JP5636766B2 - Image processing apparatus and image processing program - Google Patents

Description

  The present invention relates to an image processing apparatus and an image processing program.

There is a technique for cutting out a character image from an image.
As a technology related to this, for example, Patent Document 1 describes that character recognition is accurate for the purpose of performing accurate recognition even if the size, shape, pitch, and the like of characters are different for each document. The deterministic character detection unit detects the character that appears to be a definite character and the recognition result output unit outputs it as the recognition result, while the character rectangle information detection unit detects the rectangle information of the definite character and optimizes the rectangle evaluation function The rectangular division integration unit obtains a new basic rectangle by dividing and integrating the basic rectangles of parts that have not yet been determined based on the optimized rectangle evaluation function, and again performs processing below the candidate character selection unit It is also disclosed that the writer estimation unit optimizes the dictionary used in the candidate character selection unit using information obtained from the confirmed characters.

  Further, for example, in Patent Document 2, it is possible to cut out characters from a document image at high speed and accurately, and cut out characters in a document in which alphanumeric symbols such as kanji and hiragana are mixed at high speed and accurately. For this purpose, a cutout candidate is estimated from the shape information of each circumscribed rectangle, character recognition is performed on the estimated cutout candidate, and a cutout candidate that is determined to be cutout can be determined based on the character recognition result. For a cutout candidate that is determined as a cutout result and that is not determined to be cutout by the character recognition result, a plurality of cutout candidates based on a combination of each circumscribed rectangle are estimated, and each cutout candidate is individually A recognition evaluation value for each rectangle is obtained, and combinations for each segmentation candidate using the recognition evaluation values for these individual rectangles The cutout candidate that obtained the optimal combination evaluation value among the values is confirmed as the cutout result, and only the alphanumeric symbols are recognized after performing recognition for only the alphanumeric symbols, and then the alphanumeric symbols And cutting out characters other than the confirmed characters.

  In addition, for example, Patent Document 3 has a problem of recognizing a character string with high accuracy even when the character extraction or character string collation cannot determine how to extract the character. In the cutout process, first, a plurality of cutout methods that are judged to be more correct are selected from all cutout methods by the conventional method, and then the size of each character pattern and the pattern before and after the method of the present invention are selected. The evaluation value (rough shape penalty) for evaluating the validity of the hypothesis of how to cut out is calculated based on the positional relationship, and this assumption is made by the linear discriminant function obtained by collecting and registering in advance and learning the sample. It is disclosed that information on size and positional relationship can be easily handled by discriminating a correct clipping method.

JP 05-174187 A Japanese Patent Laid-Open No. 08-161432 Japanese Patent Laid-Open No. 09-185681

  In the present invention, when a position to cut out a character image existing in an image is determined, and the evaluation value of a candidate for a cut-out position becomes a singular value, the position to be cut out is determined only by the singular evaluation value. It is an object of the present invention to provide an image processing apparatus and an image processing program that prevent this from happening.

The gist of the present invention for achieving the object lies in the inventions of the following items.
According to the first aspect of the present invention, there is provided first calculation means for calculating a weighted linear sum for a plurality of feature amounts related to position candidates for cutting out one character image existing in an image, and the first calculation means. As an argument, the absolute value of the slope of the output is obtained when the argument converges to a predetermined value when the argument is an extreme value, or the distance between the argument and the predetermined value increases. A second calculation means for calculating an evaluation value of a candidate for a position to cut out the one character image by a nonlinear monotone function that becomes smaller or a function that approximates the nonlinear monotone function; and a calculation by the second calculation means based on the evaluation value, a clipping position determining means for determining a position for cutting out a character image that is present in the image, and receiving means for receiving training data segmentation position of the character image, the cropping position A number calculation unit that calculates the number of correct answers or errors of the cutout position by comparing the cutout position determined by the determination unit with the teacher data received by the reception unit; and the cutout position calculated by the number calculation unit Based on the number of correct answers or the number of errors, weight change means for changing the weight used in the first calculation means at the character cut-out position for one character is provided, and the weight change means is a correct answer when the current weight is used. The image processing apparatus is characterized in that the next weight is determined from the amount of change from the value based on the number or the number of errors to the number of correct answers with the changed weight or the value based on the number of errors .

According to a second aspect of the present invention , there is provided a first calculation means for calculating a weighted linear sum for a plurality of feature quantities related to a position candidate for cutting out one character image existing in an image, and the first calculation means. As an argument, the absolute value of the slope of the output is obtained when the argument converges to a predetermined value when the argument is an extreme value, or the distance between the argument and the predetermined value increases. A second calculation means for calculating an evaluation value of a candidate for a position to cut out the one character image by a nonlinear monotone function that becomes smaller or a function that approximates the nonlinear monotone function; and a calculation by the second calculation means A plurality of sets of cut position determining means for determining a position to cut out a character image existing in the image based on the evaluated value, and a set of the first calculation means and the second calculation means, Multiple second A third calculation means for calculating a weighted linear sum with respect to the evaluation value calculated by the calculation means, and a calculation result by the third calculation means as an argument, and when the argument is a limit value, A nonlinear monotonic function that approximates to the nonlinear monotonic function, which converges to a predetermined value, or whose absolute value of the slope of output decreases as the distance between the argument and a predetermined value increases. And a fourth calculating means for calculating an evaluation value of a position candidate to cut out the one character image, wherein the cutting position determining means is based on the evaluation value calculated by the fourth calculating means. the determines the position for cutting out a character image that is present in the image, and receiving means for receiving training data segmentation position of the character image, the cropping position determining extraction position before and determined by means Based on the number of correct answers or the number of errors of the cutout position calculated by the number calculating means by comparing the teacher data received by the receiving means and calculating the number of correct answers or the number of errors of the cutout position, 1 A weight changing means for changing a weight used in the first calculating means or the third calculating means at a character cut-out position for a character, wherein the weight changing means The image processing apparatus is characterized in that the next weight is determined from the amount of change from the value based on the number of errors to the number of correct answers with the weight after change or the value based on the number of errors .

According to a third aspect of the present invention , there is provided a first calculation means for calculating a weighted linear sum for a plurality of feature quantities related to a position candidate for cutting out one character image existing in an image, and the first calculation means. As an argument, the absolute value of the slope of the output is obtained when the argument converges to a predetermined value when the argument is an extreme value, or the distance between the argument and the predetermined value increases. A second calculation means for calculating an evaluation value of a candidate for a position to cut out the one character image by a nonlinear monotone function that becomes smaller or a function that approximates the nonlinear monotone function; and a calculation by the second calculation means A plurality of sets of cut position determining means for determining a position to cut out a character image existing in the image based on the evaluated value, and a set of the first calculation means and the second calculation means, Multiple second A fifth calculating means for calculating the sum of the evaluation values calculated by the calculating means, and the cutting position determining means is based on the sum of the evaluation values calculated by the fifth calculating means. A receiving unit that determines a position to cut out a character image existing in the image and receives teacher data of a cutout position of the character image, a cutout position determined by the cutout position determining unit, and a teacher data received by the receiving unit The number calculation means for calculating the correct number or error number of the cutout position and the correct number or error number of the cutout position calculated by the number calculation means, Weight changing means for changing the weight used in the first calculating means is provided, and the weight changing means is the number of correct answers or errors in the case of the current weight. An image processing apparatus characterized by the amount of change to the correct number or value based on the number of errors in the weight of the changed from a value based on the number determines the next weight.

According to a fourth aspect of the present invention , there is provided a first calculation means for calculating a weighted linear sum for a plurality of feature amounts related to a position candidate for cutting out one character image existing in an image, and the first calculation means. As an argument, the absolute value of the slope of the output is obtained when the argument converges to a predetermined value when the argument is an extreme value, or the distance between the argument and the predetermined value increases. A second calculation means for calculating an evaluation value of a candidate for a position to cut out the one character image by a nonlinear monotone function that becomes smaller or a function that approximates the nonlinear monotone function; and a calculation by the second calculation means A plurality of sets of cut position determining means for determining a position to cut out a character image existing in the image based on the evaluated value, and a set of the first calculation means and the second calculation means, Multiple second A third calculation means for calculating a weighted linear sum with respect to the evaluation value calculated by the calculation means, and a calculation result by the third calculation means as an argument, and when the argument is a limit value, A nonlinear monotonic function that approximates to the nonlinear monotonic function, which converges to a predetermined value, or whose absolute value of the slope of output decreases as the distance between the argument and a predetermined value increases. And a fourth calculating means for calculating an evaluation value of a position candidate to cut out the one character image, wherein the cutting position determining means is based on the evaluation value calculated by the fourth calculating means. , Determining a position to cut out a character image existing in the image, a plurality of sets by the first calculation means and the second calculation means, a set by the third calculation means and the fourth calculation means Have multiple Sixth calculating means for calculating a sum of evaluation values calculated by the plurality of fourth calculating means is provided, and the cut-out position determining means is a sum of evaluation values calculated by the sixth calculating means. The character image existing in the image is determined to be cut out, and accepting means for receiving the character image cut-out position teacher data, the cut-out position determined by the cut-out position determining means and the accepting means Based on the number of correct answers or the number of errors of the cutout position by comparing the received teacher data, and based on the number of correct answers or the number of errors of the cutout position calculated by the number calculation means A weight changing means for changing a weight used in the first calculating means or the third calculating means at the character cutout position; Image processing characterized in that the next weight is determined from the amount of change from a value based on the number of correct answers or the number of errors in the case of the current weight to a value based on the number of correct answers or the number of errors in the changed weight Device.

According to a fifth aspect of the present invention, there is provided a first calculating means for calculating a weighted linear sum with respect to a plurality of feature amounts relating to a position candidate for cutting out one character image existing in an image; When the calculation result of the calculation means is used as an argument, when the argument is a limit value, the value converges to a predetermined value, or when the distance between the argument and the predetermined value increases, the slope of the output Second calculation means for calculating an evaluation value of a candidate for a position to cut out the one character image by a non-linear monotone function whose absolute value becomes small or a function approximating the non-linear monotone function; and the second calculation based on the evaluation value calculated by the means, and the cropping position determining means for determining a position for cutting out a character image that is present in the image, receiving means for receiving training data segmentation position of the character image The number calculation means for comparing the cutout position determined by the cutout position determination means and the teacher data received by the reception means to calculate the correct number or the number of errors of the cutout position, and calculated by the number calculation means Based on the number of correct answers or the number of errors in the cutout position, the weight changer functions to change the weight used in the first calculation unit at the character cutout position for one character. An image processing program for determining a next weight from an amount of change from a value based on the number of correct answers or the number of errors in the case of a value to a value based on the number of correct answers or the number of errors in the changed weight .
According to a sixth aspect of the present invention, there is provided a first calculation means for calculating a weighted linear sum with respect to a plurality of feature amounts related to a position candidate for cutting out one character image existing in an image; When the calculation result of the calculation means is used as an argument, when the argument is a limit value, the value converges to a predetermined value, or when the distance between the argument and the predetermined value increases, the slope of the output Second calculation means for calculating an evaluation value of a candidate for a position to cut out the one character image by a non-linear monotone function whose absolute value becomes small or a function approximating the non-linear monotone function; and the second calculation Based on the evaluation value calculated by the means, there are a plurality of sets of cutting position determining means for determining a position for cutting out a character image existing in the image, and a set of the first calculating means and the second calculating means. A third calculation means for calculating a weighted linear sum with respect to the evaluation values calculated by the plurality of second calculation means; and a calculation result obtained by the third calculation means as an argument. A non-linear monotonic function or the non-linear function in which the absolute value of the output slope decreases as the distance between the argument and the predetermined value increases. A function that approximates a monotonic function functions as a fourth calculation unit that calculates an evaluation value of a position candidate to cut out one character image, and the cut-out position determination unit is calculated by the fourth calculation unit. Based on the evaluation value, the position for cutting out the character image existing in the image is determined, and the reception means for receiving the teacher data of the cutting position of the character image is determined by the cutting position determination means. A number calculating means for calculating the number of correct answers or errors of the cutout position by comparing the cutout position and the teacher data received by the receiving means, and the number of correct answers or errors of the cutout position calculated by the number calculating means Based on the number of characters, the weight calculation unit functions as a weight change unit that changes the weight used in the first calculation unit or the third calculation unit at the character cut-out position for one character. In this case, the image processing program is characterized in that the next weight is determined from the amount of change from the value based on the number of correct answers or the number of errors to the value based on the number of correct answers or the number of errors with the changed weight.
According to a seventh aspect of the present invention, there is provided a first calculating means for calculating a weighted linear sum with respect to a plurality of feature amounts related to a position candidate for cutting out one character image existing in an image; When the calculation result of the calculation means is used as an argument, when the argument is a limit value, the value converges to a predetermined value, or when the distance between the argument and the predetermined value increases, the slope of the output Second calculation means for calculating an evaluation value of a candidate for a position to cut out the one character image by a non-linear monotone function whose absolute value becomes small or a function approximating the non-linear monotone function; and the second calculation Based on the evaluation value calculated by the means, there are a plurality of sets of cutting position determining means for determining a position for cutting out a character image existing in the image, and a set of the first calculating means and the second calculating means. , Functioning as fifth calculation means for calculating the sum of the evaluation values calculated by the plurality of second calculation means, and the cut-out position determination means is configured to calculate the evaluation values calculated by the fifth calculation means. Based on the sum, a position for cutting out a character image existing in the image is determined, and accepting means for receiving teacher data of the cutout position of the character image; a cutout position determined by the cutout position determining means; and the receiving means The number calculation means for calculating the number of correct answers or errors at the cutout position by comparing the teacher data received by the above, and the number of correct answers or the number of errors at the cutout position calculated by the number calculation means. Functioning as a weight changing means for changing the weight used in the first calculating means at the character cutout position of the character, and the weight changing means An image processing program for determining a next weight from an amount of change from a value based on the number of correct answers or the number of errors in the case of a value to a value based on the number of correct answers or the number of errors in the changed weight .
The invention according to claim 8 is a first calculation means for calculating a weighted linear sum with respect to a plurality of feature quantities related to a position candidate for cutting out one character image existing in the image; When the calculation result of the calculation means is used as an argument, when the argument is a limit value, the value converges to a predetermined value, or when the distance between the argument and the predetermined value increases, the slope of the output Second calculation means for calculating an evaluation value of a candidate for a position to cut out the one character image by a non-linear monotone function whose absolute value becomes small or a function approximating the non-linear monotone function; and the second calculation Based on the evaluation value calculated by the means, there are a plurality of sets of cutting position determining means for determining a position for cutting out a character image existing in the image, and a set of the first calculating means and the second calculating means. A third calculation means for calculating a weighted linear sum with respect to the evaluation values calculated by the plurality of second calculation means; and a calculation result obtained by the third calculation means as an argument. A non-linear monotonic function or the non-linear function in which the absolute value of the output slope decreases as the distance between the argument and the predetermined value increases. A function that approximates a monotonic function functions as a fourth calculation unit that calculates an evaluation value of a position candidate to cut out one character image, and the cut-out position determination unit is calculated by the fourth calculation unit. The character image existing in the image is cut out based on the evaluation value, a plurality of sets of the first calculation unit and the second calculation unit, the third calculation unit, and the second calculation unit; Calculation of 4 A plurality of sets of stages, and function as sixth calculation means for calculating a sum of evaluation values calculated by the plurality of fourth calculation means, wherein the cut-out position determination means is the sixth calculation means Based on the sum of the evaluation values calculated by step (b), a position for cutting out a character image existing in the image is determined, and a receiving unit that receives teacher data for the cutting position of the character image is determined by the cutting position determination unit. The number calculation means for calculating the number of correct answers or the number of errors of the cutout position by comparing the cutout position and the teacher data received by the receiving means, and the number of correct answers or the number of errors of the cutout position calculated by the number calculation means And a weight changing means for changing a weight used in the first calculating means or the third calculating means at the character cutout position for one character, The weight changing means is configured to change from a change amount from a value based on the number of correct answers or the number of errors in the case of the current weight to a value based on the number of correct answers or the number of errors in the changed weight. An image processing program characterized by determining a weight.

According to the image processing apparatus of claim 1, when a position to cut out a character image existing in an image is determined and the evaluation value of a candidate for the cut-out position becomes a unique value, the unique evaluation is performed. It is possible to prevent the position to be cut out from being determined only by the value. In addition, when determining the position to cut out the character image existing in the image, it is possible to determine the weight to be used for calculating the evaluation value.

According to the image processing apparatus of the second aspect, it is possible to accurately determine the position where the character image existing in the image is cut out as compared with the case where the present configuration is not provided. In addition, when determining the position to cut out the character image existing in the image, it is possible to determine the weight to be used for calculating the evaluation value.

According to the image processing apparatus of the third aspect, it is possible to accurately determine the position where the character image existing in the image is cut out as compared with the case where the present configuration is not provided. In addition, when determining the position to cut out the character image existing in the image, it is possible to determine the weight to be used for calculating the evaluation value.

According to the image processing apparatus of the fourth aspect, it is possible to accurately determine the position where the character image existing in the image is cut out as compared with the case where the present configuration is not provided. In addition, when determining the position to cut out the character image existing in the image, it is possible to determine the weight to be used for calculating the evaluation value.

According to the image processing program of claim 5 , when a position to cut out a character image existing in an image is determined and the evaluation value of the candidate for the cut out position becomes a unique value, the unique evaluation is performed. It is possible to prevent the position to be cut out from being determined only by the value. In addition, when determining the position to cut out the character image existing in the image, it is possible to determine the weight to be used for calculating the evaluation value.
According to the image processing program of the sixth aspect, the position to cut out the character image existing in the image can be determined with higher accuracy than in the case where the present configuration is not provided. In addition, when determining the position to cut out the character image existing in the image, it is possible to determine the weight to be used for calculating the evaluation value.
According to the image processing program of the seventh aspect, it is possible to accurately determine the position to cut out the character image existing in the image as compared with the case where the present configuration is not provided. In addition, when determining the position to cut out the character image existing in the image, it is possible to determine the weight to be used for calculating the evaluation value.
According to the image processing program of the eighth aspect, it is possible to accurately determine the position where the character image existing in the image is cut out as compared with the case where the present configuration is not provided. In addition, when determining the position to cut out the character image existing in the image, it is possible to determine the weight to be used for calculating the evaluation value.

It is a conceptual module block diagram about the structural example of 1st Embodiment. It is a conceptual module block diagram about the structural example in the arc evaluation value determination module of 1st Embodiment. It is explanatory drawing which shows the example of a data structure of the data table for teachers. It is a conceptual module block diagram about the structural example in the arc evaluation value determination module of 2nd Embodiment. It is a notional module block diagram about the structural example in the arc evaluation value determination module of 3rd Embodiment. It is a conceptual module block diagram about the structural example in the arc evaluation value calculation module of 3rd Embodiment. It is explanatory drawing which shows the example of a data structure of the data table for teachers. It is explanatory drawing which shows the example of a relationship between an arc candidate determination module, an arc evaluation value determination module, and a character extraction position determination module. It is a notional module block diagram about the structural example of 4th Embodiment. It is a block diagram which shows the hardware structural example of the computer which implement | achieves the 1st-4th embodiment. It is explanatory drawing which shows the example of a character string image. It is explanatory drawing which shows the example of a character boundary candidate. It is explanatory drawing which shows the example of a circumscribed rectangle. It is explanatory drawing which shows the example of a character cutout result. It is explanatory drawing which shows the example of the graph expression which shows a character cutout position. It is explanatory drawing which shows the example of the pattern 2 in a graph expression. It is a conceptual module block diagram about the structural example of the image processing apparatus which performs general character cutout and character recognition. It is explanatory drawing which shows the example which added the arc evaluation value to the graph expression which shows a character cutout position. It is explanatory drawing which shows the example when it can isolate | separate with a straight line in feature-value space. It is explanatory drawing which shows the example when it cannot isolate | separate with a straight line in feature-value space.

The present embodiment is for cutting out a character image existing in an image when processing for a character image, such as character recognition, is performed.
First, before describing the present embodiment, the premise or an image processing apparatus using the present embodiment will be described. This description is intended to facilitate understanding of the present embodiment.

  For example, a character string image as shown in the example of FIG. 11 is targeted. First, this character string image is divided into character segments. A character segment is a character itself or a character portion that can be a part of a character. Here, a horizontally written character string image as shown in the example of FIG. 11 is taken as an example. In a horizontally written image, the image is divided into character segments by dividing the image with vertical lines (or lines close to vertical). For example, the character string image can be divided by the vertical lines (cut candidate 1210, cut candidate 1220) shown in FIG. 12 to obtain three character segments “I”, “HI”, and “Study”. The vertical line shown in the example of FIG. 12 is called a break candidate. The break candidate 1210 separates “I” and “HI”, and the break candidate 1220 separates “HI” and “Study”.

Next, as shown in the example of FIG. 13, circumscribed rectangles (circumscribed rectangle 1310, circumscribed rectangle 1320, circumscribed rectangle 1330) are extracted for each character segment.
Hereinafter, the technical content described in Patent Document 3 will be described as an example. Note that the terms used in the following description may be different from the terms used in Patent Document 3.
A character image is determined by integrating the character segments described above. A plurality of character segments may be integrated to form one character image, or one character segment may become one character. Determining a character image is equivalent to determining a character cutout position, and hence may be referred to as determining a character cutout position below.
There are multiple patterns of character segment integration. A final character cutout position is determined by selecting a character image having the highest evaluation from among a plurality of existing patterns.
For the example of FIG. 13, all the character cutout patterns are as shown in the example of FIG. That is, in the example of FIG. 14A, three character images (circumscribed rectangles 1310, 1320, 1330) are used as the pattern 1, and in the example of FIG. 14B, two character images (circumscribed rectangles 1310 and 1320 are used as the pattern 2. 1330) and FIG. 14C, one character image (circumscribed rectangles 1310, 1320, and 1330) is used as the pattern 3, and two character images (circumscribed rectangle 1310 are used as the pattern 4 in the example shown in FIG. 14D. , Circumscribed rectangles 1320 and 1330).

  A plurality of cutout patterns shown in the example of FIG. 14 can be represented as a graph expression indicating a character cutout position. In the example of FIG. 15, the graph is composed of four nodes, a start node 1500, an end node 1590, an intermediate node 1510 (node 1), and an intermediate node 1520 (node 2), and arcs connecting the nodes ( Connection lines between nodes are called arcs). The start point corresponds to the left end of the character string image, and the end point corresponds to the right end of the character string image. Intermediate node 1510 (node 1) and intermediate node 1520 (node 2) respectively indicate character break candidate positions (that is, break candidate 1210 and break candidate 1220 shown in the example of FIG. 12). The intermediate node 1510 (node 1) corresponds to the break candidate 1210. Further, the intermediate node 1520 (node 2) corresponds to the break candidate 1220.

A route from the start point through each node to the end point is hereinafter referred to as a “path”. The path is composed of one or a plurality of arcs. Usually there are multiple paths. The character cutout pattern shown in the example of FIG. 14 corresponds to these multiple paths. For example, the pattern 2 shown in the example of FIG. 14B corresponds to the paths (character cutout pattern 1504 and character cutout pattern 1522) shown by the thick lines in FIG.
Here, one character image candidate corresponds to any one arc. For example, a character image (character cutout pattern 1504) “K” corresponds to an arc connecting the start point node 1500 and the intermediate node 1520 (node 2). For a character corresponding to one arc, an evaluation value of the character can be determined. This is called an “arc evaluation value”.
The arc evaluation value is calculated from character shape information, recognition accuracy in character recognition, and the like. Details of this will be described later.

Here, a conceptual module configuration diagram of a configuration example of an image processing apparatus that performs general character segmentation and character recognition will be described with reference to FIG.
The image processing apparatus includes an image reception module 110, a character string extraction module 120, a character boundary candidate extraction module 130, an arc feature amount extraction module 140, a linear weighting addition module 1710, a character segmentation module 160, and a character recognition module 170. Yes.

  The image reception module 110 is connected to the character string extraction module 120, receives a target image, and passes the image to the character string extraction module 120. Accepting an image means, for example, reading an image with a scanner, a camera, etc., receiving an image from an external device via a communication line by fax, etc. And the like, and the like read out the images stored in the device etc.). The image may be a binary image or a multi-value image (including a color image). One image may be received or a plurality of images may be received. Further, as long as characters are included in the content of the image, it may be a document used for business, a pamphlet for advertisement, or the like.

  The character string extraction module 120 is connected to the image reception module 110 and the character boundary candidate extraction module 130, receives an image from the image reception module 110, extracts a character string image from the image, and converts the character string image into a character boundary. Pass to candidate extraction module 130. For the extraction of the character string image, a conventionally known technique may be used. For example, a histogram of the number of black pixels existing in the horizontal direction or the vertical direction is created, the histogram has a predetermined width, and a character string that has a predetermined distance from the adjacent histogram is separated from the adjacent histogram Extracted as an image.

  The character boundary candidate extraction module 130 is connected to the character string extraction module 120 and the arc feature quantity extraction module 140, receives a character string image from the character string extraction module 120, extracts boundary candidates of the character string image, and extracts the boundary Candidates are passed to the arc feature extraction module 140. For example, the cut candidate 1210 and the cut candidate 1220 shown in the example of FIG.

  The arc feature amount extraction module 140 is connected to the character boundary candidate extraction module 130 and the linear weight addition module 1710, receives the boundary candidate from the character boundary candidate extraction module 130, extracts the feature amount of the boundary candidate, and The feature quantity is passed to the linear weighting addition module 1710 as a feature quantity vector. The feature amount of each arc in the graph expression indicating the character cutout position is extracted. The feature amount of the arc will be described later. Generally, there are a plurality of arc feature amounts, but one may be used. Hereinafter, the arc feature quantity is also referred to as a feature quantity vector.

  The linear weighting addition module 1710 is connected to the arc feature quantity extraction module 140 and the character segmentation module 160, receives a feature quantity vector from the arc feature quantity extraction module 140, calculates an arc evaluation value of the feature quantity vector, The arc evaluation value is passed to the character cutout module 160. The calculation of the arc evaluation value will be described later.

The character segmentation module 160 is connected to the linear weighting addition module 1710 and the character recognition module 170, receives an arc evaluation value from the linear weighting addition module 1710, and selects a break candidate based on the arc evaluation value, that is, a character string image. The position where the character image existing inside is cut out is determined, the character image is cut out from the character string image (or the image received by the image receiving module 110) along the cut candidate, and the character image is passed to the character recognition module 170. “Based on the arc evaluation value” means, for example, selecting a value having the highest arc evaluation value (a value indicating that there is a high possibility of cutting out one character).
The character recognition module 170 is connected to the character cutout module 160, receives a character image from the character cutout module 160, recognizes the character image, and outputs a character code as a recognition result.

Processing of the arc feature quantity extraction module 140 and the linear weighting addition module 1710 will be described.
One path is composed of a plurality of arcs. Using a plurality of arc evaluation values, an evaluation value of a path constituted by the arcs can be calculated. This is called a “path evaluation value”.
The path evaluation value corresponds to, for example, a weighted sum of arc evaluation values. In the technique described in Patent Document 3, weighting is performed by the number of character segments in the arc.
In order to determine the character cutout position, the path with the highest path evaluation value is selected from the plurality of paths. If the path can be selected, the character cutout position is confirmed, and the character recognition result is also confirmed.
In the example of FIG. 16, it is assumed that a thick line path is selected. In this case, the character cutout positions are three points: a start node 1500, an intermediate node 1520 (node 2), and an end node 1590. In addition, the character recognition results are “formation” and “study”.

In particular, an arc evaluation value calculation method performed by the linear weighting addition module 1710 will be described.
In the technique described in Patent Document 3, a character evaluation value is calculated using a weighted linear sum of character shape information and character recognition accuracy information. More specifically, in the technique described in Patent Document 3, the arc evaluation value is calculated as follows.
First, a circumscribed rectangle of characters corresponding to each arc is created. This corresponds to the circumscribed rectangle of each character in each pattern shown in the example of FIG. Hereinafter, a circumscribed rectangle of a character corresponding to each arc is referred to as an arc circumscribed rectangle.
Next, arc feature values (plural) are calculated as follows.
f _1: the corresponding arc of the circumscribed rectangle of height f _2: the width f of the circumscribed rectangle of the corresponding arc _3: and a circumscribed rectangle of the corresponding arc of the circumscribed rectangle of the left arc distance f _4: the circumscribed rectangle of the corresponding arc, right arc enclosing rectangles spacing f _5: maximum interval f between the circumscribed rectangle of a character segment in the corresponding arc _6: number connected components in the corresponding arc further, the character similarity, and f _7.
In the technique described in Patent Document 3, the arc evaluation value V is determined by Expression (1). In this case, however, N = 7 in equation (1).

That is, the arc evaluation value is determined by the weighted linear sum of the character shape information (f _{1 to} f ₆ ) and the character recognition accuracy information f ₇ . w _i is a weight at the time of calculating the linear sum. c is a constant. In the description method of Patent Document 3, the expression (1) is described in a different format instead of the expression (1). However, although there are differences in description, they are mathematically the same.
The linear weighted addition module 1710 receives the above-described values of f _{1 to} f ₇ that are feature amounts as the feature amount vectors. Here, the number of features is N. The internal operation of the linear weighted addition module 1710 is shown by the equation (1). Then, the arc evaluation value V is passed to the character cutout module 160.

Next, a phenomenon that may occur when the technique described in Patent Document 3 is implemented will be described.
<Phenomenon 1>
In the technique described in Patent Document 3, a linear sum of feature amounts is used as an arc evaluation value. Since it is a linear sum, the range of the arc evaluation value can be a value between minus infinity and plus infinity depending on the content of the feature amount.
When the arc evaluation value is calculated using the linear sum in this way, the arc evaluation value may become a very high value or a very low value.
When the arc evaluation value becomes a very high value or a very low value, the overall path evaluation value may be dragged to the very high or very low value. For example, here, the path evaluation value is evaluated as a weighted sum of arc evaluation values. The weight is determined appropriately. Here, it is assumed that weighting is performed by the number of character segments in the arc as in the prior art.
In the example of FIG. 15, it is assumed that the evaluation values are as shown in the example of FIG.
If it is such an arc evaluation value, for example, if weighting is performed by the number of character segments in the arc,
The path evaluation value in the case of the character cutout pattern 1504 “K” and the character cutout pattern 1522 “Study” is 10 × 2 + 10 = 30.
The path evaluation value in the case of the character cutout pattern 1506 “I”, the character cutout pattern 1512 “H”, and the character cutout pattern 1522 “Study” is 1 + 100 + 10 = 111.
That is, since the arc evaluation value of “HI” is too high compared to other evaluation values, the other evaluation values are small (that is, when the character evaluation is not character-like, for example, the arc evaluation value of the character cutout pattern 1506 “I”) 1), however, the path including “hi” is selected by being dragged by the value.

次に、正解切り出し位置に対応するアーク特徴量ベクトルをｆ_Ｔ、不正解切り出し位置に対応するアーク特徴量ベクトルをｆ_Ｆとする。
アーク評価値関数が妥当であるためには、式（３）のような関係になることが望ましい。つまり、正解切り出し位置の場合のアーク評価値は、不正解切り出し位置の場合のアーク評価値よりも大きな値であることが望ましい。

正解アーク、不正解アークは複数存在するため、式（３）のような関係を得るためには、正解切り出し位置の場合のアーク評価値の最小値Ｖ_Ｔｍｉｎと、不正解切り出し位置の場合のアーク評価値の最大値Ｖ_Ｆｍａｘとの関係が、式（４）のようになる必要がある。

式（４）が成り立つとき、式（５）を満たす値Ｖ_０が存在する。

<Phenomenon 2>
Here, it is assumed that the arc feature vector f = (f ₁ ,..., F _N ) and the weight vector is w = (w ₁ ,..., W _N ). The arc evaluation value function is V (f). Formula (1) becomes like Formula (2).

Next, an arc feature quantity vector corresponding to the correct answer cut-out position is set to f _T , and an arc feature quantity vector corresponding to the incorrect answer cut-out position is set to f _F.
In order for the arc evaluation value function to be valid, it is desirable that the relationship is as shown in Expression (3). That is, it is desirable that the arc evaluation value in the case of the correct answer cut-out position is larger than the arc evaluation value in the case of the incorrect answer cut-out position.

Since there are a plurality of correct answer arcs and incorrect answer arcs, in order to obtain the relationship as in Expression (3), the minimum arc evaluation value V _{Tmin in} the case of the correct answer cut-out position and the arc in the case of the incorrect answer cut-out position The relationship between the evaluation value and the maximum value V _Fmax needs to be as shown in Equation (4).

When equation (4) holds, there is a value V ₀ that satisfies equation (5).

（６）式で表すことのできる超平面は（５）式より、正解切り出し位置の場合の特徴量ベクトル（正解特徴量ベクトル）の分布と、不正解切り出し位置の場合の特徴量ベクトル（不正解特徴量ベクトル）の分布を完全に分離することになる。このように、特徴量空間内の超平面で、正解特徴量ベクトルの分布と不正解特徴量ベクトルの分布を分離することができれば、（３）式を満たすことができて、妥当なアーク評価値関数を設計することが可能となる。
簡単のため、特徴量空間が２次元の場合を図示する。２次元の場合には、超平面は直線となる。図１９の例に示す破線１９３０のように、正解特徴量分布１９２０と不正解特徴量分布１９１０が直線で分離できる場合は問題がない。 Here, the feature vector exists in an N-dimensional space. Expression (1) or Expression (2) forms a hyperplane in this N-dimensional feature amount space. That is, a set of feature quantity vectors f such that the arc evaluation value V (f) becomes the predetermined value V ₀ exists on the hyperplane represented by the equation (6).

The hyperplane that can be expressed by the equation (6) is the distribution of the feature vector (correct feature vector) at the correct cutout position and the feature vector (incorrect solution) at the incorrect cutout position from the equation (5). The distribution of the feature vector is completely separated. Thus, if the correct feature vector distribution and the incorrect feature vector distribution can be separated on the hyperplane in the feature space, equation (3) can be satisfied, and an appropriate arc evaluation value can be obtained. It is possible to design a function.
For simplicity, the case where the feature space is two-dimensional is illustrated. In the case of two dimensions, the hyperplane is a straight line. There is no problem if the correct feature quantity distribution 1920 and the incorrect answer feature quantity distribution 1910 can be separated by a straight line as indicated by a broken line 1930 in the example of FIG.

以上で、本実施の形態の前提又は本実施の形態を利用する画像処理装置についての説明を終了する。 In the technique described in Patent Document 3, since the formula (1) is adopted, the separation surface cannot be other than a hyperplane. However, in reality, the separation plane between the correct answer and the incorrect answer is not a hyperplane, and may have a more complicated shape. The technique described in Patent Document 3 cannot cope with the case where the distribution of correct and incorrect answers has a complicated shape that cannot be separated on a hyperplane.
For example, in the case of the distribution of the correct feature amount distribution 2020 and the incorrect feature amount distribution 2010 as shown in the example of FIG. 20, it is no longer possible to separate them with straight lines. In such a case, the technique described in Patent Document 3 cannot obtain a proper arc evaluation value. That is, a phenomenon as expressed by the equation (7) occurs. When this phenomenon occurs, a high arc evaluation value is obtained despite the wrong character cutout position. As a result, the character cutout position is incorrect.

This is the end of the description of the premise of the present embodiment or the image processing apparatus using the present embodiment.

Hereinafter, examples of various preferred embodiments for realizing the present invention will be described with reference to the drawings.
<First Embodiment>
FIG. 1 is a conceptual module configuration diagram of a configuration example according to the first embodiment.
The module generally refers to components such as software (computer program) and hardware that can be logically separated. Therefore, the module in the present embodiment indicates not only a module in a computer program but also a module in a hardware configuration. Therefore, the present embodiment is a computer program for causing these modules to function (a program for causing a computer to execute each procedure, a program for causing a computer to function as each means, and a function for each computer. This also serves as an explanation of the program and system and method for realizing the above. However, for the sake of explanation, the words “store”, “store”, and equivalents thereof are used. However, when the embodiment is a computer program, these words are stored in a storage device or stored in memory. It is the control to be stored in the device. Modules may correspond to functions one-to-one, but in mounting, one module may be configured by one program, or a plurality of modules may be configured by one program, and conversely, one module May be composed of a plurality of programs. The plurality of modules may be executed by one computer, or one module may be executed by a plurality of computers in a distributed or parallel environment. Note that one module may include other modules. Hereinafter, “connection” is used not only for physical connection but also for logical connection (data exchange, instruction, reference relationship between data, etc.). “Predetermined” means that the process is determined before the target process, and not only before the process according to this embodiment starts but also after the process according to this embodiment starts. In addition, if it is before the target processing, it is used in accordance with the situation / state at that time or with the intention to be decided according to the situation / state up to that point.
In addition, the system or device is configured by connecting a plurality of computers, hardware, devices, and the like by communication means such as a network (including one-to-one correspondence communication connection), etc., and one computer, hardware, device. The case where it implement | achieves by etc. is also included. “Apparatus” and “system” are used as synonymous terms. Of course, the “system” does not include a social “mechanism” (social system) that is an artificial arrangement.
In addition, when performing a plurality of processes in each module or in each module, the target information is read from the storage device for each process, and the processing result is written to the storage device after performing the processing. is there. Therefore, description of reading from the storage device before processing and writing to the storage device after processing may be omitted. Here, the storage device may include a hard disk, a RAM (Random Access Memory), an external storage medium, a storage device via a communication line, a register in a CPU (Central Processing Unit), and the like.

  As shown in the example of FIG. 1, the image processing apparatus according to the present embodiment includes an image reception module 110, a character string extraction module 120, a character boundary candidate extraction module 130, an arc feature quantity extraction module 140, and an arc evaluation value determination module. 150, a character cutout module 160, and a character recognition module 170. Note that parts of the same type as those of the image processing apparatus shown in the example of FIG. Therefore, the arc evaluation value determination module 150 will be described in detail. However, the character string extraction module 120, the character boundary candidate extraction module 130, and the arc feature amount extraction module 140 will be described in more detail.

The character string extraction module 120 is connected to the image reception module 110 and the character boundary candidate extraction module 130.
The character string extraction module 120 extracts one column of character string images that are horizontally or vertically written from the target image. Here, the column is a column aligned horizontally in the case of horizontal writing, and a column aligned vertically in the case of vertical writing.
Some images have a plurality of character strings. Various methods for separating such a plurality of character strings into a single character string have been proposed so far, and these may be used.
As an example of separation to form a single character string, techniques described in JP-A-4-311283, JP-A-3-233789, JP-A-5-73718, JP-A-2000-90194, etc. There is. These methods and other methods may be used.

The character boundary candidate extraction module 130 is connected to the character string extraction module 120 and the arc feature amount extraction module 140.
The character boundary candidate extraction module 130 receives a single character string image and divides it into a plurality of character segments. Since there are various methods for dividing the character segment, any one of them may be used. For example, the techniques described in JP-A-5-114047, JP-A-4-100189, JP-A-4-92992, JP-A-4-68481, JP-A-9-54814, etc., Patent Document 3 In particular, a character boundary candidate extraction method described in paragraph 0021, a character cutout position determination method described in Japanese Patent Laid-Open No. 5-128308, particularly in paragraph 0005, and the like may be used. Of course, other methods are also acceptable.

The arc feature amount extraction module 140 is connected to the character boundary candidate extraction module 130 and the arc evaluation value determination module 150.
The content of the arc feature quantity vector extracted by the arc feature quantity extraction module 140 is not particularly limited. For example, the above-described feature quantities f _{1 to} f ₇ may be used. Other feature amounts may be used. The number of dimensions of the feature vector (that is, the number of types of feature quantities) is various and may be any number of dimensions.

Further, a supplementary description will be made regarding a specific example in which character recognition is performed after a character is cut out, and the character recognition accuracy is used as one of the feature quantities of the arc. This is the amount indicated as the character similarity in the technique described in Patent Document 3.
As the character recognition accuracy, it is only necessary to obtain a certainty or likelihood of the character code output at the time of character recognition. As methods for obtaining such character recognition accuracy, various methods have been proposed so far, and any one of them may be used. For example, a method described in paragraph 0024 of Patent Document 3 and a recognition evaluation value acquisition method described in Paragraph 0051 of Patent Document 2 may be used. Other methods may be used.

The arc evaluation value determination module 150 is connected to the arc feature quantity extraction module 140 and the character segmentation module 160, receives a feature quantity vector from the arc feature quantity extraction module 140, and determines an arc evaluation value using the feature quantity vector. The arc evaluation value is passed to the character segmentation module 160.
The arc evaluation value determination module 150 prevents the influence of a very large arc evaluation value when compared with other arc evaluation values and a very small arc evaluation value when compared with other arc evaluation values from increasing. It is. That is, when the arc evaluation value is large compared to other arc evaluation values, the influence of the magnitude is reduced, and when the arc evaluation value is small, the influence of the magnitude is reduced. As a technique, a nonlinear function having the following characteristics is further added to the weighted addition result of the feature amount. Nonlinear functions include: (1) a monotonic function, and (2) when the input is plus infinity or minus infinity, it converges to a predetermined value, or the more it deviates from a certain center position. , A function that decreases the absolute value of the slope.

FIG. 2 is a conceptual module configuration diagram of a configuration example in the arc evaluation value determination module 150 according to the first embodiment. The arc evaluation value determination module 150 includes a linear weighting addition module 210 and a nonlinear function module 220. The linear weighting addition module 210 and the nonlinear function module 220 are connected.
The linear weighted addition module 210 receives feature amounts 1 to N (a plurality of feature amounts relating to position candidates for cutting out one character image existing in the image) as feature vectors from the arc feature amount extraction module 140, and the above-described FIG. The calculation processing of the weighted linear sum equivalent to the linear weighted addition module 1710 in the example 17 is performed. The result is passed to the nonlinear function module 220.
The nonlinear function module 220 receives the calculation result from the linear weighted addition module 210 as an argument, and when the argument is a limit value, the nonlinear function module 220 converges to a predetermined value, or the argument and a predetermined value The evaluation value (arc evaluation value) of the candidate for the position to cut out one character image is calculated by a nonlinear monotone function that makes the absolute value of the output slope smaller as the distance increases. To do. That is, the calculation is performed using an arc evaluation function.

Assuming that the arc evaluation value function is V (f), the calculation of this configuration is expressed by equation (8). The input feature vector is f, the weight vector is w, and the scalar value of the weight is c. σ () is a nonlinear function.

By adopting such a configuration, the influence when an extremely large (or extremely small) value is taken is reduced.
In the example shown in FIG. 18, the input / output of the nonlinear function (linear weighted addition module 210) will be described, for example, in the following case.
Input: 1 Output: 1
Input: 10 Output: 2
Input: 100 Output: 3
If you do this,
Pass 1: The pass evaluation value in the case of “K” or “Study” is 2 × 2 + 2 = 6
Pass 2: The pass evaluation value for “I”, “Hi” and “Study” is 1 + 3 + 2 = 6
Thus, the influence of the value 100 can be reduced, and the evaluation values of the path 1 and the path 2 become comparable values.

  As an example of the non-linear function, any function such as a hyperbolic tangent function or a logistic sigmoid function may be used as long as it meets the above-described “non-linear function condition”.

When this embodiment is actually used, the weights (that is, w and c) in the linear weight addition module 210 must be determined. Therefore, a teacher data table 300 as shown in the example of FIG. 3 is prepared.
The teacher data table 300 includes a data number column 310, a feature amount 1 column 320, a feature amount 2 column 330, a feature amount N column 380, a correct / incorrect answer column 390, and the like. The data number column 310 stores a data number that uniquely identifies an arc. For example, an individual data number is assigned to each arc in order from 1. The feature quantity 1 column 320 to the feature quantity N column 380 store the feature quantities extracted by the arc feature quantity extraction module 140. That is, the arc feature quantity extraction module 140 extracts a feature quantity of a certain arc, and inputs it alongside the table of the teacher data table 300. Further, when the arc represents a break between correct characters, 1 is input in the correct / incorrect answer column 390, for example. If the arc does not represent a break between correct characters, for example, 0 is entered in the correct / incorrect answer column 390.

又は、（１０）式を最小とするように重みを決定すればよい。

又は、（１１）式を最小とするように重みを決定すればよい。

これらに限らず、Ｖ_ｋとｔ_ｋの値の差が小さいとき小さくなるような評価値を最小とするように重みを決定する方法であれば、どのようなものでもよい。
重み決定方式としては、一般的なロジスティック回帰の回帰係数決定方式や、単層パーセプトロンの重み決定方式を利用すればよい。 The weights may be determined so that values as close as possible to the above-mentioned correct / incorrect data can be obtained when feature values are input using the data created as described above as teacher data.
There are the following methods as evaluation methods for values as close as possible. First, let k be a data number that gives a number to each arc. Against arc data number k, and V _k arc evaluation value calculated using the arc evaluation value determination module 150. In addition, the teacher data of the arc of the number k and t _k. Here, for example,
• When the correct answer t _k = 1
・ Incorrect answer t _k = 0
And
At this time, the weight may be determined so as to minimize Equation (9).

Alternatively, the weight may be determined so as to minimize Equation (10).

Alternatively, the weight may be determined so as to minimize Equation (11).

The present invention is not limited to these, and any method may be used as long as the weight is determined so as to minimize the evaluation value that becomes small when the difference between the values of V _k and t _k is small.
As the weight determination method, a general logistic regression regression coefficient determination method or a single-layer perceptron weight determination method may be used.

  The character segmentation module 160 is connected to the arc evaluation value determination module 150 and the character recognition module 170, and based on the arc evaluation value calculated by the arc evaluation value determination module 150, character images existing in the character string image are extracted. The position to be cut out is determined, and the character image is cut out from the character string image (or the image received by the image receiving module 110).

<Second Embodiment>
In the first embodiment, since the nonlinear function is a monotone function, the magnitude of the output arc evaluation value changes, but the relative order does not change. That is, the above-described <Phenomenon 2> can occur.
The second embodiment has the configuration shown in the example of FIG. 1 and employs a configuration in which the arc evaluation value determination module 150 repeats the process based on the combination of the linear weighting addition and the nonlinear function twice.

FIG. 4 is a conceptual module configuration diagram of a configuration example in the arc evaluation value determination module 150 according to the second embodiment.
The arc evaluation value determination module 150 includes a linear weighted addition module 1-1: 411, a linear weighted addition module 1-2: 412,..., A linear weighted addition module 1-M: 41M, and a nonlinear function σ _1-1 module 421. , A non-linear function σ _1-2 module 422,..., A non-linear function σ _1-M module 42M, a linear weighting addition module 2: 430, and a non-linear function σ ₂ module 440. In FIG. 4, a plurality of feature amounts (feature amount 1 to feature amount N) are described as a feature amount vector by a single line.
The linear weighted addition module 1-1: 411 is connected to the nonlinear function σ _1-1 module 421.
Linear weighted addition Module 1-2: 412, is connected to the non-linear function sigma _1-2 module 422.
The linear weighted addition module 1-M: 41M is connected to the nonlinear function σ _1-M module 42M.
The nonlinear function σ _1-1 module 421 is connected to the linear weighted addition module 1-1: 411 and the linear weighted addition module 2: 430.
The nonlinear function σ _1-2 module 422 is connected to the linear weighted addition module 1-2: 412 and the linear weighted addition module 2: 430.
The nonlinear function σ _1-M module 42M is connected to the linear weighted addition module 1-M: 41M and the linear weighted addition module 2: 430.
Combination of linear weighted addition module 1-1: 411 and nonlinear function σ _1-1 module 421, combination of linear weighted addition module 1-2: 412 and nonlinear function σ _1-2 module 422, linear weighted addition module 1-M: The combination of 41M and the nonlinear function σ _1-M module 42M corresponds to the combination of the linear weighted addition module 210 and the nonlinear function module 220 in the first embodiment.
The linear weighted addition module 2: 430 is connected to the nonlinear function σ _1-1 module 421, the nonlinear function σ _1-2 module 422, the nonlinear function σ _1-M module 42M, and the nonlinear function σ ₂ module 440.
The nonlinear function σ ₂ module 440 is connected to the linear weighted addition module 2: 430.
The linear weighted addition module 2: 430 corresponds to the linear weighted addition module 210 in the first embodiment, and the nonlinear function σ ₂ module 440 corresponds to the nonlinear function module 220 in the first embodiment.

さらに、（１３）式に示すようにＵを定義する。

線形重み付け加算モジュール２：４３０と非線形関数σ_２モジュール４４０の組み合わせでは、（１４）式による演算を行って、出力値Ｖを得る。ｗ_２及びｃ_２は、線形重み付け加算モジュール２：４３０で用いる重みである。

前述において、非線形関数σは非線形関数σ_１−１モジュール４２１等の関数σ_１−ｉ（ただし、ｉ＝１，２，…，Ｍ）、及び関数σ_２の複数存在している。これらの関数は同じであってもよいし、異なっていてもよい。典型的にはハイパーボリックタンジェント関数やロジスティックシグモイド関数などを共通に用いればよい。
第２の実施の形態においては、その構成が３層のパーセプトロンと同等となる。そのため、正解アークと非正解アークが非線形な分離面を持っていても対応が可能となる。また、通常の誤差逆伝播方式を用いて重み係数を決定するようにしてもよい。その際の教師データとしては、図１３の例に示した教師用データテーブル３００を用いればよい。第２の実施の形態のアーク評価値決定モジュール１５０を用いて算出したアーク評価値ｙ_ｉと教師データｔ_ｉの値の差が小さいとき小さくなるような評価値の例も、第１の実施と同等である。 In the combination of the linear weighted addition module 1-i and the nonlinear function σ _1-i module (where i = 1, 2,..., M), the calculation according to the equation (12) is performed to obtain the output value U _i . w _1-i and c _1-i are weights used in the linear weighted addition module 1-i. M is the number of first-stage linear weighting addition modules.

Furthermore, U is defined as shown in equation (13).

Linear weighted addition module 2: 430 and the combination of the non-linear function sigma ₂ module 440, performs calculation according to (14) to obtain an output value V. w ₂ and c ₂ are weights used in the linear weighted addition module 2: 430.

As described above, the nonlinear function σ has a plurality of functions σ _1-i (where i = 1, 2,..., M) and the function σ ₂ such as the nonlinear function σ _1-1 module 421. These functions may be the same or different. Typically, a hyperbolic tangent function or a logistic sigmoid function may be used in common.
In the second embodiment, the configuration is equivalent to a three-layer perceptron. Therefore, even if the correct arc and the non-correct arc have a non-linear separation surface, it is possible to cope with it. Further, the weight coefficient may be determined using a normal error back propagation method. As the teacher data at that time, the teacher data table 300 shown in the example of FIG. 13 may be used. Examples of the second arc evaluation value determination module 150 arc evaluation value y _i and teacher evaluation value as reduced when the difference is small value of the data t _i calculated using the embodiment also, as in the first embodiment It is equivalent.

<Third Embodiment>
The third embodiment has the configuration shown in the example of FIG. 1, and further uses a plurality of arc evaluation value determination modules 150 of the second embodiment in the arc evaluation value determination module 150. , Take the sum.
Even if the performance of the arc evaluation value determination module 150 that is one estimator is poor, the performance can be improved by using a plurality of estimators. For example, assume that there are three estimators. It is assumed that one estimator is an incorrect answer and the remaining two estimators are correct answers. By taking the three majority decisions and adopting the correct answer, it is possible to estimate the correct answer.
In the example of the present embodiment, the majority operation is performed by addition.

FIG. 5 is a conceptual module configuration diagram of a configuration example in the arc evaluation value determination module 150 according to the third embodiment.
The arc evaluation value determination module 150 includes an arc evaluation value calculation module 1: 511, an arc evaluation value calculation module 2: 512, an arc evaluation value calculation module K: 51K, and an arc evaluation value addition module 520.
The arc evaluation value calculation module 1: 511, the arc evaluation value calculation module 2: 512, and the arc evaluation value calculation module K: 51K are connected to the arc evaluation value addition module 520, respectively.
In the third embodiment, a plurality of arc evaluation value calculation modules (arc evaluation value calculation module 1 to arc evaluation value calculation module K) are used. The output of the arc evaluation value calculation module j is V _j .

さらに、（１６）式に示すようにＵを定義する。

線形重み付け加算モジュールｊ−２と非線形関数σ_ｊ−２モジュールの組み合わせでは、（１７）式による演算を行って、出力値Ｖ_ｊを得る。ｗ_ｊ−２及びｃ_ｊ−２は、線形重み付け加算モジュールｊ−２が用いる重みである。

アーク評価値加算モジュール５２０では、アーク評価値算出モジュール１：５１１、アーク評価値算出モジュール２：５１２、・・・、アーク評価値算出モジュールＫ：５１Ｋによって計算されたアーク評価値の和を計算する。具体的には、例えば、（１８）式を用いて、アーク評価値Ｖを算出する。

このＶが、第３の実施の形態のアーク評価値決定モジュール１５０が文字切り出しモジュール１６０へ渡すアーク評価値である。 The arc evaluation value calculation module j adopts the same configuration as the arc evaluation value determination module 150 of the second embodiment.
FIG. 6 is a conceptual module configuration diagram of a configuration example in the arc evaluation value calculation module according to the third embodiment.
Linear weighted addition module j-1-1: 611, linear weighted addition module j-1-2: 612, linear weighted addition module j-1-Mj: 61M, nonlinear function σ _j-1-1 module 621, nonlinear function σ _{a j-1-2} module 622; a nonlinear function σ _j-1-Mj module 62M; a linear weighting addition module j-2: 630; and a nonlinear function σ _j-2 module 640.
The linear weighted addition module j-1-1: 611 is connected to the nonlinear function σ _j-1-1 module 621.
The linear weighted addition module j-1-2: 612 is connected to the nonlinear function σ _j-1-2 module 622.
The linear weighted addition module j-1-Mj: 61M is connected to the nonlinear function σ _j-1-Mj module 62M.
The nonlinear function σ _j-1-1 module 621 is connected to the linear weighted addition module j-1-1: 611 and the linear weighted addition module j-2: 630.
The nonlinear function σ _j-1-2 module 622 is connected to the linear weighted addition module j-1-2: 612 and the linear weighted addition module j-2: 630.
The nonlinear function σ _j-1-Mj module 62M is connected to the linear weighted addition module j-1-Mj: 61M and the linear weighted addition module j-2: 630.
The linear weighted addition module j-2: 630 includes a nonlinear function σ _j-1-1 module 621, a nonlinear function σ _j-1-2 module 622, a nonlinear function σ _j-1-Mj module 62M, and a nonlinear function σ _j-2. A module 640 is connected.
The nonlinear function σ _j-2 module 640 is connected to the linear weighted addition module j-2: 630.
The arc evaluation value calculation module j has the same operation as that of the second embodiment except that a subscript j is assigned to each component. The operation will be described below. In the combination of the linear weighting adder j-1-i and the nonlinear function σ _j-1-i (where i = 1, 2,..., Mj), the calculation according to the equation (15) is performed, and the output value U _{j− i} is obtained. w _j-1-i and c _j-1-i are weights used by the linear weighted addition module j-1-i. Mj is the number of first linear weighting addition modules.

Furthermore, U is defined as shown in equation (16).

In the combination of the linear weighted addition module j-2 and the nonlinear function σ _j-2 module, the calculation according to the equation (17) is performed to obtain the output value V _j . w _j-2 and c _j-2 are weights used by the linear weighted addition module j-2.

The arc evaluation value addition module 520 calculates the sum of the arc evaluation values calculated by the arc evaluation value calculation module 1: 511, the arc evaluation value calculation module 2: 512,..., The arc evaluation value calculation module K: 51K. . Specifically, for example, the arc evaluation value V is calculated using the equation (18).

This V is the arc evaluation value that the arc evaluation value determination module 150 of the third embodiment passes to the character segmentation module 160.

In the third embodiment, it is necessary to determine weights w _j-1-i , c _j-1-i , w _j-2 and c _j-2 . However, i = 1, 2,..., Mj, j = 1, 2,.
The weight determination method is described in the literature “J. Friedman, T. Hastie, R. Tibshirani“ Additive Logistic Regression: a Statistical View of Boosting, ”Anals of V.3. 407, 2000 "may be used.
Hereinafter, for convenience of explanation, for example, a function that takes 0 when the input is minus infinity and 1 when plus infinity is used as the nonlinear function σ. Actually, if linear transformation is performed according to a value that converges at minus infinity or plus infinity, the function may be changed to a function that takes −1 when the input is minus infinity and 1 when plus infinity.
Next, a teacher data table 700 shown in the example of FIG. 7 is prepared as teacher data. The teacher data table 700 has a data number column 710, a feature amount 1 column 720, a feature amount 2 column 730, a feature amount N column 770, a correct / incorrect answer column 780, a weight column 790, and the like. This is obtained by adding a weight column 790 to the teacher data table 300 shown in the example of FIG. Here, “weight” and “weight” are semantically the same. However, since the term “weight” has been used for the coefficient in the linear weighted addition module, the term “weight” is used for the weight of the teacher data in order to distinguish it from the term. The amount of data is G.

（イ）教師データのウエイトを更新する。
（イ−１）ｋ番目のデータに対するアーク評価値をＶ_ｊｋとする。
（イ−２）ウエイトを（２１）式で更新する。これは、アーク評価値の推定が間違ったデータのウエイトを大きくして、合っていたデータのウエイトを小さくする操作を示している。

（ウ）もしｊが所定の値以上になっているか、あるいは、評価値推定精度が十分であれば終了する。
（エ）ｊを１増大させて、（ア）に戻る。 Furthermore, the weight is determined using the following method.
Here, a symbol representing a correct / incorrect answer for data number k is y _k . In the third embodiment, for example,
・ When correct, y _k = + 1
・ Incorrect answer y _k = -1
And
Further, a logistic sigmoid function is adopted as the nonlinear function σ. At this time, V _j takes a value from 0 to 1.
1. First, all the data weights in the teacher data table 700 shown in the example of FIG.
2. Let j = 1.
(A) Using the weight of each teacher data, the weight of the arc evaluation value calculation module j is determined so as to minimize the weighted square error. The determination method is the same as that described in the description of the second embodiment. Normally, the weight is determined so as to minimize the square error, but the weight is determined so as to minimize the weighted square error weighted by the weight.
That is, in the first embodiment or the second embodiment, the weight is determined so as to minimize Equation (19), but here, Equation (20) is minimized. Determine the weight. However, there is a relationship of t _k = (y _k +1) / 2 here.

(B) Update the weight of the teacher data.
(A-1) Let the arc evaluation value for the kth data be V _jk .
(B-2) The weight is updated by equation (21). This indicates an operation of increasing the weight of data for which the estimation of the arc evaluation value is incorrect and decreasing the weight of the matched data.

(C) If j is equal to or greater than a predetermined value, or if the evaluation value estimation accuracy is sufficient, the process ends.
(D) Increase j by 1 and return to (a).

又は、（２３）式を最小とするように重みを決定すればよい。

In (a), the square error is minimized, but as described in the description of the first embodiment, the evaluation values of Equation (10), Equation (11), and others may be minimized. In this case, the weighting is the same, and each k may be multiplied by a _k . Specifically, the equation (22) is obtained. What is necessary is just to determine a weight so that this (22) Formula may be made into the minimum.

Alternatively, the weight may be determined so as to minimize Equation (23).

In the above description, a logistic sigmoid function is employed as the nonlinear function σ. At this time, since V _j takes a value from 0 to 1, 2V _jk −1 was calculated. This is to change the value range to a range of −1 to +1. This change is merely a linear transformation. Even when other nonlinear functions are used, linear transformation may be performed simply so that the range of values is in the range of −1 to +1.

In the above description, the feature quantity vector input to each arc evaluation value calculation module (arc evaluation value calculation module 1: 511, arc evaluation value calculation module 2: 512,..., Arc evaluation value calculation module K: 51K) is used. The contents are the same, but the contents may be different.
That is,
The input of the arc evaluation value calculation module 1 is only the height of the circumscribed rectangle of the character segment.
The input of the arc evaluation value calculation module 2 is set as all feature quantities.
The input of the arc evaluation value calculation module 3 is only the width of the character segment circumscribed rectangle.
・…
Etc.

In the above description, the configuration of each arc evaluation value calculator is the configuration of the arc evaluation value determination module 150 of the second embodiment (see the example of FIG. 4), but the arc evaluation value of the first embodiment. The configuration of the determination module 150 (see the example of FIG. 2) may be mixed and all may be the configuration of the arc evaluation value determination module 150 of the first embodiment. That is,
The arc evaluation value calculator 1 is a configuration of the arc evaluation value determination module 150 of the second embodiment. The arc evaluation value calculator 2 is a configuration of the arc evaluation value determination module 150 of the first embodiment. The calculator 3 is a configuration of the arc evaluation value determination module 150 of the second embodiment.
Or
The arc evaluation value calculator 1 is a configuration of the arc evaluation value determination module 150 of the first embodiment. The arc evaluation value calculator 2 is a configuration of the arc evaluation value determination module 150 of the first embodiment. The calculator 3 has the configuration of the arc evaluation value determination module 150 of the first embodiment.
Or
The arc evaluation value calculator 1 is a configuration of the arc evaluation value determination module 150 of the second embodiment. The arc evaluation value calculator 2 is a configuration of the arc evaluation value determination module 150 of the second embodiment. The calculator 3 is a configuration of the arc evaluation value determination module 150 of the second embodiment.
It is good.

<Fourth embodiment>
In the above-described embodiment, the arc evaluation value determination module 150 estimates the arc evaluation value.
The teacher data of the arc evaluation value to be estimated is, for example, 1 when the arc corresponds to the correct cutout position of the character, and 0 when the arc corresponds to the incorrect cutout position.
In that case, the following two types of optimization (weight determination) are performed.
As a two-class classification problem of class 0 and class 1, weights are determined so that errors in class classification are as small as possible.
-Minimize the square error between the estimated value existing between 0 and 1 and the teacher data (0 or 1) (it may be an evaluation value indicating an error such as an absolute value error or cross-entropy) Determine the weights to

However, even if the arc evaluation value is inaccurate, the character cutout position may be accurate. Conversely, even if the arc evaluation value is accurate, the character cutout position should not be inaccurate.
The accuracy of the arc evaluation value and the character cut-out position has a complicated relationship and is not a monotonous relationship. FIG. 8 shows a relationship example between the arc candidate determination module 810, the arc evaluation value determination module 820, and the character cutout position determination module 830.
The arc candidate determination module 810 is connected to the arc evaluation value determination module 820.
The arc evaluation value determination module 820 is connected to the arc candidate determination module 810 and the character cutout position determination module 830.
The character cutout position determination module 830 is connected to the arc evaluation value determination module 820.
In the character recognition process, first, the arc candidate determination module 810 receives an image and extracts a plurality of arc candidates as described above. Further, the arc evaluation value determination module 820 determines an arc evaluation value, and the character cutout position determination module 830 selects an optimum path from a plurality of paths as a set of a plurality of arc candidates. The character cutout position is fixed. Compared to the module configuration shown in the example of FIG. 1, the arc candidate determination module 810 corresponds to the image reception module 110 to the arc feature amount extraction module 140, and the arc evaluation value determination module 820 corresponds to the arc evaluation value determination module 150. The character cutout position determination module 830 corresponds to the character cutout module 160.
In the first to third embodiments, the arc evaluation value determination in the arc evaluation value determination module 150 is performed by referring only to the inside of the arc. An example of determining the weight used in the arc evaluation value determination module 820 in consideration of the whole will be described.
Hereinafter, it is assumed that the configuration of the arc evaluation value determination module 820 is any of those described in the description of the first to third embodiments.

The fourth embodiment relates to a method for determining the weights of the first to third embodiments. The configuration of the arc evaluation value determination module 820 is the arc evaluation value determination module 150 shown in the examples of the first to third embodiments.
FIG. 9 is a conceptual module configuration diagram of a configuration example according to the fourth embodiment.
As shown in the example of FIG. 9, the image processing apparatus according to the fourth embodiment includes an arc candidate determination module 910, a weight change module 920, an arc evaluation value determination module 930, a character cutout position determination module 940, and the number of correct cutout positions. A calculation module 950 is included. The arc candidate determination module 910 corresponds to the arc candidate determination module 810 shown in the example of FIG. 8, and the arc evaluation value determination module 930 corresponds to the arc evaluation value determination module 820 shown in the example of FIG. The position determination module 940 corresponds to the character cutout position determination module 830 shown in the example of FIG.

The arc candidate determination module 910 is connected to the arc evaluation value determination module 930, receives an image, and determines an arc candidate.
The weight change module 920 is connected to the arc evaluation value determination module 930 and the cut-out position correct answer number calculation module 950, and based on the correct number of cut-out positions calculated by the cut-out position correct number calculation module 950, The weight used in the arc evaluation value determination module 930 at the character cutout position is changed. Then, the next weight is determined from the amount of change from the correct answer number with the current weight to the correct answer number with the changed weight.

The arc evaluation value determination module 930 is connected to the arc candidate determination module 910, the weight change module 920, and the character extraction position determination module 940, receives the arc candidate from the arc candidate determination module 910, and receives the weight from the weight change module 920. Is used to determine the arc evaluation value.
The character cutout position determination module 940 is connected to the arc evaluation value determination module 930 and the cutout position correct answer number calculation module 950, and based on the arc evaluation value from the arc evaluation value determination module 930, the character image existing in the image Is determined, and the determined clipping position is passed to the clipping position correct number calculation module 950.
The cutout position correct number calculation module 950 is connected to the weight change module 920 and the character cutout position determination module 940, accepts the cutout position and the teacher data of the position to cut out the character image from the character cutout position determination module 940, and receives the character cutout position. The cutout position from the determination module 940 is compared with the teacher data to calculate the correct number of cutout positions.

Next, the flow of processing will be described.
First, the arc candidate determination module 910 receives an image and determines an arc candidate.
The initial weight used by the arc evaluation value determination module 930 may be a random number, or may be a weight determined by the method described in the description of the first to third embodiments. In any case, the weight change module 920 retains the initial weight.
Next, the arc evaluation value determination module 930 determines an arc evaluation value. Then, the character cutout position determination module 940 determines the character cutout position using the arc evaluation value.
The character cutout position after determination is passed to the cutout position correct number calculation module 950. Separately, character cutout teacher data is input to the cutout position correct answer number calculation module 950.
Here, the character cutout result is composed of, for example, a pair of a character code and a position and size of a circumscribed rectangle of the character in the image. The same applies to the character segmentation teacher data.
In the cutout position correct answer number calculation module 950,
・ Teacher characters: Multiple characters present in the character cutout teacher data (has the position, size, and character code of the circumscribed rectangle)
When,
Estimated character: Comparison with the character determined by the character cutout position determination module 940 is performed.
The cutout position correct answer number calculation module 950 calculates the character cutout position of the teacher character and the estimated character, and the number of matching character codes. Here, regarding the determination of matching between the character cutout position and the size, a minute shift may be allowed. Note that it is possible to determine only the match of the character cutout position and size without determining the match of the character code.

The number of correct answers determined as described above is passed to the weight change module 920.
The weight change module 920 determines the next weight by using the number of correct answers and the past weight held in the weight change module 920.
Here, W is a weight vector in which weights (that is, all w and c) used in the arc evaluation value determination module 930 are arranged. Let the elements of W be (W ₁ , W ₂ ,...).
In addition, the initial weight and _{W 0.} The following weight and _{W 1.} In this way, the weights are updated one after another. When the number of correct answers stops increasing, when the rate of increase in the number of correct answers falls below a predetermined value, or when the number of repetitions reaches a predetermined number of times, the weight change process is terminated. The weight at the time is output to the arc evaluation value determination module 930.

ここでαは重み更新の速度を規定するパラメタである。∇は、（２５）式を示す演算子である。

∇Ａは、Ｗ_ｍを変更したときのＡの変化量を示す。この変化の方向にＷを動かせばＡを増大させることができる。 Next, details of the processing of the weight change module 920 will be described.
First, let A be the number of correct character cutouts. A is a function of W. That is, it can be written as A (W). What is necessary is just to determine W so that A may be maximized. Now, let W _{m be the} current weight. Further, the changed weight is set to W _{m + 1} .
The weight update formula is the formula (24).

Here, α is a parameter that defines the speed of weight update. ∇ is an operator indicating the expression (25).

∇A indicates the amount of change of A when W _m is changed. If W is moved in the direction of this change, A can be increased.

又は、重みの更新の別の方法として下記の方法を用いる。ランダム、網羅的、又は予め定められたアルゴリズムを用いて更新量ｄＷを設定して、（２８）式とする。

このＷ_ｍ＋１を用いて、（２９）式の関係を有しているならば、ＷをＷ_ｍ＋１に更新する。（２９）式の関係を有していなければ、更新せずに、次のｄＷを試す。

以上で、正解個数を最大化する。 However, since the content of the function A (W) is unknown, it is impossible to calculate ∇A analytically. Therefore, an appropriate ε is determined, and ∇A is calculated numerically as Equation (26) or Equation (27).

Alternatively, the following method is used as another method of updating the weight. The update amount dW is set using a random, exhaustive, or predetermined algorithm to obtain equation (28).

If this W _{m + 1} is used and the relationship of equation (29) is satisfied, W is updated to W _{m + 1} . If the relationship of the formula (29) is not satisfied, the next dW is tried without updating.

Thus, the number of correct answers is maximized.

  In the above description, the number of correct answers is maximized, but the number of errors may be minimized. Alternatively, the error rate (that is, the number of errors / the number of correct answers, or the number of errors / (number of errors + number of correct answers)) may be minimized. Alternatively, the correct answer rate (that is, the number of correct answers / number of errors, or the number of correct answers / (number of errors + number of correct answers)) may be maximized. That is, as the values based on the number of correct answers or the number of errors, there are the number of correct answers, the number of errors, the correct answer rate, and the error rate.

  With reference to FIG. 10, a hardware configuration example of the image processing apparatuses according to the first to fourth embodiments will be described. The configuration illustrated in FIG. 10 is configured by, for example, a personal computer (PC), and illustrates a hardware configuration example including a data reading unit 1017 such as a scanner and a data output unit 1018 such as a printer.

  The CPU 1001 executes various modules described in the above-described embodiments, that is, the modules illustrated in the examples of FIGS. 1, 2, 4, 5, 6, 6, 8, 9, and 17 It is a control part which performs the process according to the computer program which described the sequence.

  A ROM (Read Only Memory) 1002 stores programs used by the CPU 1001, calculation parameters, and the like. The RAM 1003 stores programs used in the execution of the CPU 1001, parameters that change as appropriate during the execution, and the like. These are connected to each other by a host bus 1004 including a CPU bus.

  The host bus 1004 is connected to an external bus 1006 such as a PCI (Peripheral Component Interconnect / Interface) bus via a bridge 1005.

  A keyboard 1008 and a pointing device 1009 such as a mouse are input devices operated by an operator. The display 1010 includes a liquid crystal display device or a CRT (Cathode Ray Tube), and displays various types of information as text or image information.

  An HDD (Hard Disk Drive) 1011 includes a hard disk, drives the hard disk, and records or reproduces a program executed by the CPU 1001 and information. The hard disk stores images, character images, character cutout positions, character cutout position candidates, a teacher data table 300, a teacher data table 700, and the like. Further, various computer programs such as various other data processing programs are stored.

  The drive 1012 reads data or a program recorded on a removable recording medium 1013 such as a magnetic disk, an optical disk, a magneto-optical disk, or a semiconductor memory, and the data or program is read out to the interface 1007 and the external bus 1006. , The bridge 1005, and the RAM 1003 connected via the host bus 1004. The removable recording medium 1013 can also be used as a data recording area similar to a hard disk.

  The connection port 1014 is a port for connecting the external connection device 1015 and has a connection unit such as USB and IEEE1394. The connection port 1014 is connected to the CPU 1001 and the like via the interface 1007, the external bus 1006, the bridge 1005, the host bus 1004, and the like. A communication unit 1016 is connected to a network and executes data communication processing with the outside. The data reading unit 1017 is a scanner, for example, and executes document reading processing. The data output unit 1018 is a printer, for example, and executes document data output processing.

  Note that the hardware configuration of the image processing apparatus shown in FIG. 10 shows one configuration example, and the first to fourth embodiments are not limited to the configuration shown in FIG. Any configuration can be used as long as the module described in the embodiment can be executed. For example, some modules may be configured with dedicated hardware (for example, Application Specific Integrated Circuit (ASIC), etc.), and some modules are in an external system and connected via a communication line In addition, a plurality of systems shown in FIG. 10 may be connected to each other via communication lines so as to cooperate with each other. Further, it may be incorporated in a copying machine, a fax machine, a scanner, a printer, a multifunction machine (an image processing apparatus having any two or more functions of a scanner, a printer, a copying machine, a fax machine, etc.).

Note that the various embodiments described above may be combined (for example, a module in one embodiment may be applied to another embodiment, replaced, etc.), and the background art may be used as the processing content of each module. You may employ | adopt the technique demonstrated by.
In addition, although demonstrated using a numerical formula, the thing equivalent to the numerical formula is contained in a numerical formula. The equivalent includes not only the mathematical formula itself, but also transformation of the mathematical formula to the extent that the final result is not affected, or solving the mathematical formula by an algorithmic solution.

The program described above may be provided by being stored in a recording medium, or the program may be provided by communication means. In that case, for example, the above-described program may be regarded as an invention of a “computer-readable recording medium recording the program”.
The “computer-readable recording medium on which a program is recorded” refers to a computer-readable recording medium on which a program is recorded, which is used for program installation, execution, program distribution, and the like.
The recording medium is, for example, a digital versatile disc (DVD), which is a standard established by the DVD Forum, such as “DVD-R, DVD-RW, DVD-RAM,” and DVD + RW. Standard “DVD + R, DVD + RW, etc.”, compact disc (CD), read-only memory (CD-ROM), CD recordable (CD-R), CD rewritable (CD-RW), Blu-ray disc ( Blu-ray Disc (registered trademark), magneto-optical disk (MO), flexible disk (FD), magnetic tape, hard disk, read-only memory (ROM), electrically erasable and rewritable read-only memory (EEPROM), flash Includes memory, random access memory (RAM), etc. .
The program or a part of the program may be recorded on the recording medium for storage or distribution. Also, by communication, for example, a local area network (LAN), a metropolitan area network (MAN), a wide area network (WAN), a wired network used for the Internet, an intranet, an extranet, etc., or wireless communication It may be transmitted using a transmission medium such as a network or a combination of these, or may be carried on a carrier wave.
Furthermore, the program may be a part of another program, or may be recorded on a recording medium together with a separate program. Moreover, it may be divided and recorded on a plurality of recording media. Further, it may be recorded in any manner as long as it can be restored, such as compression or encryption.

41M ... Linear weighting addition module 1-M
42M: nonlinear function σ _1-M module 51K: arc evaluation value calculation module K
61M: Linear weighted addition module j-1-Mj
62M ... Nonlinear function [sigma] _j-1-Mj module 110 ... Image reception module 120 ... Character string extraction module 130 ... Character boundary candidate extraction module 140 ... Arc feature value extraction module 150 ... Arc evaluation value determination module 160 ... Character segmentation module 170 ... Character recognition module 210 ... Linear weighting addition module 220 ... Nonlinear function module 411 ... Linear weighting addition module 1-1
412 ... Linear weighting addition module 1-2
421: nonlinear function σ _1-1 module 422: nonlinear function σ _1-2 module 430: linear weighted addition module 2
440: nonlinear function σ ₂ module 511: arc evaluation value calculation module 1
512 ... Arc evaluation value calculation module 2
520 ... Arc evaluation value addition module 611 ... Linear weighting addition module j-1-1
612 ... Linear weighting addition module j-1-2
621 ... Nonlinear function σ _j-1-1 module 622 ... Nonlinear function σ _j-1-2 module 630 ... Linear weighted addition module j-2
640: Nonlinear function σ _j-2 module 810: Arc candidate determination module 820 ... Arc evaluation value determination module 830 ... Character extraction position determination module 910 ... Arc candidate determination module 920 ... Weight change module 930 ... Arc evaluation value determination module 940 ... Character Cutout position determination module 950 ... Cutout position correct answer number calculation module 1710 ... Linear weighting addition module

Claims (8)

First calculating means for calculating a weighted linear sum for a plurality of feature amounts related to a position candidate for cutting out one character image existing in the image;
The calculation result by the first calculation means is used as an argument, and when the argument is a limit value, it converges to a predetermined value, or output when the distance between the argument and the predetermined value increases Second calculating means for calculating an evaluation value of a candidate for a position to cut out the one character image by a non-linear monotone function in which the absolute value of the slope of the non-linear monotonic function or a function approximating the non-linear monotone function is reduced;
Based on the evaluation value calculated by the second calculating means, a cutting position determining means for determining a position for cutting out a character image existing in the image ;
Accepting means for accepting teacher data of the cutout position of the character image;
A number calculating means for comparing the cutout position determined by the cutout position determining means with the teacher data received by the receiving means, and calculating the correct number of answers or the number of errors of the cutout position;
Weight change means for changing the weight used in the first calculation means at the character cutout position for one character based on the correct number or the error number of the cutout position calculated by the number calculation means.
Comprising
The weight changing means determines a next weight from an amount of change from a value based on the number of correct answers or the number of errors in the case of the current weight to a value based on the number of correct answers or the number of errors in the changed weight.
The image processing apparatus characterized by. First calculating means for calculating a weighted linear sum for a plurality of feature amounts related to a position candidate for cutting out one character image existing in the image;
The calculation result by the first calculation means is used as an argument, and when the argument is a limit value, it converges to a predetermined value, or output when the distance between the argument and the predetermined value increases Second calculating means for calculating an evaluation value of a candidate for a position to cut out the one character image by a non-linear monotone function in which the absolute value of the slope of the non-linear monotonic function or a function approximating the non-linear monotone function is reduced;
Based on the evaluation value calculated by the second calculating means, a cutting position determining means for determining a position for cutting out a character image existing in the image;
A plurality of sets of the first calculation means and the second calculation means;
Third calculation means for calculating a weighted linear sum for the evaluation values calculated by the plurality of second calculation means;
The calculation result by the third calculation means is used as an argument, and when the argument is a limit value, it converges to a predetermined value, or output when the distance between the argument and the predetermined value increases A fourth calculation means for calculating an evaluation value of a candidate for a position to cut out the one character image by a nonlinear monotone function in which the absolute value of the slope of the curve becomes small or a function approximating the nonlinear monotone function;
The cut-out position determining means determines a position to cut out a character image existing in the image based on the evaluation value calculated by the fourth calculation means ,
Accepting means for accepting teacher data of the cutout position of the character image;
A number calculating means for comparing the cutout position determined by the cutout position determining means with the teacher data received by the receiving means, and calculating the correct number of answers or the number of errors of the cutout position;
A weight change for changing the weight used in the first calculation means or the third calculation means at the character cutout position for one character based on the correct number or the error number of the cutout position calculated by the number calculation means means
Comprising
The weight changing means determines a next weight from a change amount from a value based on the number of correct answers or the number of errors in the case of the current weight to a value based on the number of correct answers or the number of errors in the changed weight. An image processing apparatus. First calculating means for calculating a weighted linear sum for a plurality of feature amounts related to a position candidate for cutting out one character image existing in the image;
The calculation result by the first calculation means is used as an argument, and when the argument is a limit value, it converges to a predetermined value, or output when the distance between the argument and the predetermined value increases Second calculating means for calculating an evaluation value of a candidate for a position to cut out the one character image by a non-linear monotone function in which the absolute value of the slope of the non-linear monotonic function or a function approximating the non-linear monotone function is reduced;
Based on the evaluation value calculated by the second calculating means, a cutting position determining means for determining a position for cutting out a character image existing in the image;
A plurality of sets of the first calculation means and the second calculation means;
Fifth calculation means for calculating a sum of evaluation values calculated by the plurality of second calculation means;
The cut-out position determining means determines a position to cut out a character image existing in the image based on the sum of evaluation values calculated by the fifth calculation means ,
Accepting means for accepting teacher data of the cutout position of the character image;
A number calculating means for comparing the cutout position determined by the cutout position determining means with the teacher data received by the receiving means, and calculating the correct number of answers or the number of errors of the cutout position;
Weight change means for changing the weight used in the first calculation means at the character cutout position for one character based on the correct number or the error number of the cutout position calculated by the number calculation means.
Comprising
The weight changing means determines a next weight from a change amount from a value based on the number of correct answers or the number of errors in the case of the current weight to a value based on the number of correct answers or the number of errors in the changed weight. An image processing apparatus. First calculating means for calculating a weighted linear sum for a plurality of feature amounts related to a position candidate for cutting out one character image existing in the image;
The calculation result by the first calculation means is used as an argument, and when the argument is a limit value, it converges to a predetermined value, or output when the distance between the argument and the predetermined value increases Second calculating means for calculating an evaluation value of a candidate for a position to cut out the one character image by a non-linear monotone function in which the absolute value of the slope of the non-linear monotonic function or a function approximating the non-linear monotone function is reduced;
Based on the evaluation value calculated by the second calculating means, a cutting position determining means for determining a position for cutting out a character image existing in the image;
A plurality of sets of the first calculation means and the second calculation means;
Third calculation means for calculating a weighted linear sum for the evaluation values calculated by the plurality of second calculation means;
The calculation result by the third calculation means is used as an argument, and when the argument is a limit value, it converges to a predetermined value, or output when the distance between the argument and the predetermined value increases Fourth calculation means for calculating an evaluation value of a candidate for a position to cut out the one character image by a nonlinear monotone function or an approximate function of the nonlinear monotone function in which the absolute value of the slope of the image becomes small
Comprising
The cut-out position determining means determines a position to cut out a character image existing in the image based on the evaluation value calculated by the fourth calculation means,
A plurality of sets by the first calculation means and the second calculation means; a plurality of sets by the third calculation means and the fourth calculation means;
Sixth calculation means for calculating a sum of evaluation values calculated by the plurality of fourth calculation means,
The cut-out position determining means determines a position to cut out a character image existing in the image based on the sum of evaluation values calculated by the sixth calculation means ,
Accepting means for accepting teacher data of the cutout position of the character image;
A number calculating means for comparing the cutout position determined by the cutout position determining means with the teacher data received by the receiving means, and calculating the correct number of answers or the number of errors of the cutout position;
A weight change for changing the weight used in the first calculation means or the third calculation means at the character cutout position for one character based on the correct number or the error number of the cutout position calculated by the number calculation means means
Comprising
The weight changing means determines a next weight from a change amount from a value based on the number of correct answers or the number of errors in the case of the current weight to a value based on the number of correct answers or the number of errors in the changed weight. An image processing apparatus. Computer
First calculating means for calculating a weighted linear sum for a plurality of feature amounts related to a position candidate for cutting out one character image existing in the image;
The calculation result by the first calculation means is used as an argument, and when the argument is a limit value, it converges to a predetermined value, or output when the distance between the argument and the predetermined value increases Second calculating means for calculating an evaluation value of a candidate for a position to cut out the one character image by a non-linear monotone function in which the absolute value of the slope of the non-linear monotonic function or a function approximating the non-linear monotone function is reduced;
Based on the evaluation value calculated by the second calculating means, a cutting position determining means for determining a position for cutting out a character image existing in the image ;
Accepting means for accepting teacher data of the cutout position of the character image;
A number calculating means for comparing the cutout position determined by the cutout position determining means with the teacher data received by the receiving means, and calculating the correct number of answers or the number of errors of the cutout position;
Weight change means for changing the weight used in the first calculation means at the character cutout position for one character based on the correct number or the error number of the cutout position calculated by the number calculation means.
Function as
The weight changing means determines a next weight from an amount of change from a value based on the number of correct answers or the number of errors in the case of the current weight to a value based on the number of correct answers or the number of errors in the changed weight.
An image processing program characterized by that .   Computer
  First calculating means for calculating a weighted linear sum for a plurality of feature amounts related to a position candidate for cutting out one character image existing in the image;
  The calculation result by the first calculation means is used as an argument, and when the argument is a limit value, it converges to a predetermined value, or output when the distance between the argument and the predetermined value increases Second calculating means for calculating an evaluation value of a candidate for a position to cut out the one character image by a non-linear monotone function in which the absolute value of the slope of the non-linear monotonic function or a function approximating the non-linear monotone function is reduced;
  Based on the evaluation value calculated by the second calculating means, a cutting position determining means for determining a position for cutting out a character image existing in the image;
  A plurality of sets of the first calculation means and the second calculation means;
  Third calculation means for calculating a weighted linear sum for the evaluation values calculated by the plurality of second calculation means;
  The calculation result by the third calculation means is used as an argument, and when the argument is a limit value, it converges to a predetermined value, or output when the distance between the argument and the predetermined value increases Fourth calculation means for calculating an evaluation value of a candidate for a position to cut out the one character image by a nonlinear monotone function or an approximate function of the nonlinear monotone function in which the absolute value of the slope of the image becomes small
  Function as
  The cut-out position determining means determines a position to cut out a character image existing in the image based on the evaluation value calculated by the fourth calculation means,
  Accepting means for accepting teacher data of the cutout position of the character image;
  A number calculating means for comparing the cutout position determined by the cutout position determining means with the teacher data received by the receiving means, and calculating the correct number of answers or the number of errors of the cutout position;
  A weight change for changing the weight used in the first calculation means or the third calculation means at the character cutout position for one character based on the correct number or the error number of the cutout position calculated by the number calculation means means
  Function as
  The weight changing means determines a next weight from an amount of change from a value based on the number of correct answers or the number of errors in the case of the current weight to a value based on the number of correct answers or the number of errors in the changed weight.
  An image processing program characterized by that.   Computer
  First calculating means for calculating a weighted linear sum for a plurality of feature amounts related to a position candidate for cutting out one character image existing in the image;
  The calculation result by the first calculation means is used as an argument, and when the argument is a limit value, it converges to a predetermined value, or output when the distance between the argument and the predetermined value increases Second calculating means for calculating an evaluation value of a candidate for a position to cut out the one character image by a non-linear monotone function in which the absolute value of the slope of the non-linear monotonic function or a function approximating the non-linear monotone function is reduced;
  Based on the evaluation value calculated by the second calculating means, a cutting position determining means for determining a position for cutting out a character image existing in the image;
  A plurality of sets of the first calculation means and the second calculation means;
  Fifth calculation means for calculating the sum of the evaluation values calculated by the plurality of second calculation means
  Function as
  The cut-out position determining means determines a position to cut out a character image existing in the image based on the sum of evaluation values calculated by the fifth calculation means,
  Accepting means for accepting teacher data of the cutout position of the character image;
  A number calculating means for comparing the cutout position determined by the cutout position determining means with the teacher data received by the receiving means, and calculating the correct number of answers or the number of errors of the cutout position;
  Weight change means for changing the weight used in the first calculation means at the character cutout position for one character based on the correct number or the error number of the cutout position calculated by the number calculation means.
  Function as
  The weight changing means determines a next weight from an amount of change from a value based on the number of correct answers or the number of errors in the case of the current weight to a value based on the number of correct answers or the number of errors in the changed weight.
  An image processing program characterized by that.   Computer
  First calculating means for calculating a weighted linear sum for a plurality of feature amounts related to a position candidate for cutting out one character image existing in the image;
  The calculation result by the first calculation means is used as an argument, and when the argument is a limit value, it converges to a predetermined value, or output when the distance between the argument and the predetermined value increases Second calculating means for calculating an evaluation value of a candidate for a position to cut out the one character image by a non-linear monotone function in which the absolute value of the slope of the non-linear monotonic function or a function approximating the non-linear monotone function is reduced;
  Based on the evaluation value calculated by the second calculating means, a cutting position determining means for determining a position for cutting out a character image existing in the image;
  A plurality of sets of the first calculation means and the second calculation means;
  Third calculation means for calculating a weighted linear sum for the evaluation values calculated by the plurality of second calculation means;
  The calculation result by the third calculation means is used as an argument, and when the argument is a limit value, it converges to a predetermined value, or output when the distance between the argument and the predetermined value increases Fourth calculation means for calculating an evaluation value of a candidate for a position to cut out the one character image by a nonlinear monotone function or an approximate function of the nonlinear monotone function in which the absolute value of the slope of the image becomes small
  Function as
  The cut-out position determining means determines a position to cut out a character image existing in the image based on the evaluation value calculated by the fourth calculation means,
  A plurality of sets by the first calculation means and the second calculation means; a plurality of sets by the third calculation means and the fourth calculation means;
  Sixth calculation means for calculating a sum of evaluation values calculated by the plurality of fourth calculation means.
  Function as
  The cut-out position determining means determines a position to cut out a character image existing in the image based on the sum of evaluation values calculated by the sixth calculation means,
  Accepting means for accepting teacher data of the cutout position of the character image;
  A number calculating means for comparing the cutout position determined by the cutout position determining means with the teacher data received by the receiving means, and calculating the correct number of answers or the number of errors of the cutout position;
  A weight change for changing the weight used in the first calculation means or the third calculation means at the character cutout position for one character based on the correct number or the error number of the cutout position calculated by the number calculation means means
  Function as
  The weight changing means determines a next weight from an amount of change from a value based on the number of correct answers or the number of errors in the case of the current weight to a value based on the number of correct answers or the number of errors in the changed weight.
  An image processing program characterized by that.

Images