JP5729260B2 - Computer program for character recognition, character recognition device, and character recognition method - Google Patents

Description

  The present invention relates to a character recognition computer program, a character recognition device, and a character recognition method, for example, for recognizing each character included in a character string from an image of a character string represented on a medium.

  In recent years, optical character recognition (Optical Character Recognition, OCR) is a character recognition that recognizes a character string by analyzing an image of the character string represented on a medium such as paper or a display, and converts it into electronic data. Technology is being used. In such a character recognition technique, it has been studied to improve recognition accuracy by verifying consistency as a word or a sentence with respect to each recognized combination of characters (for example, Patent Document 1). ~ 3).

  For example, the document recognition apparatus disclosed in Patent Document 1 obtains a recognition candidate character group for each character pattern on the image, and collates the recognition candidate character group of each character pattern with a town name dictionary to identify a town name. Recognize

  Moreover, the character recognition apparatus disclosed by patent document 2 character-recognizes each character pattern, and calculates | requires the certainty degree of a some candidate character and each candidate character. The character recognition device searches for a word from the candidate character string and determines the occurrence probability of the searched word. The character recognition apparatus selects the optimum candidate character string by integrating the certainty factor of each character, the character n-gram probability, and the word occurrence probability.

  Further, the search method disclosed in Patent Document 3 generates a derived character string in which a first character included in the search character string and a second character corresponding to the first character are generated, and is obtained from the search target document. The search character string and the derived character string are searched.

JP 2000-148906 A JP-A-11-328316 JP 2010-225137 A

  However, in character recognition technology, when recognizing individual characters in a character string, one character may be misrecognized as two characters, resulting in insertion of a character that should not originally exist, Characters that should exist may be lost. For example, one character “movement” may be erroneously recognized as two characters “heavy” and “power”. In such a case, the length of the character string connecting the recognized individual characters is different from the original length of the character string. Therefore, with the technique disclosed in Patent Document 1 or Patent Document 2, there is a possibility that a word that matches the recognized character string cannot be found in the dictionary to be searched.

  On the other hand, in the technique disclosed in Patent Document 3, one recognized character is replaced with a plurality of characters corresponding to a previously registered erroneous recognition pattern, or a plurality of recognized characters are replaced with one character. By doing so, a derived character string is generated. Therefore, this technique can recognize a character string accurately even if the length of the recognized character string is different from the length of the original character string by misrecognizing each character.

  However, in the technique disclosed in Patent Document 3, the number of derived character strings increases as the character string to be recognized becomes longer. As a result, the amount of calculation required to search the search character string and the derived character string from the search target document becomes enormous, and the time required to obtain the search result may be increased.

  Therefore, the present specification can accurately recognize a character string even when there is an error in the recognition result of each character included in the character string on the image, and can reduce the amount of calculation required for character string recognition. An object is to provide a computer program for character recognition.

  According to one embodiment, a computer program for character recognition is provided. This computer program for character recognition uses a character section on an image that is presumed to be captured for each of a plurality of characters from an image of a character string including a plurality of characters represented on a medium. Is obtained for each of the plurality of character sections, and a candidate character lattice that is a set of paths representing a relative positional relationship with other character sections is obtained. At least one candidate character that is a candidate for a character in the corresponding character section is obtained, and at least a part of the first pass overlaps the first pass among the plurality of passes included in the candidate character lattice. The second path of the two or more consecutive paths including the first path candidate character and the second path candidate character, and corresponding to the same character section as the second path Third pa Or the candidate character lattice is corrected so that the first pass is inserted between the two or more consecutive passes, and the corrected candidate among a plurality of words registered in the word dictionary A word that at least partially matches a combination of candidate characters included in each of two or more consecutive paths in the character lattice is detected as a word that may be included in the character string, and the character on the image of the detected word A candidate character and a word lattice obtained by adding a word path representing a position in the sequence to the candidate character lattice are obtained, and in the candidate character and the word lattice, for each word path, the probability that the word corresponding to the word path is included in the character string. An evaluation value that represents the word path and the word path included in the candidate character and the word lattice, and the word path and the continuous word path corresponding to the entire character string selected from the path. The total value of the evaluation values is obtained for each path array, and the word path included in the array having the highest total value and the word path and the combination of words corresponding to the path and candidate characters arranged according to the path order are displayed on the image. It has an instruction which makes a computer perform presumption as a character string.

The objects and advantages of the invention will be realized and attained by means of the elements and combinations particularly pointed out in the appended claims.
It should be understood that both the foregoing general description and the following detailed description are exemplary and explanatory and are not restrictive of the invention as claimed.

  The computer program for character recognition disclosed herein can accurately recognize a character string even if there is an error in the recognition result of each character included in the character string on the image, and the calculation required for character string recognition. The amount can be reduced.

It is a schematic block diagram of the character recognition apparatus by one Embodiment. It is a block diagram which shows the function of a process part. It is a figure which shows an example of the character string reflected in the image, and the character area detected with respect to the character string. It is a figure which shows an example of the candidate character lattice corresponding to the character string shown in FIG. It is a figure which shows an example of the candidate character lattice with respect to another character string. It is an operation | movement flowchart of an example of a candidate character lattice correction process. It is an operation | movement flowchart of another example of a candidate character lattice correction process. It is explanatory drawing of the search procedure of the optimal path | pass using DP matching. It is a figure which shows an example of the candidate character and word lattice which added the path | pass of the detected word with respect to the character string shown in FIG. It is an operation | movement flowchart of a character recognition process.

  Hereinafter, a character recognition device according to an embodiment will be described with reference to the drawings. This character recognition device recognizes a character string from an image obtained by photographing a character string including a plurality of characters represented on a medium such as paper or a display screen, and converts the character string into electronic data. For this purpose, this character recognition device generates a candidate character lattice that is a set of paths corresponding to character sections in which individual characters are shown and representing the relative positional relationship of the character sections. In order to reduce the amount of computation, this character recognition device can be used to eliminate duplication when one path included in a candidate character lattice overlaps at least partly with respect to another path. By correcting these paths, the number of path combinations when searching for words is reduced. Furthermore, this character recognition device performs a fuzzy search on a combination of a plurality of candidate characters for each path when searching for a word, so that a word that may be included in a character string due to erroneous recognition of individual characters. Is prevented from being detected and the recognition accuracy of the character string is improved.

  FIG. 1 is a schematic configuration diagram of a character recognition device according to an embodiment. The character recognition device 1 includes an image acquisition unit 11, an output unit 12, a storage unit 13, a storage medium access device 14, and a processing unit 15. Furthermore, the character recognition device 1 may include an input device such as a plurality of operation buttons and a display device such as a liquid crystal display that displays a character string converted into electronic data. The processing unit 15 is connected to the image acquisition unit 11, the output unit 12, the storage unit 13, and the storage medium access device 14 via, for example, a bus.

  The image acquisition unit 11 includes, for example, a digital camera or a scanner that captures a character string represented on a medium. Then, the image acquisition unit 11 generates an image showing the character string, and outputs the image to the processing unit 15.

Alternatively, the image acquisition unit 11 may include a communication interface for connecting the character recognition device 1 to an image input device (not shown) such as a digital camera or a camera-equipped mobile phone and a control circuit thereof. Such a communication interface can be an interface in accordance with a communication standard for connecting peripheral devices such as Universal Serial Bus (Universal Serial Bus, USB).
Alternatively, the image acquisition unit 11 may include a communication interface for connecting to a communication network in accordance with a communication standard such as Ethernet (registered trademark) and its control circuit.
In this case, the image acquisition unit 11 acquires an image showing a character string from the image input device or another device connected to the communication network, and passes the image to the processing unit 15.

The output unit 12 includes, for example, a communication interface for connecting the character recognition device 1 to another device and its control circuit. Such a communication interface can be an interface according to a communication standard for connecting peripheral devices such as USB or a communication standard such as Ethernet (registered trademark).
The output unit 12 receives from the processing unit 15 a character string converted into electronic data, which is generated when the processing unit 15 recognizes the character string shown in the image, and receives the converted character string as another device. Output to. Note that the image acquisition unit 11 and the output unit 12 may be integrated.

  The storage unit 13 includes, for example, a readable / writable semiconductor memory and a read-only semiconductor memory. The storage unit 13 is a computer program executed on the processing unit 15 and various information used for recognizing the character string on the image, for example, a word dictionary in which a plurality of words to be searched are registered, The candidate character and candidate character lattice for each pass are stored. The storage unit 13 may store an image showing a character string to be recognized.

  The storage medium access device 14 is a device that accesses the storage medium 16 such as a magnetic disk, a semiconductor memory card, and an optical storage medium. For example, the storage medium access device 14 reads a computer program for character recognition executed on the processing unit 15 stored in the storage medium 16 and passes the computer program to the processing unit 15 or stores it in the storage unit 13. The storage medium access device 14 may write the character string converted into electronic data generated by the processing unit 15 into the storage medium 16.

  The processing unit 15 includes one or a plurality of processors and their peripheral circuits. Then, the processing unit 15 recognizes each character included in the character string from an image showing a character string including a plurality of characters, and arranges the character code corresponding to each character in order from the beginning of the character string, thereby electronic data. Generate a character string.

  FIG. 2 is a block diagram illustrating functions of the processing unit 15. The processing unit 15 includes a character section detection unit 21, a candidate character lattice generation unit 22, a candidate character extraction unit 23, a candidate character lattice correction unit 24, a word search unit 25, and an estimation unit 26. Each of these units included in the processing unit 15 is, for example, a functional module implemented by a computer program executed on a processor included in the processing unit 15. Alternatively, each of these units included in the processing unit 15 may be mounted on the character recognition device 1 as an integrated circuit in which circuits for realizing the respective functions are integrated.

  The character section detection unit 21 detects a character section that is a section in which the character is estimated to be captured for each of a plurality of characters included in the character string captured on the image.

  In this embodiment, the value of each pixel is a value within the range of 0 to 255 in an image obtained by photographing a character represented in black on a white background, such as a character string printed on a medium such as paper. And the higher the density, the higher the pixel value. In addition, when the luminance of the character is higher than the luminance of the background, such as a character string displayed on the display screen, the image may be generated such that the higher the luminance is, the higher the pixel value is. . The character string to be recognized is a so-called horizontally written character string and is assumed to be along the horizontal direction on the image. If the character string to be recognized is a so-called vertical character string, the horizontal direction and the vertical direction in the following description may be switched.

  As described above, in the present embodiment, the value of the pixel in which the character is shown on the image is higher than the value of the pixel in which the character is not shown. For this reason, the average value of the pixels corresponding to the separation between characters is lower than the average value of the pixels in which the characters are shown. Therefore, the character section detection unit 21 calculates the sum of the pixel values for each vertical line included in the region corresponding to the character string of one line in the image, and calculates the total value at the horizontal position corresponding to the vertical line. The projection value is a pixel value. The character section detection unit 21 sets a position where the projection value is equal to or less than a predetermined threshold as a character break. The predetermined threshold is set to a value obtained by multiplying the average value of the projection values obtained by dividing the total of the projection values by the number of pixels in the horizontal direction by a predetermined coefficient, for example. The predetermined coefficient is a positive value of 1 or less, for example, 0.5. Note that the character section detection unit 21 may detect a break between characters using a technique disclosed in Japanese Patent Application Laid-Open No. Hei 2-217978. Further, the character section detecting unit 21 binarizes each pixel of the image into a character pixel in which the character is reflected and a background pixel in which the character is not reflected, and totals the number of character pixels for each vertical line, The number of character pixels may be used as the projection value. In this case, the threshold value for binarizing the pixels may be set, for example, by discriminating and analyzing the average value of the pixel values of the entire image or the distribution of the pixel values.

  The character section detection unit 21 sets each of the sections sandwiched between two characters adjacent to each other in the horizontal direction as a character section. The character section detection unit 21 assigns a unique identification number to each character section, and stores the horizontal coordinates of the left and right ends of the character section in the storage unit 13 together with the unique identification number.

  A character string may be shown over a plurality of lines on the image. Therefore, when there is a possibility that a plurality of character strings are reflected on the image, the character section detection unit 21 may detect the separation between lines before detecting the separation between the characters. In this case, the character section detection unit 21 calculates the sum of the pixel values for each horizontal line in the image, and sets the total value as the projection value of the pixel value at the vertical position corresponding to the horizontal line. Then, the character section detection unit 21 sets the position where the projection value is the minimum value and is equal to or less than the threshold for line break detection as the line break. Note that the threshold for detecting line breaks is obtained by, for example, multiplying the average value of projection values obtained by dividing the total projection value for each line in the horizontal direction by the number of pixels in the vertical direction by a predetermined coefficient (for example, 0.5). Value is set. The character section detection unit 21 divides the image into regions that are sandwiched between two adjacent rows in the vertical direction, and uses the divided image as an image representing one character string. Then, the character section detection unit 21 detects the character section by detecting a break between characters as described above for each image representing each character string.

  Also, depending on the positional relationship between the camera that captures the character string and the character string, the row direction of the character string may not match the horizontal direction of the image. Therefore, for example, the character section detection unit 21 may rotate an image by a predetermined angle (for example, 5 degrees) by affine transformation, and obtain a projection value for each horizontal line in the rotated image. In this case, the character section detection unit 21 detects each character string using the image of the rotation angle when the projection value is the minimum value and the number of lines having a predetermined threshold value or less is the largest, and then for each character string. A character interval may be detected.

  Furthermore, the characters include characters having a plurality of components separated along the horizontal direction, such as kanji characters composed of bias and 旁. If such a character is included in the character string, the character section may be detected for each individual component of the character.

  FIG. 3 is a diagram illustrating an example of a character string shown in an image and a character section detected for the character string. A graph 310 representing the projection value of each vertical line with respect to the character string 300 “start of athletic meet” shown in FIG. 3 is shown below the character string 300. Each dotted line 320 represents a break between detected characters. For example, it can be seen that two character sections 331 and 332 are set for the character “motion”. Similarly, two character sections are set for the characters “NO”, “START”, and “RI”, respectively.

  Therefore, the character section detection unit 21 extracts a character section that may correspond to a part of one character among all the detected character sections. For this purpose, the character section detection unit 21 calculates the average value or mode value of the widths of all detected character sections in the horizontal direction as the reference character width. The character section detection unit 21 is a combination of a plurality of sets of consecutive character sections in which a set in which the total horizontal width of each character section included in the set is less than the reference character width is connected. A character interval.

  The candidate character lattice generation unit 22 generates a candidate character lattice that is a set of paths that are set for each of a plurality of character segments and that represents a relative positional relationship with other character segments. Therefore, the candidate character lattice generation unit 22 generates candidate character lattices by comparing the left end coordinates and the right end coordinates of the detected character sections and arranging them in order from the left end side of the image. .

  FIG. 4 is a diagram showing an example of a candidate character lattice corresponding to the character string 300 shown in FIG. Candidate character lattice 400 has 16 paths. Each path is represented by an arrow, and an identification number of the path, that is, an identification number of a character section corresponding to the path is shown on the path. In addition, characters corresponding to individual paths and some components of the characters are shown above or below the paths. For example, path 1 corresponds to the character “luck”. Paths 2 and 3 respectively correspond to the deviation and wrinkle of the character “motion”. On the other hand, the path 20 obtained by integrating the path 2 and the path 3 corresponds to the character “motion”.

  The candidate character extraction unit 23 extracts at least one candidate character that is a candidate for a character shown in the character section corresponding to each path. For example, the candidate character extraction unit 23 cuts out a part of an image corresponding to each character section from the image as a character image, and binarizes the character image. The threshold value for binarizing the character image can be, for example, the average pixel value of the character image. Then, the candidate character extraction unit 23 obtains a recognition distance between the character image and the template by executing template matching between the binarized character image and a template representing the shape of each character. For example, the template may be a binary image in which the value of the pixel in which the stroke of the corresponding character is reflected is “255” and the value of the other pixel is “0”. Each template is stored in the storage unit 13 in advance. A plurality of templates with different typefaces or sizes may be prepared for one character. The recognition distance can be, for example, a Hamming distance between the binarized character image and the template. In this case, the smaller the recognition distance, the higher the possibility that the character represented in the template is reflected in the character image. Alternatively, the candidate character extraction unit 23 may obtain a cross-correlation value between the character image and the template instead of the recognition distance. In this case, the higher the overall correlation value, the higher the possibility that the character represented in the template is reflected in the character image.

The candidate character extraction unit 23 extracts a predetermined number of templates in order from the smallest recognition distance for each path, and sets characters corresponding to the extracted templates as candidate characters of the path. Then, the candidate character extraction unit 23 ranks the extracted candidate characters in order from the shorter recognition distance. The predetermined number is set to any one of 1 to 5, for example. Alternatively, the candidate character extraction unit 23 may extract all templates whose recognition distance is equal to or less than a predetermined distance, and may use characters corresponding to the extracted template as candidate characters for the path. The predetermined distance can be, for example, a recognition distance when 70% to 80% of the pixels included in the template match the pixels of the character image. In this case, the candidate character extraction unit 23 preferably uses the character corresponding to the template with the smallest recognition distance as the candidate character of the path for the path without the template with the recognition distance equal to or smaller than the predetermined distance.
Note that the candidate character extraction unit 23 may obtain a candidate character for each pass by using another known technique for recognizing characters appearing on the image. However, in this case as well, it is preferable that each candidate character is ranked in order from the highest possibility that it appears in the character section corresponding to the path.

Referring again to FIG. 4, three candidate characters are shown for each pass. For example, three candidate characters “luck”, “ream”, and “achi” are shown for path 1.
The candidate character extraction unit 23 stores the code representing the candidate character extracted for each path in the storage unit 13 in association with the path together with the rank corresponding to the candidate character.

  Referring again to FIG. 4, the candidate character lattice 400 includes a plurality of consecutive arrays of paths corresponding to the entire character string 300. The paths that overlap at least partially are exclusive, and are not included in the same path array. For example, if path 2 and path 3 are selected as part of a path array corresponding to the entire character string 300, the path 20 is not included in the path array. Conversely, when the path 20 is selected as a part of the path array corresponding to the entire character string 300, the path 2 and the path 3 are not included in the path array.

  FIG. 5 is a diagram illustrating an example of a candidate character lattice for another character string. In the example illustrated in FIG. 5, a break between five characters x2 to x6 is detected for the character string 500 of “recognition”. X1 and x7 correspond to the left end and the right end of the character string 500, respectively. In this example, the boundary between each character “K” and “K” is detected as a break between characters. Therefore, even though the character string 500 includes only two characters, the candidate character lattice 510 corresponding to the character string 500 includes eight paths A1 to A8. Therefore, there are a large number of exclusive path pairs (for example, paths A1, A4 and A6, paths A6 and A7, etc.). As a result, there are also a large number of consecutive path arrays corresponding to the entire character string 500.

  As shown in FIGS. 4 and 5, as the number of exclusive path sets increases, the number of path arrays corresponding to the entire character string also increases. In order to prevent a search for a word contained in a character string from failing, a word is searched for each path array. Therefore, as the number of path arrays increases, the calculation required for the word search is increased. The amount also increases.

Therefore, the candidate character lattice correction unit 24 corrects the paths included in the exclusive path set so as to reduce the number of consecutive paths in a row corresponding to the entire character string shown in the image. Modify the candidate character lattice. In the present embodiment, the candidate character lattice correction unit 24 corrects the candidate character lattice so as to satisfy the following conditions.
(1) For a plurality of paths that are not mutually exclusive, that is, a set of paths that can be included in the same array of consecutive paths corresponding to the entire character string, the candidate character lattice correcting unit 24 Maintain the order of multiple paths and keep them non-exclusive.
(2) For a plurality of mutually exclusive paths, the candidate character lattice correcting unit 24 corrects and integrates the longer path with the shorter path. However, if the condition (1) is not satisfied by the integration, the candidate character lattice correction unit 24 corrects the longer path so as to be adjacent to the path exclusive to the path.
(3) If there are a plurality of relatively short paths exclusive to the longer path, the candidate character lattice correction unit 24 determines the longer path as any of the plurality of relatively short paths. You may integrate. Alternatively, the candidate character lattice correction unit 24 may duplicate the longer path and integrate it with each of a plurality of relatively short paths.
(4) When the candidate character lattice correcting unit 24 integrates a plurality of paths, the candidate character lattice correcting unit 24 associates the character section and the candidate character corresponding to each path with the integrated path.

  FIG. 6 is an operation flowchart of an example of candidate character lattice correction processing for correcting the candidate character lattice while satisfying each of the above conditions, which is executed by the candidate character lattice correcting unit 24. In the following description, the candidate character lattice 500 shown in FIG. 5 is referred to as appropriate for easy understanding.

  Candidate character lattice correction unit 24 classifies paths that do not overlap with other paths corresponding to character sections shorter than the character sections corresponding to the candidate character lattice correction unit α among the plurality of paths to the confirmed group α (step S101). The candidate character lattice correction unit 24 classifies the paths that have not been classified into the confirmed group α into the unconfirmed group β (step S102).

  Referring to FIG. 5 again, each path included in the candidate character lattice 500 is assigned identification numbers A1 to A8 in order from the shorter character section corresponding to each path. The paths A1 to A4 do not overlap with other paths corresponding to a character section shorter than its own character section. Therefore, the paths A1 to A4 are classified into the confirmed group α. On the other hand, the paths A5 to A8 are at least partially overlapped with other paths corresponding to a character section shorter than its own character section. For example, the character section of path A5 overlaps with the character sections of paths A2 and A3, and the character section of path A6 overlaps with the character sections of paths A1 and A4. Therefore, the paths A5 to A8 are classified into the indeterminate group β.

なお、表１には、各修正パスp[k]について、元の候補文字ラティスにおけるパスの識別番号、パスに対応する文字区間（すなわち、その文字区間の左端及び右端の座標）、そのパスについて抽出された候補文字及び他の候補パス群が対応付けられる。例えば、修正パスp[1]には、識別番号A1、パスA1の文字区間[x1,x2]、候補文字「言」が対応付けられている。なお、候補文字が複数ある場合には、すべての候補文字が対応付けられる。例えば、パスA1に対して３個の候補文字「言」、「吉」、「忘」が抽出されている場合、これらの候補文字すべてが修正パスp[1]に対応付けられる。
また、この時点では、修正パス列Pに含まれる修正パスp[k]と未確定グループに属するパスとの位置関係は調べられていないので、各修正パスについての他の候補パス群は空集合となっている。 Next, the candidate character lattice correcting unit 24 sorts the paths in the confirmed group α in the coordinate order from the beginning to the end of the character string, and corrects the corrected path string P = {p [1], p [2],. ..} is generated. Further, the candidate character lattice correcting unit 24 corresponds to each corrected path p [k] in the corrected path sequence P, that is, another candidate path group that is a set of paths exclusive to the corrected path p [k]. {Qk} is initialized as an empty set (step S103). For example, for the candidate character lattice 500, the corrected path sequence P includes paths A1 to A4 in the order of A1, A4, A3, and A2, as shown in Table 1 below.

In Table 1, for each corrected path p [k], the path identification number in the original candidate character lattice, the character section corresponding to the path (that is, the coordinates of the left and right ends of the character section), and the path The extracted candidate characters are associated with other candidate path groups. For example, the corrected path p [1] is associated with the identification number A1, the character section [x1, x2] of the path A1, and the candidate character “word”. When there are a plurality of candidate characters, all candidate characters are associated. For example, when three candidate characters “word”, “kichi”, and “forget” are extracted for the path A1, all these candidate characters are associated with the corrected path p [1].
At this time, since the positional relationship between the corrected path p [k] included in the corrected path sequence P and the path belonging to the indeterminate group has not been examined, the other candidate path groups for each corrected path are empty sets. It has become.

  The candidate character lattice correction unit 24 selects any of the paths belonging to the unconfirmed group β as the attention path A (step S104). Note that the candidate character lattice correction unit 24 selects, for example, the attention path A in order from the path having the shortest corresponding character section among the paths belonging to the unconfirmed group β. Alternatively, the candidate character lattice correcting unit 24 may select, for example, the attention path A in order from the path closest to the head of the character string among the paths belonging to the unconfirmed group β. Then, the candidate character lattice correction unit 24 selects, from among the correction paths p [k] included in the correction path sequence P, any one of the correction paths p [j] that at least partially overlaps the target path A (step S105). . The candidate character lattice correction unit 24 determines whether there is a path that does not overlap the target path A among the other candidate paths corresponding to the correction path p [j] (step S106). If there is no other candidate path corresponding to the corrected path p [j], or if there is no path that does not overlap with the target path A among the other candidate paths (step S106-No), the candidate character lattice correcting unit 24 , P [j], path A is added to another candidate path group {Qk} (step S107).

  For example, if the target path A is the path A5, the correction path p [3], that is, the path A3 or the correction path p [4], that is, the path A2 is selected. For example, if the correction path p [3] is selected, there is no other candidate path corresponding to the correction path p [3] at this time. Accordingly, the path A5 is added to another candidate path group of the corrected path p [3].

  After step S107, or when there is another candidate path corresponding to the corrected path p [j] and there is a path that does not overlap with the target path A among the other candidate paths (step S106-Yes), the candidate The character lattice correction unit 24 determines whether there is another correction path that overlaps the target path A (step S108). If there is another correction path that overlaps the target path A (step S108—Yes), the candidate character lattice correction unit 24 repeats the processing from step S105 onward.

  For example, when the correction path p [3] is first selected for the path A5 as the attention path, the candidate character lattice correction unit 24 next selects the correction path p [4], that is, the path A2. . At this time, there is no other candidate path corresponding to the corrected path p [4]. Accordingly, the path A5 is added to another candidate path group of the corrected path p [4].

  On the other hand, if there is no other correction path that overlaps the attention path A (step S108-No), the candidate character lattice correction unit 24 determines that the attention path A is one of the other candidate paths {Qj] } Is determined (step S109). If the target path A has been added to another candidate path group {Qj} of any correction path p [j] (step S109—Yes), the candidate character lattice correction unit 24 has not yet entered the unconfirmed group β. It is determined whether or not the target path remains (step S111). If an unfocused path remains (step S111—Yes), the candidate character lattice correction unit 24 repeats the processing after step S104.

図５及び表２に示されるように、パスA6、パスA7は、修正パスp[1]、p[2]に対して、他の候補パスが無いか、あるいは他の候補パス群内にパスA6、A7と重ならないパスは無いので、修正パスp[1]、p[2]の他の候補群｛Q1｝、｛Q2｝にそれぞれ追加されている。 For example, it is assumed that after the path A5, the paths A6 and A7 are sequentially selected as the target path, and the processes in steps S104 to S111 are performed. The correspondence table of the corrected path sequence P at this time is as follows.

As shown in FIG. 5 and Table 2, the path A6 and the path A7 have no other candidate paths with respect to the corrected paths p [1] and p [2], or are paths within other candidate path groups. Since there is no path that does not overlap with A6 and A7, they are added to the other candidate groups {Q1} and {Q2}, respectively, of the corrected paths p [1] and p [2].

  On the other hand, if the attention path A is not added to any other candidate path group {Qj} in any correction path p [j] in step S109 (step S109-No), the candidate character lattice correction unit 24 For the correction path sequence P, a path A is added as a new correction path between a plurality of correction paths overlapping the path A (step S110).

  For example, referring to FIG. 5 and Table 2, if the last unfocused path A8 is set as the focused path, the path A8 is the modified path p [2], that is, the path A4 and the modified path p [3], That is, it overlaps with the path A3. However, another candidate group {Q2} of the corrected path p [2] includes a path A7 that does not overlap the path A8. Similarly, the other candidate group {Q3} of the corrected path p [3] includes a path A5 that does not overlap with the path A8. Therefore, the path A8 is not added to any other candidate group of any correction path p [j]. A boundary between the character section corresponding to the correction path p [2] and the character section corresponding to the correction path p [3] is located in the middle of the character section of the path A8. Therefore, the path A8 is added between the correction path p [2] and the correction path p [3].

After step S110, the candidate character lattice correction unit 24 determines whether or not an unfocused path remains in the unconfirmed group β (step S111). If an unfocused path remains (step S111—Yes), the candidate character lattice correction unit 24 repeats the processing after step S104.
On the other hand, if there is no unfocused path remaining (step S111-No), the candidate character lattice correcting unit 24 sets all candidate characters for the other candidate path group {Qk} as candidate characters for the corrected path p [k]. It adds (step S112). Then, the correction path string P becomes a correction candidate character lattice. In this case, the modified path p [k] in which the other candidate path group {Qk} is not an empty set is included in the path in the fixed group α and the other candidate path group {Qk} corresponding to the modified path p [k]. Which will replace the path
Thereafter, the candidate character lattice correction unit 24 ends the candidate character lattice correction process.

修正候補文字ラティスの各パスには、そのパスに関連付けられた元の候補文字ラティスのパスの識別番号及びそのパスの候補文字がその順位とともに関連付けられる。表３に示されるように、修正候補文字ラティスに含まれる統合されたパスについての候補文字の数は元の個々のパスについての候補文字の数よりも増えるものの、修正候補文字ラティスは互いに排他的なパスの組を含んでいない。そのため、修正候補文字ラティスは、元の候補文字ラティスよりも簡単な構造を有している。 Table 3 is a list of paths and candidate characters of the correction candidate character lattice finally created.

Each path of the correction candidate character lattice is associated with the path identification number of the original candidate character lattice associated with the path and the candidate character of the path together with its rank. As shown in Table 3, although the number of candidate characters for the integrated path included in the correction candidate character lattice is greater than the number of candidate characters for the original individual path, the correction candidate character lattices are mutually exclusive. Does not contain a complete set of paths. Therefore, the correction candidate character lattice has a simpler structure than the original candidate character lattice.

  FIG. 7 is an operation flowchart of another example of candidate character lattice correction processing for correcting the candidate character lattice while satisfying each of the above conditions, which is executed by the candidate character lattice correction unit 24. Also in the following description, the candidate character lattice 500 shown in FIG. 5 is referred to as appropriate for easy understanding.

表４において、一番上の行は修正パスの先頭からの順序を表す。二番目の行は各修正パスに対応する区間の左端と右端の水平座標を表す。３番目の行は、修正パスに割り当てられた元の候補文字ラティスのパスの識別番号を表す。そして一番下の行は、修正パスに割り当てられたパスについての候補文字の集合を表す。この時点では、何れの修正パスにもパスは割り当てられていないので、下の２行は空欄となっている。 The candidate character lattice correction unit 24 extracts all the boundaries of the paths included in the candidate character lattice in order from the top (step S201). Then, the candidate character lattice correction unit 24 sets a path storage area including (number of path boundaries-1) correction paths in the storage unit 13 (step S202).
For example, in candidate character lattice 500, since there are seven boundaries x1 to x7 in order from the top, a path storage area including six correction paths is set as shown in Table 4 below.

In Table 4, the top line represents the order from the beginning of the correction path. The second row represents the horizontal coordinates of the left and right ends of the section corresponding to each correction path. The third row represents the path identification number of the original candidate character lattice assigned to the modified path. The bottom line represents a set of candidate characters for the path assigned to the correction path. At this time, since no path is assigned to any of the correction paths, the lower two lines are blank.

なお、表５では、簡単化のために、各パスについて一つの候補文字だけが示されている。 The candidate character lattice correction unit 24 assigns each path to a correction path having the largest overlap width between the character section corresponding to the path and the section of the correction path (step S203).
For example, x1 = 10, x2 = 25, x3 = 30, and x4 = 40. In this case, the path A6 overlaps with the correction paths M1 [x1, x2], M2 [x2, x3], and M3 [x3, x4], but the overlap width with the correction path M1 is the largest. Therefore, the candidate character lattice correction unit 24 assigns the path A6 to the correction path M1.
Table 5 shows path storage areas when all paths are assigned to any of the correction paths.

In Table 5, for simplification, only one candidate character is shown for each path.

After all paths are assigned to any of the correction paths, the candidate character lattice correction unit 24 sequentially extracts the correction paths to which one or more paths are assigned from the top, and sets the paths as the correction candidate character lattice (step). S204). Then, the candidate character lattice correction unit 24 sets the candidate character of the path assigned to the extracted correction path as the candidate character of the path of the correction candidate character lattice (step S205). In this example as well, a correction path to which a plurality of paths are assigned integrates the plurality of paths and replaces the plurality of paths.
Thereafter, the candidate character lattice correction unit 24 ends the candidate character lattice correction process.

For example, as shown in Table 5, one or more paths are assigned to the correction paths M1, M3, M5, and M6. Therefore, the correction paths M1, M3, M5, and M6 are extracted as paths for the correction candidate character lattice. Table 6 is a list of correction candidate character lattice paths and candidate characters that are finally obtained.

Note that, according to the modification, in step S203, the candidate character lattice correction unit 24 may assign each path to all the correction paths that at least partially overlap the path. In this case, for example, the path A6 is assigned to each of the correction paths M1 to M3. Table 7 shows a path storage area when all paths are assigned to any of the correction paths according to this modification.

  As described above, since any candidate character lattice correction process is executed, the correction candidate character lattice does not include an exclusive path. Therefore, there is a single array of paths that are continuous in a line corresponding to the entire character string. Become.

The word search unit 25 matches at least a part of a combination of candidate characters included in the array of at least a part of the correction candidate character lattice among a plurality of words registered in the word dictionary stored in the storage unit 13. Detect the word you want.
In the present embodiment, the word search unit 25 uses a matching method (hereinafter referred to as DP matching) based on dynamic programming in order to search for words registered in the word dictionary.

  FIG. 8 is an explanatory diagram of an optimum path search procedure using DP matching used in the present embodiment. In FIG. 8, the number obtained by adding 1 to the number of characters included in the search target word is the number of lattices on the vertical axis, and the number obtained by adding 1 to the number of paths included in at least a part of the correction candidate character lattice is the number of lattices on the horizontal axis. Grid points are arranged in a grid pattern. In this example, the search target word 810 is “athletic meet” and includes three characters. Each character is associated with a grid point in one row in order from the top in order from the top. On the other hand, the number of paths of the correction candidate character lattice is four, and the correction candidate character lattice is associated with lattice points in one column in order from the left side from the first path to the last path. Each path is associated with one or more candidate characters 811 to 814. For example, three candidate characters “luck”, “ream”, and “achi” are associated with the first path. A character string obtained by concatenating one character selected from the candidate characters of each path in order from the first path is referred to as an input character string here.

  The word search unit 25 searches for an optimum route among the routes from the upper left lattice point to the lower right lattice point. As the evaluation value, for example, the edit distance between the search target word and the input character string, or the number of matching characters between the search target word and the input character string is used. When the edit distance is used as the evaluation value, the word search unit 25 sets the route having the smallest evaluation value as the shortest route. In this case, the grid point at the upper left corner is the first focused grid point. Then, the word search unit 25 sets one of the lattice points adjacent to the right side, the lower side, and the lower right side of the target lattice point as the next target lattice point. At that time, when a transition is made from the current target lattice point to a target lattice point adjacent to the lower right, the character of the search target word corresponding to the line to which the next target lattice point belongs (for example, the third row from the top For example, it is determined whether “motion”) matches any of the candidate characters corresponding to the column to which the target lattice point belongs. If they match, the point added to the edit distance is “0”. However, if they do not match, the point added to the edit distance is “+2”. The fact that the two do not match corresponds to the replacement of the character corresponding to the next target lattice point of the search target word. Further, the addition point corresponding to character replacement may be set to “+1”.

Further, when the lattice point adjacent to the lower side of the current target lattice point becomes the next target lattice point, the point added to the edit distance is “+1”. Note that this downward transition corresponds to a lack of characters in the search target word. Furthermore, even when the grid point adjacent to the right side with respect to the current target grid point becomes the next target grid point, the point added to the edit distance is “+1”. The transition to the right corresponds to the insertion of a character into the search target word.
In the example shown in FIG. 8, the route 820 indicated by the arrow is the shortest route, and “Exercise ☆ kai” is selected as the input character string corresponding to the search target word. The ☆ mark represents the inserted character, and here is one of the candidate characters “K”, “Power”, and “Sword” corresponding to the second path from the right.

The word search unit 25 corresponds to the input character string in the character string represented by the correction candidate character lattice when the character reproduction rate and the character matching rate between the search target word and the input character string exceed a predetermined threshold. It is determined that the search target word is detected at the position.
The character reproduction rate and the character matching rate are expressed by the following equations, for example.
Character recall: (Number of characters that match between the input string and the search word)
/ Number of characters in the search target word Character precision = (Number of matching characters between the input string and the search target word)
/ Number of characters included in input character string The predetermined threshold value is set to any value within the range of 0.6 to 0.8, for example.

  Alternatively, the word search unit 25 may determine that the search target word has been extracted at a position corresponding to the input character string corresponding to the shortest path when the edit distance for the shortest path is equal to or less than a predetermined value. In this case, for example, the predetermined value may be a value obtained by multiplying the number of characters included in the search target word by any value within the range of 0.3 to 0.4.

  In addition, the word search unit 25 uses other fuzzy search methods to search for a word that at least partially matches a combination of candidate characters included in at least a part of the sequence included in the correction candidate character lattice. Also good. The word search unit 25 can use, for example, a method disclosed in Japanese Patent Laid-Open No. 2010-225137 as such an ambiguous search method. Or, the word search unit 25 was described in Kita, “VLDC pattern matching algorithm that allowed error”, IEICE technical report, COMP, Computing, 103 (622), p61-68, 2004. The method can be used.

The word search unit 25 performs the above-described processing on all words registered in the word dictionary and searches for words included in the character string. Therefore, for example, if a large number of words are registered in the word dictionary, for example, if 100,000 words are registered, the above-described processing is performed on all the words.
However, as described above, the word search unit 25 performs an ambiguous search for the correction candidate character lattice that does not include an exclusive path for each word, so that the original candidate character lattice having a plurality of combinations of paths exists. Compared to the fuzzy search, the amount of calculation can be greatly reduced. Therefore, as the number of words registered in the word dictionary increases, the effect of reducing the amount of calculation increases. In addition, since the word search unit 25 searches for a word by allowing any of character replacement, insertion, and omission, there is a misrecognition of a character for each path, or a detected character section is erroneous. Even if there is, the word search unit 25 can correctly search for a word included in the character string.

  The word search unit 25 adds a word path that is a path corresponding to the detected word to the candidate character lattice. The obtained lattice is hereinafter referred to as a candidate character / word lattice for convenience. Since the word search unit 25 knows the path targeted for DP matching when detecting the word, the word search unit 25 can specify the position to add the word path based on the position of the path. Further, for each detected word path, the word search unit 25 determines the rank of candidate characters that match each character included in the corresponding word and the identification number of the path of the original candidate character lattice corresponding to the candidate character. Associate with that word path.

  FIG. 9 is a diagram illustrating an example of a candidate character / word lattice obtained by adding a detected word path to the character string “start of athletic meet” illustrated in FIG. 3. In the candidate character / word lattice 900 shown in FIG. 9, among the paths represented by arrows, paths 1 to 23 are paths for each character, and paths 31 to 37 are added word paths.

The estimation unit 26 estimates a character string shown on the image based on the candidate character / word lattice. Therefore, for each word path included in the candidate character / word lattice, the estimation unit 26 calculates an evaluation value of a word corresponding to the word path, for example, according to the following equation.
Word evaluation value = {Σ (1-α (M _i -1))} ⁿ
Here, M _i represents the rank of candidate characters determined to match the i-th character from the beginning in the word detected by the word search unit 25. However, if there is no candidate character determined to match the i-th character, M _i is set to a value obtained by adding 1 to the maximum number of candidate characters that the candidate character extraction unit 23 extracts for each path. The α is a coefficient, and for example, the maximum number of candidate characters that the candidate character extraction unit 23 extracts for each path so that (1-α (M _i −1)) approaches 0 as the value of M _i increases. Is set to a positive value equal to or less than the reciprocal of, for example, 0.1. Therefore, the smaller the sum of the ranks of candidate characters that match individual characters included in the word corresponding to the word path, the higher the word evaluation value. N is a correction coefficient, and is set to a value of 1 or more, for example. In particular, in order to increase the word evaluation value as the number of characters included in the word increases, the correction coefficient n is preferably set to a value larger than 1, for example, set to 2. By setting the correction coefficient in this way, a word with a larger number of characters can be more easily estimated to be a word included in the character string.

Referring to FIG. 9 again, for the word “athletic meeting” corresponding to the word path 31, M ₁ = M ₂ = 1 and M ₃ = 2. Therefore, if α = 0.1 and n = 2, the word evaluation value is 8.41. In FIG. 9, the numerical value written along with the word represented by the word paths 31 to 37 is the word evaluation value for the word.

  In addition, the estimation unit 26 sets a word evaluation value for a path representing one character that is not a word path to a constant β. The constant β is set to, for example, a value that is larger than 0 and smaller than the maximum value that can be taken by a word evaluation value for a word including two characters, for example, 0.5.

In the candidate character / word lattice, the estimation unit 26 obtains a sum of word evaluation values as a concatenated path evaluation value for each word path and each path array corresponding to the entire character string. Then, the estimation unit 26 shows on the image a word path and a string of words and characters corresponding to the word path and paths arranged in accordance with the order of the paths included in the array having the maximum connection path evaluation value. Inferred as a string. When there are a plurality of exclusive paths having the same sum of word evaluation values as in paths 5 and 6 and path 21 in FIG. 9, the estimation unit 26 recognizes the recognition distance obtained by the candidate character extraction unit 23. Choose the path with the smaller sum of to be included in the array. When there are a plurality of candidate characters for the path included in the array having the maximum word evaluation value, the estimation unit 26 selects the candidate character with the highest ranking.
In the example of FIG. 9, since the connected path evaluation value of the array selected according to the order of path 31 → pass 21 → pass 34 is maximized, the estimation unit 26 has arranged words and candidate characters corresponding to these paths. The character string “start of athletic meet” is estimated as a character string reflected on the image.

FIG. 10 is an operation flowchart of the character recognition process. The processing unit 15 executes this character recognition process every time an image is received.
The character section detection unit 21 detects a character section in which the character is estimated to be captured for each individual character from the character string captured on the image (step S301). Furthermore, the character section detection unit 21 detects a section obtained by connecting two or more consecutive character sections that may correspond to one character as a character section (step S302). Thereafter, the candidate character lattice generation unit 22 generates a candidate character lattice that is a set of paths that are set for each of the plurality of character segments and that represents a relative positional relationship with other character segments (step S303).
The candidate character extraction unit 23 extracts candidate characters for each pass (step S304).

  Thereafter, the candidate character lattice correcting unit 24 generates a correction candidate character lattice by integrating mutually exclusive paths or correcting one position of the exclusive path (step S305). Specifically, the candidate character lattice correction unit 24 performs a first pass and a second pass among two or more consecutive passes that overlap at least partly with the first pass. Replacement is performed with a third path that includes candidate characters for the second path and that corresponds to the same character section as the second path. Alternatively, the candidate character lattice correction unit 24 corrects the first pass so as to be inserted between the consecutive passes.

Thereafter, the word search unit 25 selects a word that at least partially matches a combination of candidate characters included in each of two or more consecutive paths in the correction candidate character lattice from among a plurality of words registered in the word dictionary. Detection is performed (step S306). The word search unit 25 adds a word path corresponding to the detected word to the candidate character lattice to generate a candidate character / word lattice (step S307).
The estimation unit 26 estimates that a character string corresponding to an array of paths that maximizes the sum of the word evaluation values obtained for each path is a character string appearing on the image (step S308). Then, the processing unit 15 converts the character strings included in the estimated character string into electronic data by arranging the codes of the characters included in the estimated character string according to the estimated order of the character strings. The processing unit 15 outputs electronic data representing a character string to another device via the output unit 12. Alternatively, the processing unit 15 stores electronic data representing a character string in the storage unit 13. Thereafter, the processing unit 15 ends the character recognition process. In addition, the process part 15 may replace the order of the process of step S303 and step S304.

  As described above, this character recognition device corrects the structure of the candidate character lattice to eliminate mutually exclusive sets of paths before searching for words contained in the character string. The number of path arrays representing the character string has been reduced. Therefore, since this character recognition device can reduce the amount of calculation at the time of word search, the amount of calculation as the whole character recognition processing can also be reduced. Furthermore, this character recognition device searches for a word that may be included in the character string by an ambiguous search method that allows replacement, omission, or insertion of characters in the word. For this reason, the character recognition device can detect words included in the character string even if there is an error in the recognition result of each character included in the character string on the image, so that the character string can be recognized accurately.

The present invention is not limited to the above embodiment. For example, the estimation unit may increase the word evaluation value as the average value of the recognition distances for candidate characters that match each character included in the word corresponding to the word path is small.
Further, the character section detection unit may detect a plurality of character sections by changing a threshold for the projection value and detecting a character break. For example, the character section detection unit first detects a plurality of character sections by detecting a character break using a threshold value similar to that in the above embodiment. Thereafter, the character section detection unit corrects the threshold value to a lower value, for example, a value obtained by dividing the original threshold value by 1.2 to 1.5, and then compares the projection value of each horizontal position with the corrected threshold value again to correct the threshold value. A horizontal position having a projection value less than or equal to the later threshold is redetected as a character break. The character section detection unit sets a section between the redetected character breaks as a character section. This makes it easy to detect a character section including the entire character even for characters that are close to each other.

  All examples and specific terms listed herein are intended for instructional purposes to help the reader understand the concepts contributed by the inventor to the present invention and the promotion of the technology. It should be construed that it is not limited to the construction of any example herein, such specific examples and conditions, with respect to showing the superiority and inferiority of the present invention. Although embodiments of the present invention have been described in detail, it should be understood that various changes, substitutions and modifications can be made thereto without departing from the spirit and scope of the present invention.

DESCRIPTION OF SYMBOLS 1 Character recognition apparatus 11 Image acquisition part 12 Output part 13 Storage part 14 Storage medium access apparatus 15 Processing part 16 Storage medium 21 Character area detection part 22 Candidate character lattice generation part 23 Candidate character extraction part 24 Candidate character lattice correction part 25 Word search Part 26 Estimator

Claims (7)

From an image obtained by photographing a character string including a plurality of characters represented on the medium, for each of the plurality of characters, detecting a character section on the image that is estimated to include the character,
A candidate character lattice that is a set of paths representing a relative positional relationship with other character sections set for each of the plurality of character sections is obtained,
For each of the plurality of paths, obtain at least one candidate character that is a candidate for a character in the character section corresponding to the path,
Of the plurality of paths included in the candidate character lattice, a first path and a second path of two or more consecutive paths at least partially overlapping the first path, A first pass candidate character and a second pass candidate character are included and replaced with a third pass corresponding to the same character section as the second pass, or the first pass is Modifying the candidate character lattice to be inserted between the two or more consecutive passes;
Among the plurality of words registered in the word dictionary, the character string includes a word that at least partially matches a combination of candidate characters included in each of two or more consecutive paths in the modified candidate character lattice. A candidate character and a word lattice obtained by detecting a word as a potential word and adding a word path representing the position of the detected word in the character string to the candidate character lattice;
In the candidate character and the word lattice, for each word path, an evaluation value representing the probability that a word corresponding to the word path is included in the character string is obtained.
A total value of the evaluation values is obtained for each of the word path included in the candidate character and the word lattice and the word path and the array of the paths continuous in a line corresponding to the entire character string selected from the paths. The word path included in the array having the highest total value and the word path arranged according to the order of the paths and a combination of a word and a candidate character corresponding to the path are estimated as the character string.
A computer program for character recognition that causes a computer to execute the above.   Modifying the candidate character lattice means that, among the two or more consecutive paths, a path corresponding to a character section having a maximum width overlapping with the character section corresponding to the first path is defined as the second path. The computer program for character recognition according to claim 1.   The character section is obtained when the width of a correction section obtained by connecting the two or more character sections among two or more consecutive character sections corresponds to the width of one character. The computer program for character recognition according to claim 1, comprising adding a section as a character section. Obtaining the candidate character is to obtain a plurality of candidate characters for at least one of the plurality of paths, and for each of the plurality of candidate characters obtained for the path, in the character section corresponding to the path. Set the order in descending order of the probability of being reflected,
4. The evaluation value according to claim 1, wherein obtaining the evaluation value increases the evaluation value as a total of ranks of the candidate characters that match characters included in the word corresponding to the word path is smaller. The computer program for character recognition described.   The computer program for character recognition according to any one of claims 1 to 4, wherein obtaining the evaluation value increases the evaluation value as the number of characters included in the word corresponding to the word path increases. An image acquisition unit that acquires an image of a character string including a plurality of characters represented on the medium;
A storage unit for storing a word dictionary in which a plurality of words are registered;
From the image, for each of the plurality of characters, a character section detection unit that detects a character section on the image that is estimated that the character is reflected,
A candidate character lattice generation unit for obtaining a candidate character lattice that is a set of paths representing a relative positional relationship with other character intervals set for each of the plurality of character segments;
A candidate character extraction unit that obtains at least one candidate character that is a candidate for a character in the character section corresponding to the path for each of the plurality of paths;
Of the plurality of paths included in the candidate character lattice, a first path and a second path of two or more consecutive paths at least partially overlapping the first path, A first pass candidate character and a second pass candidate character are included and replaced with a third pass corresponding to the same character section as the second pass, or the first pass is A candidate character lattice correction unit for correcting the candidate character lattice to be inserted between the two or more consecutive paths;
Among the plurality of words registered in the word dictionary, the character string includes a word that at least partially matches a combination of candidate characters included in each of two or more consecutive paths in the modified candidate character lattice. A word search unit for detecting candidate words and word lattices, which are detected as possible words, and a word path representing the position of the detected word in the character string is added to the candidate character lattice;
In the candidate character and word lattice, for each word path, an evaluation value representing the probability that a word corresponding to the word path is included in the character string is obtained, and the word path included in the candidate character and word lattice and The word path included in the array having the highest total value is obtained by calculating a total value of the evaluation values for each of the word path and the array of the paths that are continuous in a line corresponding to the entire character string selected from the paths. And an estimation unit that estimates the word path arranged according to the order of the paths and a combination of a word corresponding to the path and a candidate character as the character string,
A character recognition device. Obtain an image of a character string that contains multiple characters represented on the medium,
From the image, for each of the plurality of characters, detect a character section on the image that is estimated that the character is reflected,
A candidate character lattice that is a set of paths representing a relative positional relationship with other character sections set for each of the plurality of character sections is obtained,
For each of the plurality of paths, obtain at least one candidate character that is a candidate for a character in the character section corresponding to the path,
Of the plurality of paths included in the candidate character lattice, a first path and a second path of two or more consecutive paths at least partially overlapping the first path, A first pass candidate character and a second pass candidate character are included and replaced with a third pass corresponding to the same character section as the second pass, or the first pass is Modifying the candidate character lattice to be inserted between the two or more consecutive passes;
Among the plurality of words registered in the word dictionary, the character string includes a word that at least partially matches a combination of candidate characters included in each of two or more consecutive paths in the modified candidate character lattice. A candidate character and a word lattice obtained by detecting a word as a potential word and adding a word path representing the position of the detected word in the character string to the candidate character lattice;
In the candidate character and the word lattice, for each word path, an evaluation value representing the probability that a word corresponding to the word path is included in the character string is obtained.
Obtaining the total value of the evaluation values for each word path and the array of the paths that are continuous in a row corresponding to the entire character string selected from the word path and the path included in the candidate character and word lattice; The word path included in the array having the highest total value and the word path aligned according to the order of the paths and a combination of a word and a candidate character corresponding to the path are estimated as the character string.
A method involving that.

Images