JP3548234B2 - Character recognition method and device - Google Patents

Description

[0001]
[Industrial applications]
The present invention relates to a character segmentation method for extracting an image block in character units from a document image and performing character recognition in block units, and a character recognition method and apparatus using the same.
[0002]
[Prior art]
Generally, in an optical character recognition device (OCR), an image block for each character is extracted from a document image optically read, and characters are cut out, and character recognition is performed for each cut out character. An example of a recognition process performed in a general OCR will be described with reference to FIGS.
[0003]
FIG. 8 is a flowchart showing a procedure of a general OCR character recognition process. First, in step S701, a document image is input using an image scanner or the like. Next, in step S702, image blocks are extracted from the input document image.
[0004]
FIG. 9 is a diagram illustrating a method of extracting an image block used in general character recognition processing. FIG. 9 shows an example of extracting image blocks of a vertically written document using projection. First, rows are extracted by taking projections 801 in the vertical direction. Then, after extracting the rows, the image block 803 can be extracted by taking a projection 802 for each row in a direction perpendicular to the row direction. In this state, there is a block 804 including a plurality of characters due to two or more characters touching, or a separated block 805 for a character having a separated portion.
[0005]
Therefore, in step S703, an image block (character image block) for each character is created. First, a standard character height is obtained by averaging the height of each image block in each line, extracting the most frequent height from the height distribution of the image blocks, and the like. Then, if the height of the image block is an integral multiple of the standard character height, the height is equally divided by the integer to separate the image blocks to obtain a character image block. Further, if the height of a plurality of image blocks becomes a standard character height when the image blocks are combined, a character image block is generated by combining the image blocks. For example, in FIG. 8, two image blocks are generated by dividing the image block 804 into two, and one image block is generated by combining the two image blocks with the image block 805.
[0006]
As described above, since character extraction is performed by generating a character image block in steps S702 and S703, an identification operation is performed for each character image block in step S704, and the category (character) having the highest similarity is determined. ) As the recognition result.
[0007]
[Problems to be solved by the invention]
However, in character extraction by the above-described method, horizontal characters included in a vertically written document as shown in FIG. 10 are extracted as one character image block as in 901. There is a problem that the character cannot be cut out in character units, and the recognition cannot be performed or a recognition error is caused.
[0008]
The present invention has been made in view of such a problem, and when extracting characters included in a line in a row direction, it is possible to extract a character from a character string in a column direction, and the character recognition rate is improved. And a character recognition method and apparatus using the same.
[0009]
[Means for Solving the Problems]
A character recognition device according to the present invention for achieving the above object has the following configuration. That is,
A character recognition device that performs a character recognition process on a character image included in a document image,
An acquisition unit that extracts a row based on a projection in a row direction of the document image and acquires a rectangular area including a recognition target character image based on a projection in a column direction orthogonal to the extracted row;
A first recognizing means for performing a character recognition process for each rectangular area obtained by the obtaining means to obtain a category of a recognition result and a similarity thereof;
First determining means for determining whether to divide the rectangular area in the column direction based on whether or not the similarity obtained by the first recognizing means is smaller than a predetermined threshold value;
The rectangular area determined to be divided by the first recognizing means is divided in the column direction by taking a projection in the row direction, and the rectangular area determined not to be divided by the first recognizing means. Is dividing means that does not perform the division in the column direction,
A second recognition unit that performs a character recognition process on each of the rectangular areas obtained by the division by the division unit and obtains a category of each recognition result and a similarity thereof;
Second determining means for determining whether at least one of the similarities obtained for each of the divided rectangular regions is smaller than a predetermined threshold,
For a rectangular area determined not to be divided by the first recognition means, the category obtained by the first recognition means is output as a final result, and it is determined by the first recognition means to be divided and the second recognition means For at least one of the rectangular areas determined to be smaller than the predetermined threshold value, the category obtained by the first recognizing means is output as a final result, and the first recognizing means determines to divide the area. Output means for outputting, as a final result, the category obtained by the second recognition means for a rectangular area which is determined and all of the similarities obtained by the second recognition means are equal to or larger than the predetermined threshold. Having.
[0010]
Further, the character recognition method according to the present invention for achieving the above object,
A character recognition method for performing a character recognition process on a character image included in a document image,
An acquisition step of extracting a row based on the projection in the row direction of the document image, and acquiring a rectangular area including the recognition target character image based on the projection in the column direction orthogonal to the extracted row,
A first recognition step of performing a character recognition process on each of the rectangular areas obtained in the obtaining step to obtain a category of a recognition result and a similarity thereof;
A first determination step of determining whether to divide the rectangular area in the column direction based on whether the similarity obtained in the first recognition step is smaller than a predetermined threshold value,
The projection in the row direction is performed on the rectangular area determined to be divided in the first recognition step to perform the division in the column direction, and the rectangular area determined not to be divided in the first recognition step is determined. Is a division step in which the division in the column direction is not performed,
A second recognition step of performing a character recognition process on each of the rectangular areas obtained by the division in the division step to obtain a category of each recognition result and a similarity thereof;
A second determining step of determining whether at least one of the similarities obtained for each of the divided rectangular areas is smaller than a predetermined threshold,
For the rectangular area determined not to be divided in the first recognition step, the category obtained in the first recognition step is output as a final result, and it is determined that the division is performed in the first recognition step and the second recognition step is performed. For a rectangular area in which at least one of the degrees of similarity obtained in the above is determined to be smaller than the predetermined threshold, the category obtained in the first recognition step is output as a final result, and it is determined that the category is divided in the first recognition step. An output step of outputting, as a final result, the category obtained in the second recognition step for a rectangular area that has been determined and all of the similarities obtained in the second recognition step are equal to or larger than the predetermined threshold. Having.
[0018]
In the present invention, the row direction and the column direction correspond to the vertical direction and the horizontal direction, respectively, in the case of a vertically written document. In a horizontally written document, the row direction corresponds to the horizontal direction, and the column direction corresponds to the vertical direction.
[0019]
【Example】
Preferred embodiments of the present invention will be described below with reference to the accompanying drawings.
[0020]
[Example 1]
FIG. 1 is a block diagram showing a configuration of a character recognition device for implementing the present invention. A CPU 101 performs various controls in the apparatus according to a control program stored in a ROM 102. A ROM 102 stores a control program executed by the CPU 101 and various data. A control program for realizing the processing shown in the flowchart described later is also stored in the ROM 102. A RAM 103 provides a work area when the CPU 101 executes various processes. The RAM 103 also includes an area for storing an optically read document image or the like. A display 104 displays a read image, various recognition results, and the like. A keyboard 105 is used to input various instructions and data. An image scanner 106 optically reads an image.
[0021]
Next, the operation of this embodiment executed by the character recognition device having the configuration shown in FIG. 1 will be described.
[0022]
FIG. 2 is a diagram illustrating an outline of cutting out a character image block according to the present embodiment. In the figure, reference numeral 201 denotes an image block having a standard size, which is directly used as a character image block. Reference numeral 202 denotes an image block for horizontal text, which has a horizontal width (w1) wider than the standard character width. Reference numeral 203 denotes a standard block representing a standard size of a character image block, which is calculated by a predetermined method from an image block cut out from a character image read by the scanner 106. By multiplying the horizontal width ws of the standard block 203 by a predetermined coefficient, a threshold value T for identifying an image block including horizontal characters is obtained. Then, the horizontal width of each of the extracted image blocks is compared with a threshold value T, and an image block having a horizontal width larger than T is an image block including a horizontal text, and a horizontal text division process is performed. For example, FIG. 2 shows a state in which the image block 202 satisfies w1> T, and the image block is further projected in the row direction, and as a result, is divided into three blocks.
[0023]
Next, the operation of the first embodiment will be described in more detail with reference to FIGS.
[0024]
FIG. 3 is a flowchart illustrating a character recognition process performed by the character recognition device according to the first embodiment. First, in step S201, a document image is input using the image scanner 106, and the obtained image data is stored in the RAM 103. At this time, the input image may be displayed on the display 104.
[0025]
Next, vertical writing characters are extracted from the vertical writing document, and the result is stored in the RAM 103. Next, in step S202, an image block is cut out. As described in the conventional example, vertical characters are cut out to extract lines by taking a vertical projection in step S202, and to take out image blocks by taking a horizontal projection for each line. However, if there is a horizontal character in the line as it is, the horizontal width of the horizontal character becomes the line width. Therefore, in this example, the left and right margins are cut off for each cut-out image block, and the width of each image block is corrected.
[0026]
Next, in step S203, an image block (character image block) for each character is generated. Here, as described in the description of the related art, an image block in units of one character is generated by performing separation and combination for each image block.
[0027]
Through the processing of the subsequent steps S204 to S208, it is determined whether or not all the character image blocks are horizontal characters, and the character image blocks determined to be horizontal characters are further separated.
[0028]
In step S204, the character image blocks are focused on one by one, and it is checked whether the character image blocks include horizontal characters. If it is determined that the target character image block includes horizontal characters, the process proceeds to step S205 to perform a horizontal writing character cutout process on the character image block. On the other hand, if it is determined that the character image block of interest does not include horizontal characters, the process proceeds to step S207. The method for determining whether or not the character is a horizontal character in step S204 will be described later.
[0029]
In the cutout processing for horizontal writing characters, first, an image block is extracted in step S205. Generally, the same processing is performed for the cutout processing for horizontally written characters as for the cutout processing for vertically written characters. That is, horizontal projections are extracted from the input image in the left-right direction to extract horizontal rows, and then image blocks are cut out by projecting each row in the vertical direction. However, in the case of the present embodiment, since the process is for the image block of interest, the process may be started assuming that a row in the horizontal direction has already been extracted.
[0030]
Next, in step S206, an image block for each character is generated. This can be performed in the same manner as in the case of vertical writing by replacing the “height” of the image block in the cutout of the character image block in the vertical writing character image with “width”. However, since the number of characters in a horizontally composed character of a vertically written document is small, the standard character width is often not accurately obtained. Therefore, this processing (step S206) may be omitted and the image block obtained in step S205 may be used as it is as a character image block. Then, the above result is stored in the RAM 103.
[0031]
In step S207, the character image block of interest is updated to the next block. In step S208, it is determined whether or not the investigation in step S204 has been completed for all the character image blocks, and if not completed, the flow returns to step S204 to repeat the above processing.
[0032]
On the other hand, if the investigation in step S204 has been completed for all blocks, the process proceeds to step S210. In step S210, a classification operation is performed for each character image block, and a category (character) having the highest similarity is set as a recognition result and stored in the RAM 103. Here, the recognition result may be displayed on the display 104.
[0033]
Next, the determination of whether or not the character is a horizontal character performed in step S204 will be described in more detail with reference to the flowchart of FIG. FIG. 4 is a flowchart illustrating a processing procedure for determining whether a character image block includes horizontal characters.
[0034]
First, in step S301, the widths of all the character image blocks are obtained. Next, in step S302, for example, the average of the widths of the character image blocks in the line is obtained, and the standard character width is obtained. Here, if punctuation marks, symbols, and the like include extremely small widths, the accuracy of the standard character width is reduced. For example, the following processing is performed.
[0035]
For example, based on the standard character height H obtained in the generation of the character image block in step S203 and the values m and n (0 <m <n) empirically obtained in advance, the image block width w becomes mH <w <. A character image block in the range of nH is extracted. Then, the average value is calculated using the extracted character image blocks, and the average value is used as the standard character width, whereby the accuracy of the standard character width can be improved. Alternatively, a distribution relating to the width of the character image block may be obtained, and the representative value of the class having the highest frequency may be set as the standard character width. Here, the values of m and n described above are preferably about m = 0.7 and n = 1.3.
[0036]
In step S303, using the obtained standard character width w, a threshold value T for determining whether or not each character image block includes horizontal characters is determined. Generally, in a vertically written document, the width of a character image block including horizontal characters is larger than the standard character width ws. Therefore, for example, a threshold value T = t × ws is obtained using an empirically obtained value t (t> 1), and the threshold value T is compared with the width of each character image block to obtain each character image block. It can be determined whether the block includes a horizontal character string. Here, the value of t is preferably about t = 1.2.
[0037]
Therefore, in step S304, the width w of each image block is compared with the threshold T. Here, if w> T, it is determined that the character image block includes horizontal characters, and the process proceeds to step S305. On the other hand, if w ≦ T, the process proceeds to step S306, where a normal character image block is set.
[0038]
As described above, according to the first embodiment, by utilizing the fact that the width of the character image block of the horizontal character in the vertical writing is larger than the width of the normal character image block, the character including the horizontal character is used. Image blocks can be extracted. Further, by performing the character extraction process for horizontal writing on the character image block, it becomes possible to extract horizontal characters included in a vertically written document in units of characters.
[0039]
The above embodiment is one configuration example for implementing the present invention, and it goes without saying that various applications are possible. For example, it is configured such that an operator can specify a recognition area, perform an operation of correcting a recognition result, automatically determine a recognition area without intervention of an operator, and reduce erroneous recognition after this processing. May be added.
[0040]
Further, the threshold value T may be changed according to the height of each character image block. In the first embodiment, one threshold value T is determined based on the standard character width obtained from the standard character height. However, when quadruple-width characters are mixed in full-width characters, the quadruple-width characters are displayed horizontally. It may be recognized as a kumimoji. Therefore, if the threshold T is a function of the height of the character image block (for example, threshold T = k × h, where k is a constant and h is the height of the character image block), such a problem can be solved.
[0041]
Further, it goes without saying that a program for performing the processing for carrying out the present invention may be externally provided to a general-purpose computer, and the control program of the present apparatus may be stored in the RAM.
[0042]
[Example 2]
Next, a second embodiment will be described. In the first embodiment, it is determined based on the size of the character image block whether or not the image is a character image block including horizontal characters. In the second embodiment, first, an identification operation is performed on each character image block, and a character image block in which the similarity is smaller than a predetermined value in only a category is processed as horizontal text. The configuration of the character recognition device according to the second embodiment is the same as that of the first embodiment (FIG. 1), and a description thereof will not be repeated.
[0043]
FIG. 5 is a flowchart illustrating a procedure of character recognition in the second embodiment. As in the above-described embodiment, first, a document image is input in step S401, an image block is extracted in step S402, and an image block (character image block) is created in character units in step S403. Thereafter, in step S404, an identification operation is performed for each character image block.
[0044]
Next, attention is paid to each of the character image blocks, and for each character image block, a determination is made as to whether or not to perform character cutout and identification calculation as horizontal characters (step S405). Here, if it is determined that the character image block of interest is to be reprocessed as a horizontally composed character, the process proceeds to step S406. On the other hand, if reprocessing for horizontal characters is not performed, the process proceeds to step S410. The determination processing in step S405 will be described later with reference to the flowchart of FIG.
[0045]
In step S406, the character image block for horizontal writing is subjected to the character image block to newly extract an image block. Further, in step S407, a character image block is generated from the image block obtained in step S406. However, as described in the first embodiment, in the generation of a character image block by separation and combination of image blocks in step S407, since the standard character width cannot be obtained with high accuracy, this step is omitted, and is obtained in step S406. The image block may be used as it is as a character image block.
[0046]
In step S408, an identification operation is performed on the character image block newly cut out in steps S406 and S407. In step S409, the character image block before character re-cutout (the character image block cut out in steps S402 to S403) and the character image block after re-character cut-out (character image block cut out in steps S406 to S407) ), The result with the higher reliability is adopted, and this is used as the recognition result. The determination of the reliability in step S409 will be described later with reference to the flowchart of FIG.
[0047]
In step S410, the target character image block is advanced to the next. Then, it is determined whether or not all the character image blocks have been checked in step S411, and if there is any unprocessed character image block, the process returns to step S405. On the other hand, when the processing has been completed for all the character image blocks, this processing is completed.
[0048]
Next, the determination method in step S405 described above will be described. FIG. 6 is a flowchart showing a procedure for determining whether or not each character image block is to be reprocessed as a horizontal character.
[0049]
In the identification operation in step S404, the similarity between all the categories and each character image block is calculated, and the category having the highest similarity is determined as the result of the identification operation. When trying to identify an image in which similar characters do not exist, such as when characters are cut out incorrectly, the degree of similarity is generally lower than when images are cut out correctly. In this example, it is determined whether reprocessing is necessary using this property. Here, let V be a threshold value of the similarity for determining whether or not the character cutout needs to be reprocessed.
[0050]
First, in step S501, the preset similarity threshold V is compared with the similarity v of the target character image block. If v <V, the process advances to step S503 to determine that the block is a character image block to be processed again. If v ≧ V, the process proceeds to step S502. In step S502, a misrecognition list in which characters that are easily misrecognized such as "10" and "$" or "00" and "@" are stored in advance, and the recognition result in step S404 is included in the misrecognition list. Find out if you can. If the recognition result exists in the misrecognition list, the process proceeds to step S503, and it is determined that the character image block is to be processed again. The erroneous recognition list is stored in the ROM 102 or the RAM 103.
[0051]
As described above, the character image block to be subjected to the character cutout processing is selected again.
[0052]
In step S409, it is determined which of the recognition results before and after re-character segmentation is to be used. This processing will be described below. FIG. 7 is a flowchart illustrating a procedure for determining which of the results before and after the re-processing is to be adopted for a character image block that has been re-processed for character extraction (that is, character extraction for horizontal writing).
[0053]
First, in step S601, the similarity is smaller than a predetermined threshold V as a result of the identification operation (step S408) for the image block divided by the re-processing of character extraction (character extraction for horizontal writing in steps S406 to S407). Determine if something exists. If there is even one with a low similarity, the process proceeds to step S602, and the original result before reprocessing (that is, the result obtained in step S404) is set as the final recognition result. On the other hand, if the similarities in all the newly extracted character image blocks are larger than the threshold value V, it is determined that the horizontal character extraction is successful, and the result after re-processing in step S603 (that is, the result obtained in step S408). Result) is the final result.
[0054]
As described above, according to the second embodiment, the similarity obtained by the identification operation is used to determine whether or not to perform the processing for the vertically written horizontal characters. It becomes possible to recognize the kumi character portion. Further, after the processing of the horizontal writing characters in the vertical writing is completed, it is determined whether or not to use the result, thereby reducing erroneous recognition and improving recognition accuracy.
[0055]
Further, it goes without saying that the process of determining whether or not to perform reprocessing using the size of the character image block described in the first embodiment may be performed with a configuration added to the second embodiment. .
[0056]
As described above, when recognizing a vertically written document, vertical writing that cannot be correctly performed until now can be performed by determining horizontal writing characters in vertical writing and performing processing for horizontal writing on the portion. It has the effect of enabling recognition of horizontal characters in a horizontal direction, making it possible to perform an input operation by OCR more accurately, and reducing erroneous recognition and correction of characters that cannot be recognized.
[0057]
In the above embodiment, the mixture of horizontal characters in a vertically written document has been described. However, it is needless to say that the mixed concept of vertical characters in a horizontally written document can be realized by the same concept as described above.
[0058]
The present invention may be applied to a system including a plurality of devices or an apparatus including a single device. Needless to say, the present invention can also be applied to a case where the present invention is achieved by supplying a program for causing a system or an apparatus to execute a process defined by the present invention.
[0059]
【The invention's effect】
As described above, according to the present invention, when characters included in a line in the row direction are extracted, characters can be extracted from a character string in the column direction, and the character recognition rate is improved.
[0060]
[Brief description of the drawings]
FIG. 1 is a block diagram showing a configuration of a character recognition device for implementing the present invention.
FIG. 2 is a diagram illustrating an outline of cutting out a character image block according to the embodiment.
FIG. 3 is a flowchart illustrating a character recognition process performed by the character recognition device according to the first embodiment.
FIG. 4 is a flowchart illustrating a processing procedure for determining whether or not a character image block includes a horizontal composition character.
FIG. 5 is a flowchart illustrating a procedure of character recognition according to the second embodiment.
FIG. 6 is a flowchart illustrating a procedure for determining whether or not each character image block is to be reprocessed as a horizontal character.
FIG. 7 is a flowchart illustrating a procedure for determining which recognition result before or after reprocessing is to be adopted.
FIG. 8 is a flowchart illustrating a procedure of a general OCR character recognition process.
FIG. 9 is a diagram illustrating a method of extracting image blocks used in general character recognition processing.
FIG. 10 is a diagram illustrating an example of a vertically written document including horizontal characters.
[Explanation of symbols]
101 CPU
102 ROM
103 RAM
104 display 105 keyboard 106 image scanner

Claims (2)

A character recognition device that performs a character recognition process on a character image included in a document image,
An acquisition unit that extracts a row based on a projection in a row direction of the document image and acquires a rectangular area including a recognition target character image based on a projection in a column direction orthogonal to the extracted row;
A first recognizing means for performing a character recognition process for each rectangular area obtained by the obtaining means to obtain a category of a recognition result and a similarity thereof;
First determining means for determining whether to divide the rectangular area in the column direction based on whether or not the similarity obtained by the first recognizing means is smaller than a predetermined threshold value;
The rectangular area determined to be divided by the first recognition means is divided in the column direction by projecting in the row direction, and the rectangular area determined not to be divided by the first recognition means is calculated. Is dividing means that does not perform the division in the column direction,
A second recognition unit that performs a character recognition process on each of the rectangular areas obtained by the division by the division unit and obtains a category of each recognition result and a similarity thereof;
Second determining means for determining whether at least one of the similarities obtained for each of the divided rectangular regions is smaller than a predetermined threshold,
For a rectangular area determined not to be divided by the first recognition means, the category obtained by the first recognition means is output as a final result, and it is determined by the first recognition means to be divided and the second recognition means For at least one of the rectangular areas determined to be smaller than the predetermined threshold value, the category obtained by the first recognizing means is output as a final result, and the first recognizing means determines to divide the area. Output means for outputting, as a final result, the category obtained by the second recognition means for a rectangular area which is determined and all of the similarities obtained by the second recognition means are equal to or larger than the predetermined threshold. A character recognition device comprising: A character recognition method for performing a character recognition process on a character image included in a document image,
An acquisition step of extracting a row based on the projection in the row direction of the document image, and acquiring a rectangular area including the recognition target character image based on the projection in the column direction orthogonal to the extracted row,
A first recognition step of performing a character recognition process on each of the rectangular areas obtained in the obtaining step to obtain a category of a recognition result and a similarity thereof;
A first determination step of determining whether to divide the rectangular area in the column direction based on whether the similarity obtained in the first recognition step is smaller than a predetermined threshold value,
The rectangular area determined to be divided in the first recognition step is divided in the column direction by taking the projection in the row direction, and the rectangular area determined not to be divided in the first recognition step is determined. Is a division step in which the division in the column direction is not performed,
A second recognition step of performing a character recognition process on each of the rectangular areas obtained by the division in the division step to obtain a category of each recognition result and a similarity thereof;
A second determining step of determining whether at least one of the similarities obtained for each of the divided rectangular areas is smaller than a predetermined threshold,
For the rectangular area determined not to be divided in the first recognition step, the category obtained in the first recognition step is output as a final result, and it is determined that the division is performed in the first recognition step and the second recognition step is performed. For a rectangular area in which at least one of the degrees of similarity obtained in the above is determined to be smaller than the predetermined threshold, the category obtained in the first recognition step is output as a final result, and it is determined that the category is divided in the first recognition step. An output step of outputting, as a final result, the category obtained in the second recognition step for a rectangular area that has been determined and all of the similarities obtained in the second recognition step are equal to or larger than the predetermined threshold. Character recognition method characterized by having.

Images