold value h1 is set in advance, and when r1 (x)> h1, it is determined that the dissimilarity is high (similarity is low) and the rejection is determined. In this case, when the input image is non-character, it is assumed that the degree of similarity is low even for the first candidate character, and therefore non-character rejection is assumed.
Further, the discriminant function corresponding to the second candidate character k2 is assumed to be fk2. At this time, fk2 (x) is the similarity to the character type k2. Further, fk1 (x) ≧ fk2 (x). When r2 (x) = fk2 (x) −fk1 (x), the larger the value of r2 (x), the closer the values of fk1 (x) and fk2 (x) are. At this time, the identification is ambiguous between the first candidate character and the second candidate character. Therefore, the threshold value h2 is set in advance, and when r2 (x)> h2, the identification result is rejected as ambiguous.
FIG. 13 is a diagram for illustrating an example of an image to be rejected.
In addition, in Patent Document 4, a character blurring degree r3 (x) like a character example 1301 in FIG. 13 and a character crushing degree r4 (x) like a character example 1302 are calculated and used as the basis. Describes how to make a rejection decision. A threshold value h3 is set in advance, and when r3 (x)> h3, the blur is so large that it is rejected. Further, a threshold value h4 is set in advance, and when r4 (x)> h4 is satisfied, it is rejected because the collapse is large.

JP 2010-244372 A Japanese Patent Laid-Open No. 11-53466 JP 2004-171316 A Japanese Patent Application No. 2011-212308

Mohammed Cheriet, Nawawa Kharma, Cheng lin Liu, and Ching Suen. Character Recognition Systems: A Guide for Students and Practitioners. Wiley-Interscience, 2007. Kenichiro Ishii, Noriyoshi Ueda, Eisaku Maeda, Hiroshi Murase. Pattern recognition. Ohm Publishing Office.

Various indices such as the non-character degree (dissimilarity) r1, the ambiguity r2, the faint degree r3, and the collapse degree r4 can be considered as indices for rejecting characters. However, how to combine these indicators is not clear. In the prior art, a simple method such as rejecting those rejected according to any of the criteria, or a method of combining a plurality of indicators by trial and error manually is employed.
In the former simple method, since it is necessary to calculate all the rejection indexes, the calculation cost is high. In addition, since it is rejected when any of the rejection indicators exceeds the threshold, it is generally assumed that the rejection is too strong and the correct reading rate is lowered, so that a high correct reading rate and a low misreading rate are achieved. Even for this purpose, it is not always suitable. Further, the latter manual trial and error is a method that requires a considerable amount of cost when the number of indicators increases, and it may be difficult to implement.
In view of the above, an object of the present invention is to provide a high correct reading rate, a low misreading rate, and a high-speed rejection method at a low human cost.

According to the first solution of the present invention,
A plurality of rejection value calculation units for calculating a rejection value by a preset rejection function for the recognition result of the character identified from the input image;
One or more rejection determination units for determining whether or not to reject the recognition result based on one or more rejection values calculated by any one or more of the plurality of rejection value calculation units, respectively,
With
Using the plurality of rejection value calculation units combined based on the correlation of the plurality of rejection value calculation units, the rejection determination unit makes a rejection determination of the recognition result based on a plurality of rejection values, and rejects By rejecting the determined recognition result, a character recognition device is provided, wherein the recognition result that is not determined to be rejected is stored in a storage unit or displayed on a display unit.

According to the second solution of the present invention,
A character recognition method,
For a recognition result of characters identified from the input image, using a plurality of rejection value calculation units for calculating a rejection value by a preset rejection function,
Based on one or more rejection values calculated by any one or more of the plurality of rejection value calculation units, respectively, using one or more rejection determination units to determine whether to reject the recognition result,
Using the plurality of rejection value calculation units combined based on the correlation of the plurality of rejection value calculation units, the rejection determination unit makes a rejection determination of the recognition result based on a plurality of rejection values, and rejects By rejecting the determined recognition result, a character recognition method is provided, wherein the recognition result that is not determined to be rejected is stored in a storage unit or displayed on a display unit.

According to the third solution of the present invention,
A character recognition program,
The processing unit uses a plurality of rejection value calculation units, a function of calculating a rejection value by a preset rejection function for the recognition result of the character identified from the input image,
The processing unit rejects the recognition result based on one or a plurality of rejection values calculated by any one or a plurality of the rejection value calculation units using one or a plurality of rejection determination units. A function to determine whether or not
The processing unit uses the plurality of rejection value calculation units combined based on the correlation of the plurality of rejection value calculation units, and the rejection determination unit performs rejection determination of the recognition result based on the plurality of rejection values. A character recognition program for causing a computer to execute a function of storing the recognition result that is not determined to be rejected in a storage unit or displaying it on a display unit by rejecting the recognition result determined to be rejected.

According to the present embodiment, a high correct reading rate, a low misreading rate, and a high-speed rejection method can be provided at a low human cost.

It is an example of the flowchart explaining the process of the character recognition apparatus of Example 4 of this invention. It is an example of the block diagram of a character recognition apparatus. It is a figure for demonstrating two rejection values with high independence. It is a figure for demonstrating two rejection values with low independence. It is an example of the flowchart explaining the process of the character recognition apparatus of the related technology of this invention. It is a figure for demonstrating the flow of document digitization by a character recognition apparatus. It is a figure for demonstrating the result of character identification. It is a figure which shows the example of a rejection value. It is a figure for demonstrating the system for character identification. It is a figure for demonstrating a character extraction process. It is a figure for demonstrating a character recognition and recognition result selection process. It is a figure for showing the example of a non-character and an ambiguous character. It is a figure for showing the example of the image used as rejection object. It is a figure for showing an example of processing of feature extraction. It is a figure for showing the example of the character image database for learning. It is a figure which shows the rejection area | region in the case of a serial structure. It is an example of the flowchart explaining the process of the character recognition apparatus of Example 1 and Example 2 of this invention. It is an example of the flowchart explaining the process of the character recognition apparatus of Example 3 of this invention. It is explanatory drawing (1) about the gradient feature extraction method. It is explanatory drawing (2) about a gradient feature extraction method. It is explanatory drawing of a rejection function. It is a flowchart of the composition process of a rejection value.

Hereinafter, examples will be described with reference to the drawings.

1. Overview

In this embodiment, if an example is given,
The character recognition device
A document imaging unit for obtaining a document image by optically scanning the document;
A pre-processing unit having means for removing noise and background from the input image and binarizing to generate a binary image;
A layout analysis unit having means for analyzing the document structure and chart structure of the binary image;
A character string extraction unit having means for extracting an image of a character string unit from the binary image;
A character cutout unit having means for cutting out character-by-character images from each of the extracted character string images;
A character identification unit having means for recognizing characters in the image of each character unit cut out by the character cutout unit;
If a rejection value calculation means having a plurality of rejection value calculation means and having a high rejection capability is arranged earlier and is determined to be rejection based on the rejection value calculated by the previous rejection value calculation means, A rejection determination unit having means for determining whether or not to reject the recognition result by omitting the rejection value calculation process;
A recognition result selection unit having means for selecting a recognition result of each character string image based on the recognition result and the rejection determination result;
A retry determination unit having means for determining whether to perform reprocessing of recognition based on the recognition result;
A post-recognition processing unit having means for storing a recognition result and outputting the result to a display device;
Have

In the character recognition device of this embodiment, in the rejection determination unit, the higher the rejection efficiency based on the strength of the rejection ability and the rejection value calculation cost, the higher the rejection efficiency, and based on the rejection value calculated by the previous rejection value calculation means When it is determined as rejection, it may be characterized by determining whether or not to reject the recognition result by omitting a subsequent rejection value calculation process.

The character recognition device of this embodiment is
In the above-described rejection determination unit, a new rejection value is generated based on each of the rejection values of the plurality of rejection value calculation means arranged in parallel, and the rejection determination is performed based on the rejection value. Good.

The character recognition device of this embodiment is
The above-described rejection determination unit may include means for determining independence of a plurality of rejection values, and the rejection value calculation means having high independence may be processed in series.

The character recognition device of this embodiment is
The rejection determination unit described above may include means for determining the independence of a plurality of rejection values, and the rejection value calculation means having low independence may be processed in parallel.

The character recognition device of the present embodiment has means for determining the independence of a plurality of rejection values in the rejection determination unit described above, and as a means for determining the independence, the rejection image database based on the rejection values and the correct reading A function for identifying the image database is learned by a cost function based on the identification error, and the identification error by the function is compared with the identification error when the rejection value is configured in series, and the difference between the two is determined in advance. It may be characterized that it is determined that the independence is low when the threshold is equal to or greater than the threshold value, and that the independence is high in other cases.

2. Embodiment

An embodiment of a character recognition device having a rejection method will be described with reference to a diagram. The character recognition apparatus according to the present embodiment is an apparatus that digitizes an input document by detecting and recognizing characters in an input document image and encoding the characters. In addition to general documents, input documents include, for example, forms and specifications.
FIG. 2 is a configuration diagram illustrating an example of a character recognition apparatus according to the present embodiment.
A character recognition device 201 of this embodiment performs, for example, stamp recognition and form recognition, and includes an input device 202, a display device 203, an image acquisition device 204, a communication device 205, a calculation device (CPU) 206, and an external storage device. 207. The external storage device 207 includes a correctly read image database 211 and a reject image database 212.
The input device 202 is a keyboard, a mouse, or the like for inputting commands and the like. The input device 202 is a device for inputting a command executed for control of a program executed by the arithmetic unit (CPU) 206 and other control of connected devices.
The display device 203 is a device such as a display that appropriately displays processing contents.
The image acquisition device 204 is an image acquisition device such as a scanner. The acquired image may be stored in an external storage device or the like.
The communication device 205 is used for exchanging data from an external device such as a PC or a server. The communication device 205 is used for purposes such as acquisition of an execution command by a user from an external device and acquisition of information such as images and text from an external device. The communication device 205 is also used for purposes such as sending the processing content of the stamp recognition and form recognition device 201 to an external device.
An arithmetic unit (CPU) 206 is an arithmetic unit that executes processing such as generation of a recognition dictionary used for character recognition in a document image.
The external storage device 207 is an external storage device such as an HDD or a memory. The external storage device 207 stores various data such as a form image, a seal image, and a seal recognition dictionary. The external storage device is also used for temporarily storing data generated during processing executed by the arithmetic unit (CPU) 206.
The input device 202, the display device 203, the image acquisition device 204, and the communication device 205 may be omitted. When there is no input device 202, the process is started by an instruction from an external device using the communication device 205, or automatically by time designation or the like. If there is no display device 203, the processing result is transmitted to an external device using the communication device 205 or stored in the external storage device 207.
The output and input of the module that executes processing may be performed via the external storage device 207. That is, when the processing unit 1 outputs a processing result to the processing unit 2 and the processing unit 2 receives the processing result as an input, the processing unit 1 actually outputs the processing result to the external storage device 207 and stores it. In addition, the processing unit 2 may acquire the output result of the processing unit 1 stored in the external storage device 207 as an input.

Next, a description will be given of processing performed by the character recognition device 201 in the present embodiment.
Below, the process of the character recognition apparatus by the related technique of this invention is demonstrated using FIG. Thereafter, the processing of this embodiment will be described with reference to FIG.
First, processing of the character recognition device according to the related art of the present invention will be described.
FIG. 5 shows a typical example of the flow of document digitization by the character recognition device.
In document imaging (scanning) 101, the CPU 206 of the character recognition device 201 reads a document with a scanner or the like and converts it into an image. At this time, when background printing is printed in color, the CPU 206 may perform processing such as color dropout for optically removing printing of a specific color. Input documents include general documents, forms, and mark sheet paper created for the purpose of processing by a character recognition device from the beginning.
In the pre-processing 102, the CPU 206 performs processing such as binarization (monochromeization) of a color image of a document image, noise removal, and removal of unnecessary portions such as background printing. The binary image after the preprocessing is, for example, a form image 601 in FIG.
In the layout analysis 103, the CPU 206 performs a layout analysis of the binary image, and recognizes the position of the chart, the paragraph structure, the position of items and data, and the like. Regarding the position of the item and the data, for example, in the case of reference number 602 in FIG. 6, the CPU 206 determines that the payment amount above the column of the reference number 602 is the item name due to the table structure, and It is analyzed that the frame in which 890 and 123 are described is a data frame. In the case of a thesis or technical report, metadata extraction such as recognizing the position where the title, author, abstract, page number, etc. are written may be performed from the structure and positional relationship of the document.
In the character string extraction 104, the CPU 206 extracts a character string unit image from the document image. The CPU 206 extracts an image in units of character strings such as an image for one line in the case of a general document and an image in a frame in the case of a table. For example, an image within a table frame is extracted as indicated by reference numeral 602 in FIG.
In a series of processes of character extraction 105, character recognition 503, and recognition result selection 114, characters in each extracted character string image are recognized. In this process, as shown by reference number 603 in FIG. 6, the character string image is divided into character units, and characters in each character image are recognized. Convert to a code that can be handled by a computer.
The processing from the character extraction 105 to the recognition result selection 114 after the character string extraction 104 will be described with an example.

FIG. 10 is a diagram for explaining the character cutting process.
First, the character extraction 105 will be described. For example, assume that a character string image such as an image 1001 in FIG. 10 is obtained by character string extraction. First, in the process of character cut-out 105, the CPU 206 creates cutting candidate points based on points where the character lines intersect each other, points where the character lines are interrupted, or the like. An image 1002 in FIG. 10 shows division by cutting candidate points. In this example, it is divided into four images. The combination of each divided image and a plurality of adjacent images becomes a character image candidate. In the example of the image 1003 in FIG. 10, the first and second images from the left and the second and third images from the left also obtain six character image candidates as character image candidates, respectively. Each route from the left end point to the right end point from the left to the right is a cutout candidate of the character string 1001.
FIG. 7 is a diagram for explaining the result of character identification.
Next, in character recognition 503, the CPU 206 recognizes characters in individual character images that are candidates. Here, for example, as shown in FIG. 7, the correct candidate character (first candidate character type) for each character image and the similarity (likelihood and reliability) for the correct candidate character are obtained.
Next, the CPU 206 creates a network as a recognition result candidate as indicated by reference numeral 1101 in FIG. 11 based on the correct candidate character obtained by the character recognition 503 and the similarity. A reference numeral 1102 is obtained by removing the image. Each route from the left end point to the right end point from the left to the right is a recognition result candidate. Here, if the CPU 206 determines that the reliability of the recognition result of the character image is low, the CPU 206 performs a rejection process and sets a rejection flag for the recognition result to confirm that the reliability of the recognition result is low. Inform later user or user.
The internal processing of this character recognition 503 will be described. Here, the CPU 206 recognizes characters drawn in individual character images. In addition, the recognition result is rejected.
First, the character identification 106 will be described. Here, first, the CPU 206 performs a feature extraction process for converting a character image into a vector value. If the number of dimensions of the vector value is N, one character image is expressed as an N-dimensional vector by the feature extraction process. By expressing the character image as a vector value, the distribution of the character image can be statistically handled.

FIG. 14 is a diagram illustrating an example of feature extraction processing.
Feature extraction will be described with reference to FIG. First, the CPU 206 normalizes the character image. In general, input character images have different sizes. Therefore, in normalization, the sizes of character images are made uniform so that they can be handled uniformly in later processing. Also, the input character image may have a large character shape even if it is a character of the same character type due to differences in writing tools, writers, fonts, and the like. This causes a reduction in recognition accuracy. Therefore, in the normalization process, the size of the input character image and the deformation of the character shape are reduced to reduce the variation in character shape among the same character type. An image 1401 in FIG. 14 is an example of an input character image, and an image 1402 is an image deformed to a size of 64 × 64. There are various methods for normalization processing, which are described in detail in Non-Patent Document 1, for example.
Next, feature extraction is performed to convert the normalized image generated by normalization into a vector value. There are various methods for feature extraction, which are described in detail in Non-Patent Document 1, for example. Here, description will be made using the simplest example of pixel feature extraction. In pixel feature extraction, the normalized image is divided into small regions. In the example of FIG. 14, the normalized image 1402 is divided into 64 small regions. A state of division is shown in an image 1403. Next, it is converted into a vector value having the number of black pixels in each small region as an element. Since there are 64 small regions, a 64-dimensional vector value is generated as in the image 1404.
As another example of a widely used feature extraction method, a gradient feature extraction method will be described.

19 and 20 are explanatory diagrams (1) and (2) for the gradient feature extraction method.
Here, it is assumed that the normalized image generated by normalization has a white edge for one pixel. Further, the pixel value of the normalized image at the pixel point (i, j) is set to f (i, j). At this time, the CPU 206 calculates the gradient vector g = (gx, gy) at each pixel point (i, j) of the normalized image as follows. This corresponds to applying the filter shown in FIG.
gx (i, j) = {f (i + 1, j + 1) + 2f (i, j + 1) + f (i-1, j + 1) -f (i + 1, j-1) -2f (i, j-1) -f (i -1, j-1)} / 8
gy (i, j) = {f (i + 1, j + 1) + 2f (i + 1, j) + f (i + 1, j-1) -f (i-1, j + 1) -2f (i-1, j) -f (i -1, j-1)} / 8
However, in the above formula, when the pixel point (i, j) is at the edge of the image, the surrounding pixel points may be outside the image area. At that time, the value of f in the area outside the image is considered to be 0, and the above formula is calculated. As a result, a gradient vector g = (gx, gy) of pixel values is obtained at each pixel point (i, j).
Next, the CPU 206 converts the vector g (i, j) into eight directions g0 (i, j), g1 (i, j),..., G7 (i, j) at intervals of 45 degrees indicated by reference numeral 2001 in FIG. Disassembled into The decomposition is performed in two directions close to the direction of g (i, j). However, if the direction of g (i, j) completely matches any of the eight directions, there is no need for decomposition, and if it matches the direction 0, g0 (i, j) = vector g It is assumed that the length is (i, j), and g1 (i, j) =... = G7 (i, j) = 0 for the other directions. The decomposition method will be described with reference to the reference numeral 2002 in FIG. When g (i, j) exists between direction 0 and direction 1 as indicated by reference numeral 2002, the CPU 206 decomposes the vector g (i, j) into components of direction 0 and direction 1. At this time, assuming that the length of the component in the direction 0 is p0 and the length of the component in the direction 1 is p1, g0 (i, j) = p0, g1 (i, j) = p1, p2 (i, j) = ... = p7 (i, j) = 0.
As described above, eight direction images g0 (i, j),..., G7 (i, j) are generated. In order to improve robustness against deformation of characters, the image may be blurred by a Gaussian filter. In this case, the blurred direction images are collected and set as g0 (i, j),..., G7 (i, j). Next, the CPU 206 divides each direction image gi (x, y) into small areas, and generates a vector having the total value of the pixel values of each small area as an element. Now, assuming that each direction image is divided into 64 small regions, 64 values are obtained from each direction image. Since this is obtained for each direction, a total of 64 × 8 = 512 values are obtained in 8 directions. Using this as a vector component, a 512-dimensional vector is generated.
The above is the description of the gradient feature extraction method.

As described above, the CPU 206 converts the character image into a vector value. In the following, it is assumed that N is the number of dimensions of the vector value generated by feature extraction. Thus, each character image is expressed as a point on the N-dimensional space, and the same character type is distributed in a close region. The situation is schematically shown in FIG.
FIG. 9 is a diagram for explaining a system for character identification. Circles, triangles, and squares represent N-dimensional vector points extracted from character images corresponding to character type A, character type B, and character type C, respectively. For example, each circle represents a vector extracted from images of different character types A.

Next, the CPU 206 refers to a character identification dictionary created in advance, and identifies a character drawn in the character image based on a vector value extracted from the character image.
Here, first, the character identification dictionary will be described. The character identification dictionary stores, for example, an identification function fk (x) that takes an N-dimensional vector as an argument and takes a real value as a value for each character type k to be identified. The identification function fk (x) has a large value for an N-dimensional vector x generated from a character image in which the character type k is drawn, and an N-dimensional value generated from a character image in which other character types are drawn. The vector x is generated by learning in advance so as to take a small value. The value of the discriminant function fk (x) is called the similarity, likelihood, etc., of the vector x to the character type k. For example, in the case of recognition for numbers, there are ten identification functions f0 (x), f1 (x),..., F9 (x) corresponding to 10 character types of 0-9.
The CPU 206 can create this identification function using, for example, a learning character image database including character images and character labels.
FIG. 15 is a diagram illustrating an example of a learning character image database. As shown in the figure, the character label is a coded correct label indicating a character drawn in the character image. The learning character image database can be created by collecting character images, for example, by letting a person write a designated character in a designated frame. The CPU 206 converts each image included in the learning character image database into an N-dimensional vector by the same method as described above. Here, based on the N-dimensional vector and the correct answer label, the CPU 206 sets the discrimination function fk (x) to a large value for the N-dimensional vector corresponding to the character type k and to other character types. The N-dimensional vector is generated by learning so as to take a small value. Various methods such as SVM (Support Vector Machine), neural network, Gaussian model, and LVQ (Learning Vector Quantization) can be used as the learning method of the discriminant function.
In character identification, the CPU 206 calculates the value of the identification function fk (x) for each character type using the N-dimensional vector x extracted from the character image. Since the value of the discriminant function fk (x) is the similarity to the character type k, the character type k having the largest value of fk (x) is the first candidate for the recognition result. Similarly, the character type k for the discriminant function having the second largest value is the second candidate for the recognition result. In this way, recognition results are obtained up to the nth candidate. For example, the recognition of the character image cut out by the character cutout 603 in FIG. 6 is as shown in FIG. As described above, a recognition result is obtained as indicated by reference numeral 604 in FIG. 6 and converted into a code such as a character code that can be handled by the computer.
The above is the description of the character identification 106.

The character identification described above is a process of calculating a similarity between a character image and each recognition target character type, and obtaining a candidate character based on the calculated similarity. In order to increase the usefulness of the OCR device, the accuracy of this character identification is important. However, if the recognition result is suspicious, it is also important to reject the recognition result to inform it.
FIG. 12 is a diagram for illustrating examples of non-characters and ambiguous characters. Examples of objects to be rejected include non-characters as indicated by reference number 1201 in FIG. 12 and ambiguous characters as indicated by reference number 1202. Non-characters include, for example, a part of characters due to a mistake in character extraction, an image in which a plurality of characters are combined, or a mixture of disturbance factors such as dirt. The ambiguous character includes, for example, an indistinguishable 7 and 9 such as the image at the left end of the reference number 1202.
If the rejection process is elaborate, there are several advantages. One is that if the result is saved with erroneously recognized characters, the entire recognized result must be manually rechecked to remain incorrect or to correct this. On the other hand, if the recognition result is suspicious and can be notified to the user, the user only has to correct that portion. Also, if the rejection can be performed with high accuracy, it is determined that there is a possibility that the previous processing may have failed, such as preprocessing, character line extraction, character extraction, etc. You can try the process again by changing the processing method and processing conditions. Thereby, recognition accuracy can be raised.
Hereinafter, the rate of correctly recognizing characters in a character image is referred to as a correct reading rate, the rate of erroneously recognizing is referred to as a misreading rate, and the rate of rejecting recognition results is referred to as a rejection rate. In general, if the rejection is made too strong, not only the misreads are rejected, but also some of the correct readings are rejected, so both the correct reading rate and the misreading rate are lowered. For this reason, it is desirable that the rejection does not decrease the correct reading rate as much as possible and reduces the misreading rate.
Below, the non-character rejection 501 and the ambiguous character rejection 502 which are the processes of a rejection determination part are demonstrated.
The non-character rejection 501 will be described. Let an N-dimensional vector extracted from the input character image be x. Also, the identification function corresponding to the first candidate character k1 is assumed to be fk1. At this time, fk1 (x) is the similarity to the character type k1. If r1 (x) = − fk1 (x), r1 (x) can be regarded as a dissimilarity with respect to the character type k1. For this reason, the CPU 206 determines a threshold value h1 in advance, and when r1 (x)> h1, determines that the dissimilarity is high (similarity is low) and is rejected. In this case, when the input image is non-character, it is assumed that the degree of similarity is low even for the first candidate character, and therefore non-character rejection is assumed.
Next, the ambiguous character rejection 502 will be described. The identification function corresponding to the second candidate character k2 is assumed to be fk2. At this time, fk2 (x) is the similarity to the character type k2. Further, fk1 (x) ≧ fk2 (x). When r2 (x) = fk2 (x) −fk1 (x), the larger the value of r2 (x), the closer the values of fk1 (x) and fk2 (x) are. At this time, the identification is ambiguous between the first candidate character and the second candidate character. Therefore, the CPU 206 determines a threshold value h2 in advance, and rejects that the identification result is ambiguous when r2 (x)> h2. This process may be skipped if the rejection determination has already been made in the non-character rejection 501.

The above is the description of the processing in the character recognition 503. This process is performed for each character image.
In the recognition result selection 114, the CPU 206 refers to a word dictionary or the like, and selects a final recognition result from recognition result candidates while comprehensively determining the recognition similarity (reliability) for each character. . For example, when the address is recognized, the word dictionary can be a dictionary that stores a list of addresses in advance. In the case of recognition of a general document, it becomes a word or the like.
The above is the processing from character extraction 105 to recognition result selection 114. This process is performed on each character string image.

Next, in the retry determination 115, the CPU 206 determines whether to change the process and reprocess the recognition. For example, the reprocessing may be performed on the entire document image, or may be performed in character string image units or character image units. For example, when there is a character with low similarity (likelihood or reliability) in the character string recognition result, when a result matching the word dictionary is not obtained, or when there is a character that cannot be read, the CPU 206 Do reprocessing. In the case of performing reprocessing, the CPU 206 tries recognition again by changing the processing method or changing processing conditions from any of the previous processing. For example, the binarization of the preprocessing 102 or the noise removal method is changed. Finally, in post-recognition processing 116, the CPU 206 performs processing such as saving the recognition result in a storage device or displaying the result on a display.
The above is the processing flow of the character recognition apparatus according to the related art of the present invention.

FIG. 13 is a diagram for illustrating an example of an image to be rejected. In addition to the above-described r1 and r2, the CPU 206 determines that the rejection index includes a character blurring degree r3 (x) as indicated by reference numeral 1301 in FIG. ) Is calculated and a rejection determination is made based on the calculation. A threshold value h3 is set in advance, and when r3 (x)> h3, the blur is so large that it is rejected. Further, a threshold value h4 is set in advance, and when r4 (x)> h4 is satisfied, it is rejected because the collapse is large. In addition, the center of gravity of the character image, the average value of the line width of the character line, and the like can also be used. For example, in the case of the barycentric position, if the character identification result is 8, but the barycentric position is greatly deviated from the center, a determination is made such as rejection.
Here, examples of the blurring degree r3 (x) and the squashing degree r4 (x) are given. In the above description, x is a vector extracted by feature extraction. Here, x is a normalized image. For each character type, the average total pixel value m of the normalized image is calculated in advance from the learning DB. For the input image, r3 (x) is a value obtained by subtracting the total pixel value of the normalized image of the input image from m, and r4 (x) is a value obtained by subtracting m from the total pixel value of the normalized image of the input pixel. Value. As a result, r3 is large when the total pixel value of the normalized image of the input image is smaller than m, and r4 is small when it is larger.
However, how to combine these indicators has not been clear so far. In the prior art, a simple method such as rejecting those rejected according to any of the criteria, or a method of combining a plurality of indicators by trial and error manually is employed.
In the former simple method, since it is necessary to calculate all the rejection indexes, the calculation cost is high. In addition, since it is rejected when any of the rejection indicators exceeds the threshold, it is generally assumed that the rejection is too strong and the correct reading rate is lowered, and that a high correct reading rate and a low misreading rate are achieved. Even for the purpose of rejection, it is not always suitable. Further, the latter manual trial and error is a method that requires a considerable amount of cost when the number of indicators increases, and it may be difficult to implement.

3. Character recognition

In the present embodiment, a rejection method that effectively combines a plurality of rejection indexes can be automatically configured. Thereby, the human cost for combining a plurality of rejection indicators can be reduced. Further, the misreading rate can be reduced while maintaining the correct reading rate at a high level, and a precise and high-speed rejection method can be configured.
Processing of the character recognition device of this embodiment will be described with reference to the drawings.
FIG. 17 is an example of a flowchart for explaining processing of the character recognition apparatus according to the embodiment of the present invention.
Document imaging 101, preprocessing 102, layout analysis 103, character string extraction 104, character extraction 105, character identification 106, recognition result selection 114, retry determination 115, and post-recognition processing 116 are described in FIG. As described above, the processing is the same as that of the related art character recognition device of the present invention.

Hereinafter, processing from processing 1701 to processing 1706, which is a rejection determination unit inside the character recognition 1707, will be described. In the rejection process, the CPU 206 makes a rejection determination using the result of the character identification 106 and the rejection value. When it is determined that the rejection is made, the CPU 206 sets a rejection flag on the character recognition result to notify the user of the subsequent processing or the user so that the result can be used.
In the configuration of the reject combination of this embodiment, a reject image database in which image samples to be rejected and a correct image database in which image samples to be correct are collected are prepared in advance. The reject image database is a database in which image samples to be rejected, such as samples misread by the character identification 106, non-character images, ambiguous character images, blurred images, collapsed images, and the like are collected. The correct reading image database is a database in which character image samples that are to be read correctly, such as those that can be correctly identified by the character identification 106 process, are collected. In the following, the proportion of correctly read image database samples that are rejected is referred to as an erroneous rejection rate, and the proportion of rejected image database samples that are not rejected is referred to as an erroneous acceptance rate. The smaller the error rejection rate and the error acceptance rate, the better the accuracy of the rejection determination.
In the following, assuming that there are n rejection value calculation units, the rejection values are numbered such as a rejection value 1, a rejection value 2,..., A rejection value n. Further, a function (rejection function) that outputs the rejection value with the image x as an input will be written as r1 (x), r2 (x),.
The nature of the rejection value will be briefly described. The rejection function ri (x) is configured to have such a property that it takes a high value for a sample that is desired to be rejected and a low value for a sample that is not desired to be rejected. For example, as described above, the degree of blur, the degree of collapse, the non-character level calculated using the value of the discriminant function, the degree of ambiguity, and the like. A threshold value h1 is provided, and is used as a rejection when ri (x)> h1. At this time, if h1 is too large, it cannot be rejected sufficiently and the misreading rate increases. On the other hand, if h1 is too low, the misreading rate decreases, but the correct reading rate also falls. Therefore, h1 is adjusted according to the user's request so as not to decrease the correct reading rate as much as possible and to reduce the misreading rate.

FIG. 16 shows a region of values determined to be rejected by hatching when it is determined that the reject value is exceeded when the threshold value is exceeded for any of the two reject values. When the rejection value 1 exceeds the threshold value 1 or when the rejection value 2 exceeds the threshold value 2, the rejection region is as shown by the hatched portion in FIG. 16.
In this embodiment, these n rejection values are arranged in descending order of the rejection strength. A high rejection strength means that the rejection rate of the rejection determination based on the rejection value is high. Here are some examples of how to set the rejection strength.
Take the first example. First, the sum e of the false rejection rate and the false acceptance rate is designated. For each rejection function ri, hi is set so that the sum e of the erroneous rejection rate and the erroneous acceptance rate when the rejection determination is performed by ri (x)> hi is minimized. At this time, the rejection values are selected in descending order of the rejection rate of the learning character image database samples when the rejection determination is performed by rejecting when ri (x)> hi.
Here is a second example. It is assumed that a threshold value hi is designated in advance by the user for each rejection function ri. At this time, rejection values are selected in descending order of the rejection rate of the learning character image database when rejection determination is performed by rejecting when ri (x)> hi.
Now, it is assumed that there are three rejection values and the rejection rates are high in the order of r1, r2, and r3, that is, the rejection strength is high. At this time, processing is performed in the order of processing 1701 to processing 1706 in FIG. In other words, the rejection value r1 (x) for the input image x is calculated in the rejection value 1 calculation 1701, and if r1 (x)> h1 in the rejection determination 1 (1702), it is determined as rejection, otherwise, Do not reject. If it is determined to be rejected, processing from processing 1703 to processing 1706, which is subsequent rejection processing, is skipped. If it is not determined to be rejected, the process proceeds to the next process 1703. Hereinafter, similarly, the processing of rejection determination 2 or the processing of rejection determination 2 and rejection determination 3 is continued. In the example, the case where there are three rejection values has been described, but the same applies to cases where there are two or more rejection values.
In the present embodiment, the processing can be completed when it is determined to be rejected. Furthermore, since the rejection rate is arranged first in order of high rejection rate, the calculation cost is efficient.

FIG. 2 is a configuration diagram illustrating an example of the character recognition apparatus according to the present exemplary embodiment, which is the same as that of the first exemplary embodiment. FIG. 17 shows a processing flow of the character recognition apparatus of this embodiment. Document recognition 101, pre-processing 102, layout analysis 103, character string extraction 104, character extraction 105, character identification 106, recognition result selection 114, retry determination 115, post-recognition processing 116 are also performed by the character recognition device. Similar to Example 1. The character identification 106 is the same as that in the first embodiment.
In a present Example, the flow of a process of each 1701-1706 of a rejection determination part differs.
In Example 1, the rejection value calculation process and the rejection determination process are arranged in descending order of the rejection intensity. This method is sufficient when there is not much difference in the calculation cost for calculating the rejection value, but it may be inefficient otherwise. For example, even if the rejection rate is high, if there is a rejection value calculation process with a high calculation cost for calculating the rejection value, a rejection value with a high calculation cost is always calculated. Here, the calculation cost is obtained as, for example, an average processing time for calculating a rejection function when processing an image included in the learning character image database.
Therefore, in this embodiment, the processing order is determined in consideration of the calculation cost (processing time) for calculating each rejection value. That is, based on the rejection efficiency determined based on the rejection rate of the rejection value and the calculation cost (processing time), the higher the rejection efficiency, the higher the rejection efficiency. The rejection efficiency can be calculated by, for example, rejection rate × calculation cost (average processing time).
FIG. 21 is an explanatory diagram of the rejection function.
When the rejection configuration of this embodiment is shown in a table, it is as shown in Table 2101 of FIG. Each row (horizontal direction) in the table indicates a parallel arrangement, and a rejection function to be combined and its synthesis function, and a column direction (vertical direction) indicates a series arrangement. In this embodiment, since all the rejection functions are connected in series, each column is one rejection function. Rejection value 1 calculation 1701, Rejection value 2 calculation 1703, Rejection value 3 calculation 1705 calculate f1 (r1 (x)), f2 (r2 (x)), and f3 (r3 (x)), respectively, and reject When there is only one rejection function in the parallel direction as in this embodiment, f1, f2, and f3 are assumed to be identity functions, for example, f1 (r1 (x)) = r1 (x) As good as

FIG. 2 is a configuration diagram illustrating an example of the character recognition apparatus according to the present exemplary embodiment, which is the same as that of the first exemplary embodiment. FIG. 18 shows the flow of processing of the character recognition apparatus of this embodiment. Document recognition 101, pre-processing 102, layout analysis 103, character string extraction 104, character extraction 105, character identification 106, recognition result selection 114, retry determination 115, post-recognition processing 116 are also performed by the character recognition device. Similar to Example 1. The character identification 106 is the same as that in the first embodiment.
In the present embodiment, processes 1801 to 1804 for performing rejection determination in character recognition 1805 are different. In this embodiment, as shown in processes 1801 to 1803, a plurality of rejection values are calculated in parallel, and a rejection determination process is performed in process 1804 based on these values.
First, the reason why the rejection value calculation is connected in parallel will be described.
FIG. 16 shows a region of values determined to be rejected by hatching when it is determined that the reject value is exceeded when the threshold value is exceeded for any of the two reject values. When the rejection value 1 exceeds the threshold value 1 or when the rejection value 2 exceeds the threshold value 2, the rejection region is as shown by the hatched portion in FIG. 16. This corresponds to the case where the rejection value calculation and the rejection determination are sequentially performed as in the first and second embodiments, and the processes are connected in series.
FIG. 4 schematically shows two rejection values, distributions of samples to be rejected, and samples to be correctly read. A triangle represents a sample of the reject image database, and a circle represents a sample of the correct reading image database. In the case of such a distribution, the boundary between the distribution of the sample of the correct reading image database and the distribution of the sample of the rejection image database is as shown in FIG. 4, and the sample to be rejected is the boundary line. It is located on the upper right side. On the other hand, when the rejection is performed in series, a rejection region as shown in FIG. 16 is obtained, and in this example, a large number of samples to be rejected cannot be rejected. If the values of threshold 1 and threshold 2 are made small so that these samples to be rejected can be rejected, many round samples to be correctly read will be rejected.
For this reason, in this embodiment, the rejection determination is performed based on both the rejection value 1 and the rejection value 2. That is, when the value of the rejection value 1 is x1, and the value of the rejection value 2 is x2, a new rejection value is determined by a function f (x1, x2) that takes these as arguments, and the value of f (x1, x2) is Reject if it is above a certain threshold. For example, f (x1, x2) = x1 + x2 can be used as f (x1, x2). Another example of how to define the function f (x1, x2) will be described.

The function f (x1, x2) is defined as a quadratic function f1 (x1, x2) = a11x1x1 + a22x2x2 + a12x1x2 + a1x1 + a2x2 + a0 of x1, x2 having a11, a22, a12, a1, a2, a0 as parameters. The parameters a11, a22, a12, a1, a2, and a0 are set so as to take a negative value with respect to the sample of the correct image database and to take a positive value with respect to the sample of the reject image database. . However, since it is generally impossible to set a parameter that satisfies this condition for all samples, a cost function (loss function) indicating the degree of not satisfying the condition by taking the parameter as an argument ) (Or a cost function based on the discrimination error between the sample of the correctly read image database and the sample of the reject image database) c (f) is defined, and learning is performed by machine learning so that this value becomes small. For example, assuming that f is 1 for samples in the reject image database and f is -1 for samples in the correct image database, c (f) is applied to all samples from these values. The sum of squared errors. c (f) is, for example, calculated from the sample of the correct reading image database by v1 = Σ | f−1 | ^ 2, which is the sum of the square error between the value of f calculated from the sample of the rejection image database and 1. The sum of square errors between the value of f and −1 is v2 = Σ | f + 1 | ^ 2, and c (f) = v1 + v2 (sum of square errors) or the like. For example, a neural network or SVM can be used. The contour line at which f = 0 created in this way becomes the boundary line between the distribution of the correct reading image database and the sample of the rejection image database, like the boundary line in FIG. Here, f has been described by taking a quadratic function as an example, but a more general function, for example, a higher-order function, a neural network, a linear combination of radial basis functions, or the like can also be used.
As mentioned above, in order to simplify explanation, although the case where it had two rejection values was demonstrated, it is the same also in the case of three or more rejection values. FIG. 18 shows the flow of processing when there are three rejection values. In processing 1801, processing 1802, and processing 1803, a rejection value 1, a rejection value 2, and a rejection value 3 are calculated, respectively. The rejection values are x1, x2, and x3. In rejection determination 3 (1804), f (x1, x2, x3) is larger than a predetermined threshold based on the new rejection value f (x1, x2, x3) created as described above. If it is not, it will be rejected, otherwise it will not be rejected.
The method of the present embodiment can perform rejection with higher accuracy than connecting in series. However, all of the rejection values must be calculated, and the value of f needs to be calculated based on the rejection values. Therefore, the case where the calculation cost concerning rejection becomes large is assumed.
A table of the rejection configuration of this embodiment is shown in Table 2102 of FIG. Each row (horizontal direction) in the table indicates a parallel arrangement, and a rejection function to be combined and its synthesis function, and a column direction (vertical direction) indicates a series arrangement. In the case of this embodiment, since all the rejection functions are connected in parallel, there is one line. The composite function is f, and the values calculated in the rejection determination 1804 are f (r1 (x), r2 (x), r3 (x)). f is, for example, a function created by the method described above.

FIG. 2 is a configuration diagram illustrating an example of the character recognition apparatus according to the present exemplary embodiment, which is the same as that of the first exemplary embodiment. FIG. 1 shows the flow of processing of the character recognition apparatus of this embodiment. Document recognition 101, pre-processing 102, layout analysis 103, character string extraction 104, character extraction 105, character identification 106, recognition result selection 114, retry determination 115, post-recognition processing 116 are also performed by the character recognition device. Similar to Example 1. The character identification 106 is the same as that in the first embodiment.
In the present embodiment, the combination of processes for performing rejection determination in the character recognition 117 (corresponding to the portions 107 to 113) is different.
In the configuration of the reject combination of this embodiment, a reject image database in which image samples to be rejected and a correct image database in which image samples to be correct are collected are prepared in advance. The reject image database is a database in which image samples to be rejected, such as samples misread by the character identification 106, non-character images, ambiguous character images, blurred images, collapsed images, and the like are collected. The correct reading image database is a database in which character image samples that are to be read correctly, such as those that can be correctly identified by the character identification 106 process, are collected. In the following, the proportion of correctly read image database samples that are rejected is referred to as an erroneous rejection rate, and the proportion of rejected image database samples that are not rejected is referred to as an erroneous acceptance rate. The smaller the error rejection rate and the error acceptance rate, the better the accuracy of the rejection determination.
In the following, assuming that there are n rejection value calculation units, the rejection values are numbered such as a rejection value 1, a rejection value 2,..., A rejection value n. In addition, a function that receives an image x and outputs a rejection value is written as r1 (x), r2 (x),..., Rn (x), and the like.

In this embodiment, these n rejection values are combined in consideration of the high independence between the rejection values, the low independence (high correlation), and the rejection efficiency. The policy of combination of the critical value calculators of the present embodiment connects the critical value calculators in series or in parallel. The combination of the independence value calculators with high independence is combined in series, the independence value calculators with low independence (high correlation) are combined in parallel, and the rejection value calculator with strong rejection strength comes first. Deploy. Further, when combining in parallel, a new rejection value is determined based on a plurality of combined rejection values, and a rejection determination is made based on the new rejection value. Furthermore, the processing with higher rejection efficiency is arranged earlier.
FIG. 16 shows a region of values determined to be rejected by hatching when it is determined that the reject value is exceeded when the threshold value is exceeded for any of the two reject values. When the rejection value 1 exceeds the threshold value 1 or when the rejection value 2 exceeds the threshold value 2, the rejection region is as shown by the hatched portion in FIG. 16.
First, the independence between rejection values will be described with reference to FIG. FIG. 3 schematically shows two rejection values, distributions of samples to be rejected, and samples to be correctly read. A triangle represents a sample of the reject image database, and a circle represents a sample of the correct reading image database. In such a distribution, the boundary line between the distribution of the sample of the correct reading image database and the distribution of the sample of the rejection image database has a large convex shape in the upper right direction as shown in FIG. In such a case, the two rejection values will be called highly independent. Such a situation may occur when a rejection value is calculated based on an event in which two rejection values are highly independent. For example, the rejection value 1 is a character blurring degree calculation, and the rejection value 2 is a case where a deviation degree of a character center of gravity position from a standard center of gravity position is calculated.

In this embodiment, when the rejection value is highly independent, the rejection value 1 calculation and the rejection value 2 calculation are processed in series. That is, first, after calculating the rejection value 1, if the value is higher than the threshold value 1, it is determined to be rejection. If it is determined to be rejected, the reject process is finished. If it is not determined to be rejected, a reject value 2 is calculated, and if the value is higher than threshold 2, it is determined to be rejected. If it is determined to be rejected, the reject process is finished. If it is not determined to be rejected, the process proceeds to the next reject process. As shown in FIG. 3, threshold 1 and threshold 2 are set, and when the rejection value 1 exceeds the threshold 1 or when the rejection value 2 exceeds the threshold 2, it is determined that the rejection is made, so that the rejection can be efficiently performed. it can. Such rejection values may be processed in series.
Next, the low independence between reject values (high correlation) will be described with reference to FIG. FIG. 4 schematically shows two rejection values, distributions of samples to be rejected, and samples to be correctly read. A triangle represents a sample of the reject image database, and a circle represents a sample of the correct reading image database. In such a distribution, when the boundary between the distribution of the sample of the correct reading image database and the distribution of the sample of the rejection image database is not as large as the case of FIG. Or, conversely, when it becomes convex in the lower left direction, the two rejection values will be referred to as having low independence. Such a situation may occur when a rejection value is calculated based on an event in which two rejection values are highly correlated. For example, the rejection value 1 is a case where the non-character level based on the discrimination function as described above is calculated, and the rejection value 2 is a case where the ambiguity level is calculated based on the discrimination function. In such a case, since the rejection value is calculated based on the discriminant function, they are related to each other and have a distribution as shown in FIG.
In this embodiment, when the independence of the rejection value is low, the rejection value 1 calculation and the rejection value 2 calculation are processed in parallel. That is, when the rejection value 1 is x1 and the rejection value 2 is x2, a new rejection value is determined by a function f (x1, x2) that takes these as arguments, and the value of f (x1, x2) is equal to or greater than a certain threshold value. Reject in case of. If it is determined to be rejected, the reject process is finished. If it is not determined to be rejected, the process proceeds to the next reject process. The method for defining the function f is the same as in the third embodiment. In the case of the distribution of FIG. 4, for example, if f (x1, x2) = x1 + x2, a threshold boundary line can be defined in an oblique direction from the upper left to the lower right, and a sample of the correctly read image database and the reject image database Samples can be separated. In the case of the distribution as shown in FIG. 4, when the rejection values are connected in series, only the region where the rejection value 1 is greater than the threshold value 1 or the rejection value 2 is greater than the threshold value 2 is rejected. In addition, triangular samples distributed below the threshold value 2 cannot be rejected. Further, if the values of the threshold value 1 and the threshold value 2 are lowered in order to reject these, a large number of round samples to be read correctly will be rejected. Therefore, such rejection values need to be connected in parallel.
As described above, in this embodiment, reject values with high independence are processed in series, and reject values with high parallelism are processed in parallel.

Here, an example of a method for determining whether two rejection values are high independence or low independence (high correlation) will be given. Two rejection values are set as x1 and x2, respectively. Two functions g1 (x1, x2) and g2 (x1, x2) are defined with these two rejection values as arguments.
The function g1 (x1, x2) is a quadratic function as in the third embodiment, and takes a negative value with respect to the sample of the correct image database and takes a positive value with respect to the sample of the reject image database. Based on the cost function c, it is set by machine learning. For example, as shown in FIGS. 3 and 4, g1 is a function in which a contour line where g1 = 0 is a boundary line, is negative in a lower left area from the boundary line, and is positive in an upper right area.
The function g2 (x1, x2) has two values h1 and h2 as parameters, and when x1> h1 or x2> h2, g2 (x1, x2) = 1, otherwise g2 (x1, x2) ) = − 1. That is, a region where g2 (x1, x2)> 0 is a rejection region. However, since it is generally impossible to set a parameter that satisfies this condition for all samples, a cost function c (h1, h2) that indicates the degree of not satisfying the condition by taking the parameter as an argument is used. Define and learn by machine learning so that this value becomes small. c (h1, h2) is, for example, the number of rejected image database samples with g2 = −1 being v1, the number of correctly read image database samples with g2 = 1 being v2, and c (h1, h2). ) = V1 + v2 (number of samples that do not satisfy the condition). For example, a neural network or SVM can be used. The boundary line between g2 = 1 and g2 = −1 of g2 created in this way is parallel to the axis of rejection value 1 or rejection value 2, and separates the distribution of the correct reading image database from the sample of the rejection image database. It becomes a boundary. 3 and 4, assuming that threshold 1 is a dotted line indicating h1 and threshold 2 is h2, g2 is a region to the left of threshold 1 and below threshold 2, g2 = -1, threshold 1 In the region on the right side or on the upper side of the threshold 2, g2 = 1.

The boundary between g2 = 1 and g2 = −1 by the function g2 (x1, x2) corresponds to the boundary of the rejection area when the rejection processing is connected in series. On the other hand, the contour line of g1 = 0 by the function g1 (x1, x2) corresponds to the boundary of the rejection area when the rejection processing is connected in parallel by the method of the third embodiment.
Here, the accuracy of the rejection region generated by the function g1 is compared with the accuracy of the rejection region generated by the function g2. The number of samples in the reject image database where g2 = −1 is v1, the number of samples in the correct image database where g2 = 1 is v2, and the number of samples in the reject image database where g1 <0. Is w1, and the number of samples in the correctly read image database satisfying g1 ≧ 0 is w2. v1 and w1 correspond to the number of false acceptances, and v2 and w2 correspond to the number of false rejections. The loss function in the case of the number of erroneous acceptance p1 and the number of false rejections p2 is set as h (p1, p2). h is a monotonically increasing function of p1 and p2. For example, h (p1, p2) = p1 + p2. In this case, it is the sum of the number of false acceptances and the number of false rejections. It can be considered that the smaller the value of h, the better the accuracy of rejection.
Next, h (v1, v2) and h (w1, w2) are compared. In general, the rejection region based on the function g1 has better accuracy, and h (v1, v2) is larger than h (w1, w2). Here, D = h (v1, v2) −h (w1, w2) represents a difference in loss between when connected in series and when connected in parallel. If this is greater than or equal to a certain value, it is determined that the rejection region by g2 is insufficient, and the independence between the rejection value 1 and the rejection value 2 is low. Conversely, when D = h (v1, v2) −h (w1, w2) is equal to or greater than a certain value, it is determined that the rejection value 1 and the rejection value 2 are highly independent.
In the above, for the sake of simplicity, two rejection values have been described, but the same applies to the case of three or more rejection values.

In this embodiment, processes with high independence are arranged in parallel, and processes with low independence are arranged in series. As in the case of the third embodiment, the above function g1 can be used as the rejection value when arranged in parallel. Moreover, like Example 2, it arrange | positions earlier, so that a rejection efficiency is high.
When there are n rejection values, an example of the method of constructing the rejection values will be given.
FIG. 22 is a flowchart of the reject value configuration process. This processing may be executed by the CPU 206 of the character recognition device 201 or another processing device other than the character recognition device 201. First, the CPU 206 or another processing device selects a pair having the lowest independence (the value of D is large) from among the n rejection values. When the value 206 for determining the independence is lower than a predetermined value, the CPU 206 or another processing device places the n reject values in series because the selected pair is highly independent. When the CPU 206 or another processing device determines that the selected pair is less independent, the selected pair is connected in parallel, and a new rejection value based on these rejection values is the same as in the method of the third embodiment. Determine. When the rejection values connected in parallel are regarded as one rejection value, there are n-1 rejection values. Similarly, the CPU 206 or another processing device selects a pair having the lowest independence from the n-1 rejection values. When it is determined that the selected pair is highly independent, the CPU 206 or another processing device arranges the n-1 rejection values in series. If it is determined that the selected pair is less independent, the CPU 206 or another processing device connects the selected pairs in parallel, and sets a new rejection value based on these rejection values in the same manner as in the method of the third embodiment. Determine. At this time, if the rejection values (referred to as r1 and r2) constituting the selected pair are made up of a plurality of rejection values in parallel, the CPU 206 or another processing device is the element constituting the rejection value. These rejection values are decomposed and connected in parallel, and new rejection values based on these rejection values are determined in the same manner as in the method of the third embodiment. For example, if r1 was originally configured by connecting two reject values s1 and s2 in parallel, the CPU 206 or another processing device decomposes r1 into the original reject values, and converts s1, s2, and r2 into Connect in parallel. As described above, the CPU 206 or another processing apparatus continues until there are finally no pairs that are determined to have low independence.

In FIG. 1, a rejection value 1 and a rejection value 2 are connected in parallel, a rejection value 3, a rejection value 4, and a rejection value 5 are connected in parallel, and the former set and the latter set are connected in series. Is shown.
The rejection configuration in the case of FIG. 1 is shown as a table 2103 in FIG. Each row (horizontal direction) in the table indicates a parallel arrangement, and a rejection function to be combined and its synthesis function, and a column direction (vertical direction) indicates a series arrangement. In the case of FIG. 1, first, the rejection value 1 and the rejection value 2 are connected in parallel, and the rejection value 3, the rejection value 4, and the rejection value 5 are connected in parallel. The rejection function 2 and the rejection function 1, the rejection function 2, and the rejection function 3 are arranged in the next line. The synthesis functions f1 and f2 can be created, for example, by the method of creating g1 described above.

In the second embodiment, the third embodiment, and the fourth embodiment, when the parallel arithmetic device can be used, the rejection functions arranged in parallel may be calculated in parallel. Moreover, even if they are arranged in series, they may be calculated if the next rejection function can be calculated. In that case, when the calculation result of the next rejection function becomes unnecessary, the result may be discarded.

4). Effects of the embodiment

According to the present embodiment, a rejection method combining a plurality of rejection indexes can be automatically configured. Thereby, the human cost for combining a plurality of rejection indicators can be reduced. Further, according to the present embodiment, the misreading rate can be reduced while maintaining the correct reading rate at a high level, and a precise and high-speed rejection method can be configured.

Further, in this embodiment, a plurality of rejection indicators are configured in such a way that highly independent ones are configured in series, and those having low independence are configured in parallel on the basis of the independence between the rejection indicators. Rate, low misreading rate, and fast rejection method can be provided at low human cost.

5. Additional remarks The present invention is not limited to the above-described embodiments, and includes various modifications. For example, the above-described embodiments have been described in detail for easy understanding of the present invention, and are not necessarily limited to those having all the configurations described. Further, a part of the configuration of one embodiment can be replaced with the configuration of another embodiment, and the configuration of another embodiment can be added to the configuration of one embodiment. Further, it is possible to add, delete, and replace other configurations for a part of the configuration of each embodiment.
Each of the above-described configurations, functions, processing units, processing means, and the like may be realized by hardware by designing a part or all of them with, for example, an integrated circuit. Each of the above-described configurations, functions, and the like may be realized by software by interpreting and executing a program that realizes each function by the processor. Information such as programs, tables, and files for realizing each function can be stored in a memory, a hard disk, a recording device such as an SSD (Solid State Drive), or a recording medium such as an IC card, an SD card, or a DVD.
Further, the control lines and information lines indicate what is considered necessary for the explanation, and not all the control lines and information lines on the product are necessarily shown. Actually, it may be considered that almost all the components are connected to each other.

The character recognition method or character recognition apparatus / system of the present invention includes a character recognition program for causing a computer to execute each procedure, a computer-readable recording medium storing the character recognition program, and an internal memory of the computer including the character recognition program Can be provided by a program product that can be loaded on the computer, a computer such as a server including the program, and the like.

201 character recognition device 202 input device 203 display device 204 image acquisition device 205 communication device 206 arithmetic device (CPU)
207 External storage device (HDD, memory)

Claims (15)

A plurality of rejection value calculation units for calculating a rejection value by a preset rejection function for the recognition result of the character identified from the input image;
One or more rejection determination units for determining whether or not to reject the recognition result based on one or more rejection values calculated by any one or more of the plurality of rejection value calculation units, respectively,
With
Using the plurality of rejection value calculation units combined based on the correlation of the plurality of rejection value calculation units, the rejection determination unit makes a rejection determination of the recognition result based on a plurality of rejection values, and rejects By rejecting the determined recognition result, the recognition result that is not determined to be rejected is stored in a storage unit or displayed on a display unit.
The character recognition device according to claim 1,
Character recognition, characterized in that, when it is determined to be rejected based on the rejection value calculated by the previous rejection value calculation unit, the calculation of the rejection value by the subsequent rejection value calculation unit is skipped. apparatus.
The character recognition device according to claim 1,
A character recognition device characterized in that the rejection value calculation unit that calculates a rejection value having a high rejection capability or rejection rate is arranged in advance and performs calculation processing.
The character recognition device according to claim 1,
A character recognition device, characterized in that a calculation process is performed by arranging in series the rejection value calculation units that calculate a rejection value with high independence of a plurality of rejection values.
The character recognition device according to claim 1,
A character recognition device characterized in that a calculation process is performed by arranging in parallel the rejection value calculation units for calculating a rejection value having a low independence of a plurality of rejection values.
The character recognition device according to claim 1,
The rejection value calculation unit for calculating a rejection value with high independence of a plurality of rejection values is arranged in series, and the rejection value calculation unit for calculating a rejection value with low independence is arranged in parallel to perform calculation processing. A character recognition device characterized by being configured to perform.
The character recognition device according to claim 1,
The rejection function is a function for calculating a rejection value that takes a high value for the recognition result desired to be rejected and takes a low value for the recognition result not desired to be rejected. Character recognition device.
The character recognition device according to claim 1,
Rejected image database that collects image samples you want to reject in advance,
It has a correct image database that collects image samples you want to read correctly,
One or more threshold values for determining rejection are determined in comparison with a rejection value so that the rejection rate based on the correct image database is relatively small and the rejection rate based on the rejection image database is relatively large. Character recognition device.
The character recognition device according to claim 1,
The higher the rejection efficiency or the rejection value calculation cost based on the rejection value calculation cost, the higher the rejection efficiency calculation unit, the earlier the rejection value calculation unit, and based on the rejection value calculated by the rejection determination unit by the rejection determination unit If it is determined to be rejected, the character recognition device is characterized by omitting the process of calculating the reject value by the subsequent reject value calculating unit.
The character recognition device according to claim 1,
A character recognition device, wherein a new rejection function is determined based on a rejection value of each of the plurality of rejection value calculation units arranged in parallel, and a rejection determination is performed based on the new rejection function.
The character recognition device according to claim 1,
Rejected image database that collects image samples you want to reject in advance,
It has a correct image database that collects image samples you want to read correctly,
Judging the independence of multiple rejection values,
A function having the rejection value as an argument for discriminating between the image sample stored in the reject image database and the image sample stored in the correct image database is learned by a function based on an identification error, and the identification error by the function And the discrimination error when the rejection value is configured in series, and if the difference between the two is greater than or equal to a predetermined threshold, it is determined that the independence is low, otherwise A character recognition device characterized by determining that the independence is high.
The character recognition device according to claim 1,
A rejection value is calculated in parallel by the plurality of rejection value calculation units arranged in parallel, and / or a rejection value is calculated in parallel by the plurality of rejection value calculation units arranged in series. A character recognition device.
The character recognition device according to claim 1,
A document imaging unit for obtaining a document image by optically scanning the document;
A preprocessing unit that removes noise and background from the document image and binarizes to generate a binary image;
A layout analysis unit for analyzing the document structure and chart structure of the binary image;
A character string extraction unit that extracts an image of a character string unit from the binary image;
A character cutout unit that cuts out character-by-character images from each of the extracted character string images;
A character identification unit for recognizing a character in an image of each character unit cut out by the character cutout unit and outputting the recognition result;

A recognition result selection unit that selects the recognition result of each character string image based on the recognition result by the character identification unit and the rejection determination result by the rejection determination unit;
Based on the recognition result, a retry determination unit that determines whether to perform recognition reprocessing;
A post-recognition processing unit for storing and / or outputting the recognition result to a display device;
A character recognition device comprising:
A character recognition method,
For a recognition result of characters identified from the input image, using a plurality of rejection value calculation units for calculating a rejection value by a preset rejection function,
Based on one or more rejection values calculated by any one or more of the plurality of rejection value calculation units, respectively, using one or more rejection determination units to determine whether to reject the recognition result,
Using the plurality of rejection value calculation units combined based on the correlation of the plurality of rejection value calculation units, the rejection determination unit makes a rejection determination of the recognition result based on a plurality of rejection values, and rejects A character recognition method comprising: rejecting the determined recognition result, and storing the recognition result that is not determined to be rejected in a storage unit or displaying the recognition result on a display unit.
A character recognition program,
The processing unit uses a plurality of rejection value calculation units, a function of calculating a rejection value by a preset rejection function for the recognition result of the character identified from the input image,
The processing unit rejects the recognition result based on one or a plurality of rejection values calculated by any one or a plurality of the rejection value calculation units using one or a plurality of rejection determination units. A function to determine whether or not
The processing unit uses the plurality of rejection value calculation units combined based on the correlation of the plurality of rejection value calculation units, and the rejection determination unit performs rejection determination of the recognition result based on the plurality of rejection values. A character recognition program for causing a computer to execute a function of storing the recognition result that is not determined to be rejected in a storage unit or displaying it on a display unit by rejecting the recognition result determined to be rejected.

Images