JP2014130510A - Method and device for recognizing character - Google Patents

Abstract

PROBLEM TO BE SOLVED: To obtain the result of character recognition by segmenting individual characters from a handwritten string.SOLUTION: The result of character recognition is obtained by performing the steps of: obtaining a string image and dividing the string image into a plurality of basic segments on the basis of the outline thereof; converting the basic segments into character segments while generating character segments into which a plurality of basic segment are combined; calculating a character cost representing the degree of similarity between the character segment and the result of character recognition, and performing any one of a process of giving a connecting line cost to character segment having a possibility of being a connecting line in addition to the character cost and a process of giving a noise cost to a character segment having a possibility of being noise in addition to the character cost; searching for an optimal path for combining character segments on the basis of the character cost and one or both of connecting line cost and noise cost; and segmenting individual characters from the string on the basis of the character segment forming the searched optimal path.

Description

  The present invention relates to a character recognition method and a character recognition device that extract each character from a character string image obtained by imaging a handwritten character string and obtain a character recognition result.

  Conventionally, an apparatus for automatically recognizing characters entered in a form from an image obtained by capturing the form has been used. In recent years, there is an increasing number of forms in which a plurality of characters can be freely entered in one entry frame instead of a character frame in which characters are entered one by one in the frame. In such a form entry frame, it is often written continuously so that adjacent characters are connected to each other. When performing character recognition, a line connecting each character is not necessary, and it is desirable to cut out only each character. For this reason, in order to cut out each character from the handwritten character string written continuously, various techniques are utilized.

  For example, Patent Document 1 discloses a method of cutting out each character by extracting a continuous line connecting adjacent characters. Specifically, a pattern that is long in the horizontal direction is extracted from the pixels that form the character string, and it is determined whether or not this is a horizontal continuous line. If it is a horizontal continuous line, the contour of the character string is searched from the horizontal continuous line, and the position of the vertical separation line to be divided in the vertical direction is determined. After dividing each character by a vertical separation line, if the obtained character is zero, each character is cut out by erasing the continuous line.

  Patent Document 2 discloses a method of cutting out each character by obtaining a combination of candidate characters. Specifically, the character string is divided into a plurality of portions, and a plurality of candidate characters such as a candidate character including a character line connecting adjacent characters and a candidate character not including a character line are generated. Then, a candidate character lattice is generated from the combination of each candidate character, and an optimum route is searched on the candidate character lattice based on an evaluation value for evaluating the character recognition result of each candidate character. Then, each character is cut out based on the optimum route.

  Non-Patent Document 1 discloses a method of cutting out each character by dividing a character string into basic segments finer than one character and then obtaining a combination of basic segments forming each character. Specifically, a candidate lattice is generated by forming a graph structure of combinations of basic segments obtained by finely dividing a character string. Then, an optimum route for combining the basic segments is searched for while evaluating the character recognition result of each candidate character configured by combining the basic segments. Then, each character is cut out based on the optimum route.

Japanese Patent No. 3188580 JP 2009-199102 A

Yojiro Tonouchi, “High-precision online frameless character recognition technology”, Toshiba Review Vol. 66, no. 4, (2011), p. 56-57

  However, according to the above prior art, each character may not be cut out accurately depending on the state of the handwritten character. For example, in the method described in Patent Document 1, since a continuation line (hereinafter referred to as a connection line) connecting adjacent characters is faintly written, the connection line is not recognized as a connection line, and each character is In some cases, it could not be cut out accurately. Specifically, for example, if the connection line is written in a state where the connection line is faint and cut in the middle, the connection line may be recognized as a character and cut out as a character “-(hyphen)”.

  In addition, in the methods described in Patent Document 2 and Non-Patent Document 1, a DP matching method (a matching method based on dynamic programming) is used to search for an optimum route on the lattice, and each candidate character is searched. Based on the character recognition result, the optimum route is searched. For example, a recognition cost indicating a smaller value is used as the candidate character is more likely to be a character obtained as a character recognition result. Then, based on the recognition cost of each candidate character, the optimum route is searched by following the route connecting the candidate characters so that the recognition cost is minimized. In other words, in this method, the optimum route cannot be searched unless the recognition cost of each candidate character is determined. For this reason, it is possible to accurately determine whether it is a connection line or a hyphen at the time of character recognition because the connection line and the hyphen are very similar in a character string that may contain the character “-(hyphen)”. In some cases, the optimum route is searched in a state of being erroneously recognized, and an erroneous character recognition result is finally obtained.

  The present invention has been made to solve the above-described problems caused by the prior art. Even when a character string includes characters connected by a connection line, the connection line and the character are divided to accurately identify each character. An object of the present invention is to provide a character recognition method and a character recognition device that can be cut out to obtain a character recognition result.

  In order to solve the above-described problems and achieve the object, the present invention is a character recognition method for recognizing and recognizing each character from a handwritten character string, and an image for acquiring a character string image including the character string An acquisition step, an image dividing step of dividing the character string image into a plurality of basic segments based on a contour shape of the character string image, and character segments obtained by combining the basic segments as character segments and combining the basic segments. A character segment generation step to be generated, a character cost indicating the similarity between each generated character segment and a character obtained by character recognition of the character segment, and a character segment having a possibility of a connection line In addition to the character cost, a process that gives a connection line cost in addition to the character cost and a character segment that may cause noise Optimal combination of character segments based on a segment evaluation process that performs at least one of the process of assigning costs, a character cost obtained in the segment evaluation process, and at least one of a connection line cost and a noise cost An optimum route searching step for searching for a route; and a character cutting step for cutting out each character from the character string based on the character segment forming the optimum route searched in the optimum route searching step. .

  Further, in the above invention, the present invention provides the above-described invention, wherein, in the optimum route search step, when calculating an evaluation cost for evaluating whether or not the route is an optimum route, the character passes through a character segment that may be a connection line. Two evaluation costs are calculated from the cost and the connection line cost, and are used for a subsequent route search, and two evaluation costs are calculated from the character cost and the noise cost in a route passing through a character segment that may be noisy. And is used for the subsequent route search.

  Further, according to the present invention, in the above-described invention, in the optimum route searching step, a penalty is added to the character cost of the character segment according to a connection state between the character segment recognized as a character and an adjacent character segment. It is characterized by that.

  In the above-mentioned invention, the present invention is characterized in that, in the optimum route searching step, an evaluation cost for evaluating whether or not the route is an optimum route is calculated based on the number of characters recognized on the route. .

  Also, in the present invention according to the above-described invention, the image dividing step includes a contour extracting step of extracting a contour from the character string image, and a pixel on the extracted contour as a target pixel, and is separated from the target pixel by a predetermined distance. An angle calculating step of calculating an angle formed by a line segment connecting two points on the contour as an angle of the target pixel, and a cutting start specified as a cutting start point when the angle of the target pixel is equal to or less than a predetermined value The distance between the point specifying step, the cutting start point specified in the cutting start point specifying step, and the pixel on the contour is calculated, and the pixel showing the minimum value is specified as the cutting end point from the change in the distance A cutting end point specifying step, and a step of cutting and dividing the character string image by a cutting line connecting the cutting start point and the cutting end point.

  Further, the present invention is a character recognition device that cuts out and recognizes each character from a handwritten character string, and an input image acquisition unit that acquires a character string image including the character string; and an outline shape of the character string image An input image dividing unit that divides the character string image into a plurality of basic segments, and a segment generation unit that generates a character segment that is a combination of a plurality of the basic segments and the basic segment as a character segment, A character cost indicating the degree of similarity between each character segment and the character obtained by character recognition of the character segment is obtained, and a connection line cost is added to the character segment that may be a connection line in addition to the character cost. And at least a process of assigning a noise cost to a character segment with a possibility of noise in addition to the character cost. Optimal route search for searching for an optimum route for combining character segments based on a segment evaluation unit that performs one of the above, a character cost obtained by the segment evaluation unit, and at least one of a connection line cost and a noise cost And a character cutout unit that cuts out each character from the character string based on each character segment that forms the optimum route searched by the optimum route search unit.

  According to the present invention, when a character string is likely to be a connection line when a character string is divided into a plurality of character segments that may be characters and an optimum path for combining the character segments is searched for. Can search for the optimum route with the possibility of being both a character and a connection line. Similarly, when there is a possibility that the character segment is noise, the optimum route can be searched in a state where both the case of being a character and the case of being a noise remain. As a result, it is possible to accurately determine whether the character segment is a character, a connection line, or noise, and cut out each character from the character string to obtain an accurate character recognition result.

  In addition, according to the present invention, when searching for the optimum path that passes through each character segment, in the character segment that may be a connection line or noise, the character cost in the case of a character and the case of a connection line By using the connection line cost or the noise cost in the case of noise, it is possible to calculate an evaluation cost for evaluating whether or not the route is optimal. And, with this being held as an intermediate cost, it is possible to search for the optimum route in consideration of the character segment downstream on the route, so whether the character segment is a character, a connection line, or noise Can be accurately determined, and each character can be cut out from the character string to obtain an accurate character recognition result.

  In addition, according to the present invention, in order to avoid a partial area obtained by dividing one character from being erroneously recognized as a single character, a penalty is added to the character cost when it is connected to an adjacent character segment, thereby reducing the evaluation cost. By calculating, it is possible to accurately determine whether each character segment is one character or a partial region obtained by dividing one character.

  Further, according to the present invention, for example, when evaluating whether or not the route is an optimal route by using the average character cost of each character segment, the route search is performed in consideration of the number of characters included in the route. The route can be searched accurately.

  Further, according to the present invention, when the character string image is divided into basic segments, the cutting start point where the direction of the handwriting may have changed from the pixels forming the outline of the character string image, and the cutting start point Since the character string image is divided by the cutting line connecting the cutting start point and the cutting end point, the character string image is divided into partial areas that form characters or characters, and the optimum route is determined. Can be accurately searched.

FIG. 1 is a block diagram illustrating a schematic configuration of a character recognition device according to the present embodiment. FIG. 2 is a diagram for explaining a method of dividing a character string image into basic segments. FIG. 3 is a diagram illustrating a cutting start point and a cutting end point that constitute a cutting line that divides a character string image into basic segments. FIG. 4 is a diagram illustrating a method for collecting cutting lines for cutting character string images. FIG. 5 is a diagram illustrating an example in which a character string image is divided into basic segments. FIG. 6 is a diagram for explaining a method of generating a character segment from a character string image and setting a search path connecting the nodes. FIG. 7 is a diagram for explaining the cost of each character segment including a connection line. FIG. 8 is a diagram for explaining information stored in each node when searching for an optimum route. FIG. 9 is a diagram illustrating a method for searching for an optimum route based on the cost of each character segment. FIG. 10 is a diagram for explaining a method of back-tracing the route obtained as the optimum route. FIG. 11 is a diagram for explaining a method for searching for the optimum route from the average value of the cost of each character segment. FIG. 12 is a diagram for explaining a method for generating a character segment from a character string image to be processed in the second embodiment and setting a search path in consideration of the number of characters. FIG. 13 is a diagram illustrating a method for searching for an optimum route using a search route in consideration of the number of characters. FIG. 14 is a diagram showing a back trace of a route obtained as the optimum route. FIG. 15 is a diagram illustrating a character string image to be processed in the third embodiment. FIG. 16 is a diagram illustrating a result of searching for an optimum route for the image illustrated in FIG. 15 by the method according to the second embodiment. FIG. 17 is a diagram for explaining a penalty given to a character segment. FIG. 18 is a diagram illustrating a result of searching for an optimum route for the image illustrated in FIG. 15 in consideration of a penalty. FIG. 19 is a diagram for explaining that the optimum route search result varies depending on whether there is a penalty or local restriction. FIG. 20 is a diagram illustrating an example of another character string image to be processed in the third embodiment. FIG. 21 is a diagram showing search paths set in the character string image shown in FIG. FIG. 22 is a diagram showing the character recognition result and cost of each character segment generated from the character string image shown in FIG. FIG. 23 is a diagram for explaining the route search performed in the region D1 shown in FIG. FIG. 24 is a diagram for explaining the route search performed in the region D2 shown in FIG.

  Exemplary embodiments of a character recognition method and a character recognition apparatus according to the present invention will be explained below in detail with reference to the accompanying drawings. The character recognition method according to the present invention can be applied to various characters such as hiragana, katakana, kanji, numbers, alphabets, etc., but the following description will be made taking numbers as an example.

  FIG. 1 is a block diagram illustrating a schematic configuration of a character recognition device 1 that implements the character recognition method according to the present embodiment. The character recognition device 1 receives a character string image input from the outside, cuts out each character recognized from the character string image and outputs it as a character image, or outputs a result of character recognition of each character. It has a function to do.

  The character recognition device 1 includes an input image acquisition unit 11, an input image division unit 12, a segment generation unit 13, a segment evaluation unit 14, an optimum route search unit 15, and a recognition result output unit 16. . The input image acquiring unit 11 has a function of receiving a character string image input from the outside, and the input image dividing unit 12 is a function of dividing the character string image acquired by the input image acquiring unit 11 into basic segments of the minimum unit. Have The segment generation unit 13 has a function of generating a character segment by combining a plurality of basic segments while making the basic segment divided by the input image dividing unit 12 a character segment. The segment evaluation unit 14 has a function of performing character recognition of each character segment generated by the segment generation unit 13 and calculating a character cost for evaluating the similarity between each character segment and the character recognition result. The optimal route search unit 15 has a function of searching for an optimal route from search routes set as combinations of character segments. Each character segment on the optimum route searched by the optimum route search unit 15 becomes a character recognition result of each character cut out from the inputted character string image, that is, each character included in the character string. The recognition result output unit 16 functions as a character cutout unit, cuts out a character image including each character from the character string image based on the optimum route searched by the optimum route search unit 15, and outputs the character image to the outside. And a function of outputting a character recognition result of each extracted character image. The character recognition device 1 may be implemented by dedicated hardware and software programs, or may be implemented by executing dedicated software programs on a general-purpose computer device. I do not care.

  In the following, from the process of dividing the character string image acquired by the input image acquisition unit 11 into basic segments by the input image dividing unit 12, the optimum route searching unit 15 searches for the optimum route, and each character included in the character string is searched. Details of the processing until the result of character recognition is obtained will be described.

  Next, a process of dividing the input character string image into basic segments will be described with reference to FIGS. FIG. 2 is a diagram illustrating processing for determining a cutting start point that is a starting point of a cutting line that divides a character string image into basic segments.

  FIG. 2A shows a character string image acquired by the input image acquisition unit 11. In this character string image, a character string “200” consisting of three characters is written in a state of being connected by a connection line connecting adjacent characters. Specifically, the first character “2” and the second character “0” are connected by a connecting line between the second character “0” and the third character “0”. ing.

  As shown in FIG. 2B, the input image dividing unit 12 extracts only the pixels forming the contour from the character string image. Then, at each pixel position forming the contour, it is determined whether or not the pixel is a cutting start point.

  FIG. 2C is a diagram for explaining processing for determining whether or not the pixel A1 indicated by the arrow in FIG. 2B is a cutting start point. In FIG. 2C, each pixel is shown in a circular shape. As shown in FIG. 2C, two pixels on the contour located at a predetermined distance from the pixel A1 to be determined are defined as a pixel P1 and a pixel P2, respectively. An angle between a line segment connecting the pixel A1 and the pixel P1 and a line segment connecting the pixel A1 and the pixel P2 (an angle outside the contour shown in FIG. 2B) is θ. When the angle θ is equal to or smaller than the predetermined angle, the pixel A1 is determined to be a cutting start point. Thus, when the determination is made at each pixel position forming the outline of the character string and the pixel that becomes the cutting start point is extracted, for example, a plurality of cutting start points are extracted as shown by white circles in FIG. The

  The angle θ shown in FIG. 2C indicates the amount of change in the tangential direction of the pixels forming the contour. That is, as the angle θ is smaller, the contour is bent at an acute angle at the pixel position. For example, the direction of the handwriting changes while writing a character, the direction of the handwriting changes to write a connection line after writing the character, the direction of the handwriting changes to write a character after writing the connection line, etc. Thus, the angle formed by the contour becomes an acute angle. Since all connection lines are divided from the left and right characters, when the angle θ is equal to or smaller than the predetermined angle, all the pixels including the pixels included in one character are extracted as cutting start points regardless of the connection lines. Is.

  After extracting the cutting start point, the input image dividing unit 12 searches for a cutting end point corresponding to the cutting start point. The character string image is cut using the line segment connecting the cutting start point and the corresponding cutting end point as a cutting line, and each area after cutting becomes a basic segment. FIG. 3 is a diagram illustrating a method for searching for a cutting end point corresponding to the cutting start point A1. As indicated by arrows in FIG. 3, the distance from the cutting start point A1 is determined at each pixel position while sequentially tracing the pixels forming the contour from the cutting start point A1. Then, when the obtained distance from the cutting start point A1 is graphed, the pixel at the position where the distance takes the minimum value is selected as the cutting end point. In FIG. 3, the cutting start point is shown in a circular shape and the cutting end point is shown in a rectangular shape. Thus, after searching for a cutting end point corresponding to each cutting start point, each cutting candidate pair is evaluated with the cutting start point and cutting end point as a cutting candidate pair.

  For example, as shown in FIG. 3, when a cutting start point A1 and a cutting end point a11, which are cutting candidate pairs, are connected by a line segment, as shown by a broken line, outside the outline of the character string image (FIG. 2 ( In the case of passing through a region other than black pixels of the character string image shown in a), it is determined that this cutting candidate pair does not cut the character string image. Then, the pixel a11 is excluded from the cutting end point. On the other hand, in the cutting candidate pair of the cutting start point A1 and the cutting end point a12, the line segment connecting them passes only from the outline of the character string image to the inside of the character string image. It is determined that it is suitable for cutting the row image, and the pixel a12 is maintained as the cutting end point. Thus, after unnecessary cutting end points are excluded, cutting lines connecting the respective cutting start points (white circles) and the corresponding cutting end points (white rectangles) are set as indicated by solid lines in FIG.

  Even after removing unnecessary cutting end points, there may be multiple cutting lines in close proximity, or multiple cutting lines may exist in an intersecting state. To decide. FIG. 4 is a diagram illustrating a method for collecting cutting lines. FIG. 4B is an enlarged view of the rectangular area B shown in FIG. 4A, and FIG. 4C shows the result of the cutting lines shown in FIG. In FIG. 4, the cutting start points (A2 and A3) are indicated by white circles, and the cutting end points (a21, a22, a23, a31, a32 and a33) are indicated by white rectangles. In addition, cutting lines c3 to c8 connecting each cutting start point and each corresponding cutting end point are indicated by solid lines, and pixels located on the cutting lines are indicated by rectangles.

  Here, the proximity refers to a line segment formed by the cutting start point H and the cutting end point I, which is a cutting candidate pair, as a cutting line HI, and a line segment formed by another cutting candidate pair, the cutting start point J and the cutting end point K, as a cutting line JK. In this case, the length of the line segments HJ and IK is equal to or less than a predetermined value, or the length of the line segments HK and IJ is equal to or less than the predetermined value. When these lengths are less than or equal to a predetermined value set in advance, the longer one of the cutting lines HI and JK is excluded. Specifically, for example, as shown in FIG. 4B, the cutting line c3 connecting the cutting start point A3 and the cutting end point a33, which are cutting candidate pairs, and the cutting start point A3 and the cutting end point a32 are connected. When there is a cutting line c4, the cutting start point A3 is common (distance zero), and the distance between the cutting end points a32 and a22 is a predetermined value or less, the two cutting lines c3 and c4 are Determined to be close. Then, it is determined that the longer one of the two cutting lines c3 and c4 is excluded, and the pixel a33 is excluded from the cutting end point. Similarly, when there are intersecting cutting lines, the longer cutting line is also excluded.

  When cutting lines are aggregated, if there is a cutting line close to the horizontal, this cutting line is preferentially excluded. The character string is formed by characters arranged in the horizontal direction, and since the direction of cutting each character and the direction of cutting the character and the connecting line are both vertical, the horizontal cutting line is Exclude it. That is, cutting lines in the same direction as the arrangement direction of characters forming the character string are excluded. For example, in FIG. 4B, the cutting lines c7 and c8 are close to each other. In this case, it is determined that the horizontal cutting line c7 is excluded, and the pixel a21 is excluded from the cutting end point. Similarly, it is determined that the horizontal cutting line c5 is excluded from the intersecting cutting lines c5 and c6, and the pixel a31 is excluded from the cutting end point. As a result of consolidating the cutting lines in this way, the six cutting lines c3 to c8 shown in FIG. 4B are collected into three cutting lines c4, c6 and c8, and as shown in FIG. The character string image of B is divided into four areas B1 to B4.

  FIG. 5 is a diagram illustrating an example of a character string image divided into basic segments by a cutting line. In this way, when the cutting line for dividing the character string image is determined by the input image dividing unit 12, for example, the character string image shown in FIG. 2A is converted into 15 B0 to B14 as shown in FIG. Divided into regions. Each area divided at this time becomes a basic segment.

  In the following, a method for cutting out each character forming the character string from the character string and obtaining a character recognition result will be described in detail. FIG. 6 is a diagram for explaining a method of generating a character segment from a character string image and setting a search path connecting the nodes. 6 to 14, an example in which two characters “0 (zero)” are written by connecting with a connection line will be described for the sake of simplicity.

  6A shows a character string image acquired by the input image acquisition unit 11, and FIG. 6B shows a result of dividing the character string image into basic segments S1 to S3 by the input image dividing unit 12. ing. Here, S0 shown in FIG. 6B is a temporary segment indicating the starting point when searching for the optimum route. FIG. 6C is a diagram showing a search route set by the optimum route search unit 15 based on the character segment generated by the segment generation unit 13. The segment generation unit 13 sets each basic segment as a character segment and generates a character segment by combining a plurality of basic segments. That is, all combinations that can be taken in the basic segment are set as character segments. For example, in FIG. 6C, in addition to the character segments S1 to S3, a combination of S1 and S2, a combination of S2 and S3, and a combination of all S1 to S3 are set as character segments. . Similarly, for the search route, all routes that can be taken from combinations of character segments are set as the search route.

  N0 to n3 illustrated in FIG. 6C indicate nodes corresponding to the basic segments S0 to S3. There is a corresponding character segment on the path connecting the nodes n0 to n3, and the character segment combination is expressed by following the path to the right. For example, when it is determined that the optimal route is a route from the node n0 to the node n1, it is determined that the first character included in the character string image is the character segment S1 on the route. In the case where it is determined that the optimum route is a route from the node n0 to the node n2, the first character included in the character string image is a character segment combining the basic segments S1 and S2 on the route. It is determined that it is determined that (S1 + S2).

  Thus, a character segment is generated from the basic segment obtained by dividing the character string image, and a search path is set based on each character segment, while the segment evaluation unit 14 evaluates the character segment. Specifically, the character recognition process of each character segment is performed, a cost that is an evaluation value for evaluating the similarity between each character segment and the obtained character recognition result is calculated, and this is used as the character cost. Note that the character cost is not particularly limited as long as it is an evaluation value for evaluating the possibility that the character segment is a character obtained as a character recognition result, and there are various points such as points, distances, probabilities, etc. Numerical values can be used. In the following, the description will be continued assuming that the character cost is a numerical value indicating a smaller value as the character recognition result and the character segment are more similar (as the possibility of correct character recognition is higher).

  FIG. 7 is a diagram showing the relationship between the cost of each character segment obtained by the segment evaluation unit 14 and the search path shown in FIG. In this embodiment, the cost for obtaining the optimum route is defined by the sum of the costs of each character segment on the route, and the route with the smallest sum of costs among the possible routes on the searched route is the optimum route. It becomes.

  Here, as shown on the path from the node n1 to the node n2 in FIG. 7, the character segment S2 is treated as a character “-(hyphen)” having a cost value of 102 or a connection line having a cost value of 10. It is. Hereinafter, in order to distinguish from the character cost, the cost in the case of the connection line is described as the connection line cost.

  The segment evaluation unit 14 obtains a result that there is a possibility that the connection line has been erroneously recognized as a character recognition result of the character segment, such as the character “-(hyphen)” or the kanji “one”. In this case, a connection line cost as a connection line is given to this character segment in addition to the character cost based on the character recognition result. In this embodiment, the value of the connection line cost is 10. As described above, when the character segment is either a hyphen or a connection line, the processing is performed in a state where both possibilities remain without determining which one is the character recognition stage.

  In addition to the connection line, when there is a possibility that noise is erroneously recognized, this character segment is similarly given a noise cost as noise in addition to the character cost. For example, when the size of the character segment is equal to or smaller than a predetermined size, a noise cost as noise is given to the character segment in addition to the character cost. Note that the value assigned as the noise cost is changed depending on the size of the character segment regarded as noise. For example, in this embodiment, a value of 10 to 40 is assigned as the noise cost depending on the magnitude of noise.

  Although details will be described in a later embodiment, a cost may be added as a penalty depending on a character recognition result of a character segment and a connection state with an adjacent character segment. Since the numerical value of each cost varies depending on the type of character to be processed, the tendency to appear in handwritten characters, and the like, it is desirable to prepare a sample in advance and determine it experimentally. For example, a large number of samples are acquired from a form to be processed, and the cost value is experimentally determined so that a correct processing result can be obtained by the method described in this embodiment. Thereby, thereafter, using the determined numerical value, it is possible to accurately extract each character included in the character string and obtain a character recognition result.

  When the search route and cost for searching for the optimum route are determined, the optimum route search unit 15 starts searching for the optimum route. FIG. 8 shows information stored in each node when searching for the optimum route. The “hypothesis” of the item shown in FIG. 8 is based on the character recognition result of each character segment, information is obtained under each assumption in the case where the character recognition result is a character, a connection line, or a noise. Indicates to save. For “evaluation cost”, if the character segment is a character, it is in the character column, if it is a connection line, it is in the connection line column, if it is noise, it is in the noise column, The sum of the minimum costs until is saved. For the “previous node”, the information of the previous node for which the minimum cost is obtained is stored. Hereinafter, a method for searching for the optimum route while storing the information shown in FIG. 8 using the searched route shown in FIG. 7 will be specifically described.

  FIG. 9 is a diagram for explaining an optimum route searching method. In the node n0 in the initial state, there is no route from the left side to this node. Therefore, as shown in the node n0, the evaluation cost is 0 (zero) for all characters, connection lines, and noise, which corresponds to the previous node. No information is included.

  Since the node n1 has only the route from the node n0, information is stored based on the character recognition result of the character segment S1 on this route. Specifically, as shown in the node n1, since the character segment S1 is recognized as the character “0 (zero)”, the evaluation cost is stored in the character cost column. The stored evaluation cost value is obtained as the sum of the evaluation cost value 0 (zero) of the node n0 and the character cost value 130 of the character segment S1 to be “130”. Further, “n0” indicating the node n0 is stored as the information of the previous node.

  Next, in the node n2, there are a route from the node n1 and a route from the node n0. In the path from the node n1, there are two costs: a character cost when the character segment S2 is a character and a connection line cost when the character segment S2 is a connection line.

  Assuming that the character segment S2 is the character “-(hyphen)”, the evaluation cost of the character reaching the node n2 is the sum of the evaluation cost value 130 of the node n1 and the character cost value 102 of the character segment S2. 232 is obtained. In the node n2, since the character segment (S1 + S2) is recognized as the character “5” even in the route from the node n0, the character evaluation cost is also obtained here. The evaluation cost at this time is 170 as the sum of the evaluation cost value 0 (zero) of the node n0 and the character cost value 170 of the character segment (S1 + S2). Since the route indicating the minimum cost is searched for as the optimum route, when a plurality of evaluation costs relating to characters are obtained in this way, the smaller value is selected. Therefore, here, the evaluation cost value 232 from the node n1 is compared with the evaluation cost value 170 from the node n0, and the node n0 having a smaller value is selected. As a result, as shown in the node n2, “170” is stored as the character evaluation cost and “n0” is stored as the previous node information.

  The path from the node n1 has a connection line cost when it is assumed that the character segment S2 is a connection line. Therefore, as shown in the node n2, the sum “140” of the evaluation cost value 130 of the node n1 and the connection line cost value 10 is stored as the evaluation cost of the connection line. Then, “n1” is stored as the previous node information.

  As described above, for the character segment S2 having the costs of both the case of the character and the case of the connection line, the search for the optimum route is advanced using both costs. Specifically, at the node n2 downstream of the character segment S2 on the path, from the character evaluation cost obtained from the character cost when the character segment S2 is a character and the connection line cost when it is a connection line The route search is performed in a state where both the obtained connection line evaluation cost and the intermediate cost at the intermediate position of the route are maintained.

  In addition, although the example of the connection line was shown here, in this example, in addition to the character cost in the case of the character, the connection line cost in the case of the connection line or the noise cost in the case of the noise is given. In the character segment, the search for the optimum route proceeds using the character cost and the connection line cost or the noise cost.

  Next, in the node n3, there is a route from the nodes n0 to n2. In the route from the node n0, the evaluation cost value 0 (zero) of the node n0 and the character segment (S1 + S2 + S3) are obtained as the sum of the character cost value 250 recognized as the character “0 (zero)”, and the character The value of the evaluation cost is 250. Similarly, in the route from the node n1, the evaluation cost value 130 of the node n1 and the character cost value 156 when the character segment (S2 + S3) is recognized as the character “0 (zero)” are obtained. The value of the character evaluation cost is 286.

  In the route from the node n2, the character evaluation cost value 170 of the node n2 and the character segment S3 are obtained as the sum of the character cost value 108 recognized as the character “0 (zero)”, and the character The value of the evaluation cost is 278. Further, since the path from the node n2 includes the result when the character segment S2 is a connection line, the evaluation cost is also calculated for this. This evaluation cost value is 248, which is obtained as the sum of the evaluation cost value 140 of the connection line of the node n2 and the character cost value 108 of the character segment S3.

  Since the character segment S3 is recognized as the character “0 (zero)”, the result obtained when the character segment S2 that is the previous segment is used as a connection line is also used. If the character is recognized as other than zero), the result using the character segment S2 as a connection line is not used. That is, when a subsequent character segment is recognized as a character that will not be written after the connection line, the result of using the immediately preceding character segment as the connection line is not used.

  Thus, the evaluation cost of the character of the node n3 is the evaluation cost from the node n0, the evaluation cost from the node n1, the evaluation cost when the character segment S2 is a character at the node n2, and the character segment S2 as a connection line at the node n2. In this case, four of the evaluation costs are obtained, and among these, a route having a minimum evaluation cost value is selected. As a result, the path when the character segment S2 is the connection line at the node n2 is selected, and as shown in the node n3, the character evaluation cost “248” and the previous node information “n2” are stored.

  When the minimum cost to the node n3 which is the final node is determined, a process of back-tracing the searched route is performed. FIG. 10 is a diagram for explaining the contents of the backtrace performed by the optimum route search unit 15 after the optimum route search result is obtained as shown in FIG. When the evaluation cost to the final node is calculated, the specific character recognition result of each character segment that forms the optimal path is not specified, and the character recognition result of each character segment is specified by this backtrace process. The First, since the previous node of the minimum cost 248 of the node n3 is the node n2, the path returning to the node n2 is selected.

  The node n2 has two evaluation costs for characters and connection lines. Since the recognition result of the character segment S3 is the character “0 (zero)”, both are valid, and the connection line with the lower cost is selected, and the previous node is the node n1. At this time, for example, if the character segment S3 is a character other than 0 (zero), it is determined that the connecting line is not written before the character segment S3, the character is selected, and the previous node is the node n0. It is said. That is, in the character recognition method according to the present embodiment, the search for the optimum route is performed in a state where both costs are maintained without determining whether the character segment S2 is a character or a connection line, and the downstream side of the route The result of the character segment S2 is finally determined in consideration of the result of the character segment S3.

  Since the node before node n1 is only node n0, node n0 is selected unconditionally. In this manner, the optimal path and the character recognition result of each character segment are determined by back-tracing with the nodes n3, n2, n1, and n0. Specifically, the character recognition result of each character segment on the determined optimum route becomes each character forming the character string. In the result of FIG. 10, each character is determined to be the character “0 (zero)”, the connection line, and the character “0 (zero)”. Based on this result, the character string shown in FIG. 6A can be divided into two characters 0 (zero) and a connection line connecting them. Moreover, the character recognition result that the character string shown in FIG. 6A is the character “0 (zero)” and the character “0 (zero)” can be obtained except for the connection line.

  In this embodiment, the smaller the character cost value, the more similar to the character recognition result obtained for each character segment, and the optimum route is searched for by calculating the sum of the cost of the character segment in each route. Shown when possible. However, the present embodiment is not limited to this. For example, when the cost monotonously increases or decreases monotonically as the route is advanced, the route search can be performed by the same method based on the sum of the costs, as in this embodiment. Depending on the situation, it is possible to search for the optimum route while confirming the product of the cost of each character segment. Specifically, if the character cost of each character segment is expressed by a probability indicating that the obtained character recognition result is correct, the optimum path can be searched based on the product of the probabilities.

  In this embodiment, the information indicating the previous node is stored in each node. However, when back tracing is performed, the cost of each node of the back trace source and the back trace destination and the character segment on this path are stored. By using the cost, it is possible to search the previous node as the backtrace destination. Therefore, it is not essential to include information indicating the previous node in the storage item when searching for a route.

  In the present embodiment, a case will be described in which the evaluation cost for searching for the optimum route is defined by the average value of the cost of each character segment. As in the case of the first embodiment, the character cost of each character segment indicates that the smaller the value, the more similar the character recognition result and the character segment.

  Since the configuration of the character recognition device 1 and the method of dividing the character string image into basic segments are the same as those in the first embodiment, description thereof will be omitted. In the following, details of processing contents different from those in the first embodiment will be described. I decided to.

  If the evaluation cost when searching for the optimum route is simply defined as the average value of the cost of each character segment, the optimum route to the intermediate node in the middle is not the optimum route when viewed from the entire character string There is.

  FIG. 11 shows an example in which the optimum route is searched based on the average value of the character cost of each character segment. For example, as in the case of FIG. 6, the character string is divided into three basic segments S1 to S3, and a path between nodes n0 to n3 set corresponding to these basic segments is shown in FIG. ). The numerical value shown in FIG. 11A indicates the character cost of each character segment.

  According to the route and the character cost shown in FIG. 11A, the route that follows the nodes n0-n1-n2-n3 is the optimum route, and the minimum cost defined by the average value is shown in FIG. As shown as the evaluation cost, 120 (= (105 + 105 + 150) / 3) is calculated as the average value of the character costs of the character segments S1, S2 and S3. On the other hand, in the route in the case of following the node n0-n2-n3, the minimum cost is obtained as an average value of the character costs of the character segment (S1 + S2) and the character segment S3, and is calculated as 125 (= (100 + 150) / 2). It turns out that it is not the optimal route.

  However, when the evaluation cost up to the node n2 is calculated, 105 (= (105 + 105) / 2) is calculated as the average value of the character costs of the character segments S1 and S2 on the optimum path traced to the nodes n0-n1-n2. On the other hand, in the route traced to the nodes n0-n2, the character cost of the character segment (S1 + S2) is 100, and this route has a smaller evaluation cost value. Therefore, simply calculating the average value of the costs for each route and simply following the route for which the evaluation cost has the minimum value results in an error in the optimum route.

  For this reason, in the character recognition method according to the present embodiment, the optimum route is searched by calculating the average cost in consideration of the number of recognized characters. Hereinafter, the character recognition method according to the present embodiment will be specifically described.

  FIG. 12 is a diagram for explaining a method for generating a character segment from a character string image to be processed in the present embodiment and setting a search path in consideration of the number of characters. In order to simplify the description, the description will be continued by taking as an example a case where two characters “0 (zero)” are connected and connected.

  FIG. 12A shows a character string image acquired by the input image acquisition unit 11, and FIG. 12B shows a result of dividing the character string image into basic segments S1 to S3 by the input image dividing unit 12. ing. FIG. 12C shows a search route set by the optimum route search unit 15 in consideration of the number of characters based on the character segment generated by the segment generation unit 13. In this way, each node is arranged two-dimensionally, assuming that the node corresponding to the character segment Sx and the number of characters y is nxy.

  In FIG. 12C, the path that follows horizontally in the right direction is a path when it is assumed that it is not a character such as a connection line or noise. If the character segment is a connection line or noise, the number of characters does not increase, and therefore a horizontal path to the right is obtained. On the other hand, when the character segment is a character, the number of characters increases, and therefore, the route extends rightward and downward. For example, in the path from the node n11, when the character segment S2 is a connection line, the path reaches the node n21 horizontally in the right direction. On the other hand, when the character segment S2 is a character, the route reaches the node n22 rightward and downward.

  Next, a method for searching for the optimum route using the cost of each character segment will be specifically described. FIG. 13 is a diagram for explaining a method for searching for the optimum route on the searched route shown in FIG. In the initial state node n00, since there is no path from the left side to n00, all evaluation costs are 0 (zero), and information corresponding to the previous node is not included.

  The node n11 has only a route from n00. On this path, the character segment S1 is recognized as the character “0 (zero)”, and the value of the character cost is 130. Therefore, in the node n11, the character cost value “130” and the previous node information “n00” are stored in the character evaluation cost column.

  In the node n21, there are a route from the node n11 and a route from the node n00. In the path from the node n11, the character segment S2 is a connection line, and the value of the connection line cost is 10. Therefore, the sum “140” of the character cost value 130 and the connection line cost value 10 of the node n11 is stored as the evaluation cost of the connection line of the node n21, and “n11” is stored as the previous node information. In the route from the node n00, the character segment (S1 + S2) is recognized as the character “5”, and the character cost value is set to 170. Therefore, the sum “170” of the evaluation cost value 0 (zero) of the node n00 and the character cost value 170 of the character segment (S1 + S2) is stored as the evaluation cost of the character of the node n21. n00 "is saved. As a result, in the node n21, two pieces of information, that is, the evaluation cost and the previous node information of the path passing through the character segment S2 that is the connection line, and the evaluation cost and node information of the path that passes through the character segment (S1 + S2) that is the character. Will be saved.

  Similarly, for n22, n31, n32, and n33, the minimum cost and the information of the previous node when this minimum cost is obtained are stored. When the calculation of the minimum cost is completed for all the nodes, the average value of the minimum costs is then calculated at the final nodes n31, n32, and n33. As a result, the value of the evaluation cost for each path is 247 for the 1-character node n31, 124 for the 2-character node n32, and 147 for the 3-character node n33, and the path to n32 indicating the minimum value is the optimal path. It is determined that there is.

  When the optimum route is searched with the final node as the node n32, processing for back-tracing the optimum route is performed. When the evaluation cost to the final node is calculated, the specific character recognition result of each character segment that forms the optimal path is not specified, and the character recognition result of each character segment is specified by this backtrace process. The FIG. 14 is a diagram for explaining the contents of the backtrace performed by the optimum route search unit 15 after the optimum route search result is obtained as shown in FIG. First, since the previous node of the evaluation cost value 248 of the node n32 is the node n21, a route returning to the node n21 is selected.

  The node n21 has two evaluation costs for characters and connection lines. Since the recognition result of the character segment S3 is the character “0 (zero)”, both are effective results, and the connection line having the smaller cost value is selected, and the route going back to the node n11 is selected. That is, at the backtrace stage, it is determined that the character recognition result of the character segment S2 is a connection line. As in the first embodiment, for example, when the character segment S3 is a character other than 0 (zero), it is determined that no connection line is written before the character segment S3, and the character is selected. The previous node is assumed to be node n00.

  Since the node before node n11 is node n00, node n00 is selected. In this way, the backtracking is performed with the nodes n32, n21, n11, and n00, and the optimum path and the character recognition result of each character segment are determined. Specifically, the character recognition result of each character segment on the determined optimum route becomes each character forming the character string. In the result of FIG. 14, it is determined that each character is a character “0 (zero)”, a connection line, and a character “0 (zero)”. Thereby, the character string shown in FIG. 12A can be correctly divided into two characters “0 (zero)” and a connection line connecting between the two characters “0 (zero)”. A character image including each character can be accurately cut out from the character string image. Moreover, the character recognition result that the character string shown in FIG. 12A is the character “0 (zero)” and the character “0 (zero)” can be obtained except for the connection line.

  In the present embodiment, an example in which a penalty is used in addition to the method shown in the second embodiment will be described. Here again, as in the case of the first and second embodiments, the cost of each character segment indicates that the smaller the value, the more similar the character recognition result and the character segment.

  A penalty is a numerical value added to a cost when a character segment satisfies a predetermined condition. In a route including a contact state that is unlikely between character segments, a penalty is added to the evaluation cost to increase the value of the evaluation cost (lower the evaluation), thereby preventing an erroneous route search.

  In order to lower the evaluation based on the evaluation cost, a penalty is added to increase the cost when the better result is shown as the cost is lower, and the cost is reduced when the better result is shown as the cost is higher. Penalties are added. In Examples 1 and 2, the value 10 of the connection line cost when the character segment is a connection line is also a kind of penalty. However, in this example, a penalty is added even when the character segment is recognized as a character. This is different from Examples 1 and 2. A case where a penalty is added will be described later.

  In the present embodiment, the effect of local restrictions will also be described. As described in the first and second embodiments, the local restriction is a condition in which the number 0 (zero) is written after the connection line is written. Specifically, when a character segment is recognized as a character other than the character 0 (zero), processing is performed with a local restriction that the immediately preceding character segment is not a connection line. In other words, when a character segment is processed as a connection line, the immediately following character segment must be character 0 (zero). Hereinafter, a specific example will be described.

  First, an image to be processed in the present embodiment will be described. Normally, each character that forms the character string is cut out from the character string, but in this embodiment, the explanation is simplified and the contents and effects of the penalty and local restrictions are easily understood. The description will be continued by taking the case of recognizing one character as an example.

  FIG. 15 is a diagram illustrating the character “7” to be processed by the character recognition device 1 and the result of dividing the character by the input image dividing unit 12 in this embodiment. FIG. 15A shows the relationship between the input image and the basic segments S1 to S3, and FIG. 15B shows the basic segments S1 to S3.

  FIG. 16 is a diagram illustrating a result of searching for an optimum route by the method described in the second embodiment on the searched route by setting a searched route based on the character segment generated by the segment generating unit 13. The method for obtaining the minimum cost and storing the information of the previous node in each node n00 to n33 is the same as that in the second embodiment, and thus the description thereof is omitted. However, in the search result shown in FIG. 16, the cost of the node n33 is the minimum. The optimum route is n00-n11-n22-n33. As a result, from the character “7” shown in FIG. 15A, three characters of the character “1”, the character “− (hyphen)”, and the character “1” are cut out. As described above, when the character string to be processed includes characters for which the optimum route is erroneously searched in the method according to the second embodiment, the optimum route is correctly searched by the method shown in the present embodiment. be able to.

  In this embodiment, an optimum route is searched using a penalty. FIG. 17 is a diagram illustrating types of penalties and values added to the character cost as penalties. For example, no. 1 is that if the character segment is recognized as the character “-(hyphen)” but the left side is in contact with the immediately preceding character segment on the left side, the connecting line is erroneously recognized as a character. Therefore, a value of 50 is added to the character cost as a penalty. No. Similarly, 2 indicates that the connection line may be misrecognized as a character, so that a value of 50 is added to the character cost as a penalty. No. 3 and no. 4 is a partial area obtained by dividing a character when the character segment is recognized as the character “1”, but the upper or lower portion of the character is in contact with another character segment. Indicates that the value 50 is added to the character cost as a penalty.

  Penalty No. 5 indicates that when the character segment is noise, the noise cost value is set to 50 in this embodiment. No. 6 indicates that the connection line cost value is set to 10 when the character segment is a connection line, as shown in the first and second embodiments. No. 5 and no. Reference numeral 6 denotes a case where noise and connection lines are used instead of characters. When the character segment is a noise or a connection line, the number of characters in the character recognition result does not increase as in the character segment S2 shown between the nodes n11 and n21 in FIG. Follow the path horizontally in the direction.

  FIG. 18 shows a result of correcting the optimum route search result shown in FIG. 16 using the penalty shown in FIG. That is, it is a result of performing an optimum route search based on the penalty and local constraint conditions according to the present embodiment.

  In the path from the node n11 to the node n22, the character segment S2 is a character “-(hyphen)”. As shown in FIG. 15A, the character segment S2 is in contact with the character segment S1 on the left side. 1 applies. Similarly, since the right side is also in contact with the character segment S3, the penalty No. 2 applies. Therefore, as shown in the lower part of the character segment S2 on the path from the node n11 to the node n22 in FIG. 18, P = 100 (= 50 × 2) is added to the character cost as a penalty. Therefore, the evaluation cost of the character at the node n22 is calculated as the sum of the evaluation cost value 164 of the node n11, the character cost value 118 of the character segment S2, and the penalty cost value 100, and becomes 382.

  Similarly, in the path from the node n22 to the node n33, the character segment S3 is the character “1”. As shown in FIG. 15A, the character segment S3 is the upper character segment. Since it is in contact with S2, penalty no. 3 is applied and P = 50 is added to the cost. Therefore, the character cost at the node n33 is calculated as the sum of the evaluation cost value 382 of the node n22, the character cost value 100 of the character segment S3, and the penalty value 50, which is 532. Since there are three characters on this route, the evaluation cost at the node n33 is finally calculated as 177 (= 532/3). Similarly, a penalty is applied to the path from the node n11 to the node n32 and the path from the node n21 to the node n32, and a predetermined value is added to the character cost.

  16 and 17 show an example in which local restrictions are applied to the path from the node n21 to the node n32. On the path from the node n21 to the node n32, the character segment S3 is recognized as the character “1”. That is, since it is not recognized as the character “0 (zero)”, the local restriction that it should not be a connection line is applied to the immediately preceding node n21. In FIG. 16 and FIG. 17, a path cut off due to local restrictions is indicated by a cross. Due to this local restriction, in the path from the node n21 to the node n32, the result of using the character segment S2 of the immediately preceding node n21 as the connection line (result of the cost 124 and the previous node being n11) cannot be used.

  As described above, as a result of searching for the optimum route including the penalty and the local constraint, the cost of the node n31 shows the minimum value as shown in FIG. Then, the back trace is performed from the node n31, and it is determined that the optimum route is n00-n31. Then, from the contents of the character segment (S1 + S2 + S3) existing on this route, the character shown in FIG. 15A can be correctly recognized as one character “7”.

  FIG. 19 is a diagram for explaining a difference in search results that appear depending on the presence or absence of local restrictions and penalties in the route search shown in FIG. FIG. 19A shows the results when there are no local constraints and penalties, and FIG. 19B shows the results when there are local constraints and penalties. As shown in FIG. 19A, if there are no local restrictions and penalties, no. The route shown in FIG. 1 is set as the optimum route, and an erroneous character recognition result is obtained. Also, here, applying local constraints, No. Even if the route 3 is removed, the wrong No. One optimal path is maintained. On the other hand, as shown in FIG. 19B, when both the penalty and the local constraint are applied, the No. 3 is removed, and the optimum route is No. 3. Five routes are selected. As described above, by applying both the constraint condition and the penalty, it is possible to accurately search the optimum route. And the correct character recognition result can be obtained.

  15 to 18, the case of a single character has been described as an example for the sake of simplicity. Needless to say, the character recognition method according to the present embodiment can accurately recognize each character using a character string as a processing target. Yes. Hereinafter, an example of a result of searching for an optimum route from a character string including a plurality of characters by the character recognition method according to the present embodiment will be described.

  20-24 is a figure which shows another example to which the character recognition method shown in the present Example is applied. FIG. 20 is a diagram for explaining a character string to be processed by the character recognition device 1. 20A shows a character string image acquired by the input image acquisition unit 11, and FIGS. 20B and 20C show that the character string image is divided into basic segments S1 to S6 by the input image dividing unit 12. FIG. Results are shown.

  FIG. 21 shows a search route set by the optimum route search unit 15 based on each character segment generated by the segment generation unit 13. As described above, when divided into six basic segments, a large number of nodes and paths are set.

  FIG. 22 shows the character cost of each character segment obtained by the segment evaluation unit 14. As described above, the segment evaluation unit 14 calculates the character cost for each of the 21 character segments generated from the six basic segments by the segment generation unit 13.

  Then, as described above in the present embodiment, the optimum route searching unit 15 performs a process for searching for the optimum route. FIG. 23 and FIG. 24 are diagrams showing a part of the evaluation cost obtained when searching for the optimum route on the route shown in FIG. 21 and the result of storing the previous node information. FIG. 23 shows the result of the region D1 in FIG. 21, and FIG. 24 shows a part of the result included in the region D2 of FIG. When searching for the optimum route on the search route shown in FIG. 21, as described above, in a character segment that may be a connection line or noise, in addition to the character cost, the connection line cost or noise cost. The evaluation cost is calculated using Since the character cost of each character segment is as shown in FIG. 22, in FIG. 23 and FIG. 24, the display of the character cost is omitted in the character segment that may be a connection line or noise. Specifically, when the character segment is a character or a connection line (segments S3 and S5), only the connection line cost is shown, and even when the character segment is a character or noise (segment S2), only the noise cost is shown. Show.

  For example, as shown in FIG. 20B, since the character segment S2 is smaller than the other character segments, the segment evaluation unit 14 adds noise to the character segment S2 in addition to the character cost shown in FIG. Noise cost. As a result, as shown in the path from the node n11 to the node n21 in FIG. 23, the character segment S2 is regarded as noise with the noise cost 50, and the noise evaluation cost of the node n21 is calculated.

  As described above, when the evaluation cost is obtained for all the routes shown in FIG. 21 using the character cost of each character segment shown in FIG. 22, the node n00-n31-n42-n52-n63 is reached as shown in FIG. A route is determined as the optimum route. The evaluation cost at this time is 124 (= 371/3) from the value shown in FIG. Based on the result obtained by back-tracing the optimum route, the character “2” of the character segment (S1 + S2 + S3) on the route from the character string to the node n31 from the character string, and the route from the node n31 to the node n42 The character “0 (zero)” of the upper character segment S4, the connecting line of the character segment S5 on the path from the node n42 to the node n52, and the character “0 (zero) of the character segment S6 on the path from the node n52 to the node n63 ) ”Will be cut out. That is, the character “2”, the character “0 (zero)”, the connection line, and the character “0 (zero)” are cut out from the character string shown in FIG. Then, the character recognition result that the character string shown in FIG. 20A is the character “2”, the character “0 (zero)”, and the character “0 (zero)” can be obtained except for the connection line. it can.

  As described above, according to the present embodiment, even when characters connected by connecting lines are included in a character string, a character segment is generated from a basic segment obtained by dividing the character string, and a search for connecting the character segments is performed. A route is set, and an optimum route is searched based on the character cost of each character segment. At this time, for a character segment that may be a character or a character other than a character (connection line or noise), it is possible to proceed with the process of searching for an optimum route with both possibilities remaining. Then, it is possible to determine whether the character segment is a character or a character other than the character while using the information of the character segment on the downstream side in the search direction by performing a back-trace process on the obtained route. As a result, it is possible to accurately search the optimum route and obtain a character recognition result obtained by accurately cutting out each character from the character string.

  In addition, when setting a route by character segment, the cost of each search route is calculated in consideration of the number of characters by setting the search route in two dimensions: the direction in which the number of characters increases and the direction in which the number of characters does not increase. The route can be searched accurately.

  Further, when searching for the optimum route, the optimum route is accurately searched by adding a penalty according to the character recognition result of the character segment and the contact state between the character segment adjacent to the left and right of the character segment. be able to.

  Also, when searching for the optimum route, for example, by setting a local constraint based on the content that the character “0 (zero)” comes after the connection line, the impossible route is excluded and the optimum route is excluded. The load related to the route search can be reduced and the optimum route can be searched accurately.

  As described above, the present invention is a useful technique for accurately recognizing characters by cutting out each character from the character string when the handwritten character string includes characters written by connecting with a connection line. .

DESCRIPTION OF SYMBOLS 1 Character recognition apparatus 11 Input image acquisition part 12 Input image division | segmentation part 13 Segment generation part 14 Segment evaluation part 15 Optimal route search part 16 Recognition result output part

Claims (6)

A character recognition method for recognizing and extracting each character from a handwritten character string,
An image acquisition step of acquiring a character string image including the character string;
An image dividing step of dividing the character string image into a plurality of basic segments based on a contour shape of the character string image;
A character segment generating step for generating a character segment in which the basic segment is a character segment and a plurality of the basic segments are combined;
A character cost indicating the degree of similarity between each generated character segment and a character obtained by character recognition of the character segment is obtained, and a connection line cost is added to the character segment that may be a connection line in addition to the character cost. A segment evaluation step of performing at least one of the process of giving noise cost in addition to the character cost to the character segment having the possibility of noise and the character segment,
An optimum route search step of searching for an optimum route for combining each character segment based on the character cost obtained in the segment evaluation step and at least one of a connection line cost and a noise cost;
And a character extracting step of extracting each character from the character string based on the character segment forming the optimal route searched in the optimal route searching step. In calculating the evaluation cost for evaluating whether or not the optimum route in the optimum route searching step,
In the route through the character segment with the possibility of a connection line, two evaluation costs are calculated from the character cost and the connection line cost, and used for the subsequent route search,
2. The character recognition method according to claim 1, wherein two evaluation costs are calculated from the character cost and the noise cost in a route passing through a character segment having a possibility of noise and used for a subsequent route search.   3. The optimum route searching step adds a penalty to a character cost of the character segment according to a connection state between the character segment recognized as a character and an adjacent character segment. The character recognition method described in 1.   The said optimal path | route search process calculates the evaluation cost which evaluates whether it is an optimal path | route based on the number of the characters recognized on the path | route, The Claim 1, 2, or 3 characterized by the above-mentioned. Character recognition method. The image dividing step includes
A contour extraction step of extracting a contour from the character string image;
An angle calculation step of calculating an angle formed by a line segment connecting two points on the contour separated from the target pixel by a predetermined distance as a pixel on the extracted contour as the target pixel;
A cutting start point specifying step for specifying a cutting start point when the angle of the target pixel is equal to or smaller than a predetermined value;
A cutting end point specifying step of calculating a distance between the cutting start point specified in the cutting start point specifying step and a pixel on the contour, and specifying a pixel indicating a minimum value from the change in the distance as a cutting end point; ,
The character recognition method according to claim 1, further comprising a step of cutting and dividing the character string image by a cutting line connecting the cutting start point and the cutting end point. . A character recognition device for recognizing and recognizing each character from a handwritten character string,
An input image acquisition unit for acquiring a character string image including the character string;
An input image dividing unit that divides the character string image into a plurality of basic segments based on a contour shape of the character string image;
A segment generation unit that generates a character segment in which the basic segment is a character segment and a plurality of the basic segments are combined;
A character cost indicating the degree of similarity between each generated character segment and a character obtained by character recognition of the character segment is obtained, and a connection line is added to the character segment that may be a connection line in addition to the character cost. A segment evaluation unit that performs at least one of a process of giving a cost and a process of giving a noise cost in addition to the character cost to a character segment that has a possibility of noise;
An optimum route search unit that searches for an optimum route that combines each character segment based on the character cost obtained by the segment evaluation unit and at least one of a connection line cost and a noise cost;
A character recognition device comprising: a character cutout unit that cuts out each character from the character string based on each character segment that forms the optimum route searched by the optimum route search unit.

Images