JP5630689B2 - Character recognition method and character recognition device - Google Patents

Description

  The present invention relates to a character recognition method and a character recognition device for optically reading characters on a medium such as paper or plastic.

  Conventionally, there is an apparatus for recognizing a character string printed on the surface of a medium such as a check. For example, a character part is detected from binarized image data, the character part is cut out along a circumscribed rectangular frame, a feature amount is extracted from the obtained character pattern, and the feature vector of the input pattern and the feature dictionary are extracted. There is known a character recognition device that obtains a similarity to a reference vector for each character and selects a candidate character according to the magnitude of the similarity.

  Among such character recognition devices, there have been various devices such as binarization based on the recognition result in order to improve recognition accuracy (see, for example, Patent Documents 1 and 2). .

  More specifically, in the technique disclosed in Patent Document 1, it is necessary to determine whether or not it is necessary to rebinarize a multi-valued image based on the result of character recognition and the result of counting the number of black spots. For example, a threshold value different from the initial threshold value is set, and the multi-valued image is binarized using the new threshold value. Moreover, in the technique disclosed in Patent Document 2, when a multi-value image is binarized for each character and the result of character recognition cannot be collated, graphic information such as a character line width is extracted, for example. Re-binarization according to the value is performed.

Japanese Patent Application Laid-Open No. 09-016715 Japanese Patent Laid-Open No. 09-138834

  However, in the technique disclosed in Patent Document 1, whether or not to perform binarization is determined according to the counting result of the number of black spots of the binarized character image. For this reason, there is no major problem when the character is similar to the standard character shape. However, if the character shape changes greatly due to deformation during printing, chipping after printing, or the like, there is a possibility that a correct determination cannot be made. That is, there is a problem in terms of ensuring the reliability of the recognition result. Further, even the technique disclosed in Patent Document 2 has a problem that it is difficult to set a threshold for re-binarization depending on graphic information to be extracted. For example, when determining the threshold for re-binarization, the character line width is not necessarily thick or thin uniformly, and it is difficult to set an accurate threshold for re-binarization.

  The present invention has been made in view of these points, and an object of the present invention is to provide a character recognition method and a character recognition device that can improve the reliability of the recognition result of character recognition.

  In order to solve the above problems, the present invention provides the following.

(1) A binarization processing step for binarizing image data obtained by imaging a medium, a character cutting step for obtaining a region including a recognition target character from the binarized image data, and the region A feature extraction step for extracting a feature vector of a character to be recognized, a feature vector extracted in the feature extraction step is compared with a given reference feature vector for each of a plurality of reference characters, and a similarity with each reference feature vector is determined. Based on a similarity calculation step to be calculated, a first character determination step for comparing a plurality of similarities calculated in the similarity calculation step with a similarity threshold, and determining a candidate character, and the first character determination step In order to separate and evaluate each of the plurality of candidate characters and the recognition target character when each of the similarities of the plurality of candidate characters clears the similarity threshold , A separation processing step for obtaining a separation evaluation value by a combination of feature vectors in at least two or more divided regions in the region divided into n × m divided regions, and comparing each of the separation evaluation values with a predetermined separation evaluation threshold value and a character recognition method which comprises a, a second character judgment step of judging candidate characters by the magnitude relationship between the separation evaluation value and the predetermined separation evaluation threshold.

  According to the present invention, image data obtained by imaging a medium is binarized, a region including a recognition target character is obtained from the binarized image data, and a feature vector of the recognition target character in the region is extracted. The extracted feature vector is compared with each given reference feature vector for each of a plurality of reference characters, a similarity with each reference feature vector is calculated, and the calculated plurality of similarities are compared with a similarity threshold. A candidate character is determined, and when there are a plurality of candidate characters based on the determination result, a separation evaluation value for separating and evaluating each of the plurality of candidate characters and the recognition target character is obtained, and the separation evaluation value is separated. Since it was decided to determine the candidate character in comparison with the evaluation threshold, it is possible to make a determination by another feature vector (separation evaluation value) for those that could not be distinguished by the determination of the first candidate character, It is possible to improve the reliability of the identification results.

In the character recognition method according to the present invention, the separation evaluation value in the separation processing step is to extract at least two or more character feature values in the divided area after dividing the area into n × m divided areas. It is calculated | required by doing .

  According to the present invention, the separation evaluation value in the separation evaluation step described above is obtained by dividing an area into n × m divided areas and then extracting at least two or more character feature quantities in the divided areas. Therefore, the accuracy of the separation evaluation value can be changed by setting the values of n and m (integers) to appropriate values. For example, if n or m is set to a large value, the separation evaluation value is obtained in a more finely divided area, so that the accuracy can be further improved. Note that n and m may be different values or the same value (n = m).

( 2 ) A binarization processing unit that binarizes image data obtained by imaging a medium, a character cutout unit that obtains a region including a recognition target character from the binarized image data, and the region A feature extraction unit that extracts a feature vector of a character to be recognized, and a feature vector extracted by the feature extraction unit is compared with a given reference feature vector for each of a plurality of reference characters, and a similarity with each reference feature vector is determined. A character recognition unit to be calculated; a determination unit that compares a plurality of similarities calculated by the character recognition unit with a similarity threshold to determine a candidate character; and each similarity of a plurality of candidate characters based on the determination unit If there are together clearing the similarity threshold, of the plurality to candidate characters evaluated separated from each and the recognized characters, n × m at least two or more in the area that is divided into divided areas Divided area With obtaining the separation evaluation value by a combination of definitive feature vectors, each said separation evaluation value is compared with a predetermined separation evaluation threshold, determining a candidate character by the magnitude relationship between the said separation evaluation value given separation evaluation threshold And a separation processing unit.

  According to the present invention, when there are a plurality of candidate characters based on the determination result of the determination unit, a separation evaluation value for separating and evaluating each of the plurality of candidate characters and the recognition target character is obtained, and the separation evaluation value Since the candidate character is determined by comparing with the separation evaluation threshold, the second character with a different feature vector (separation evaluation value) for those that could not be distinguished by the determination of the candidate character in the first character determination The determination can be performed, and the reliability of the recognition result can be improved.

  According to the character recognition method and the character recognition device of the present invention, even if there are a plurality of candidate characters in the first character determination, the second character determination is performed using the separation evaluation value that is another feature vector. Therefore, the reliability of the recognition result of character recognition can be easily increased.

  The best mode for carrying out the present invention will be described below with reference to the drawings.

[Character recognition device]
FIG. 1 is a block diagram showing an electrical configuration of a character recognition device 1 according to an embodiment of the present invention.

  The character recognition device 1 optically reads an image of a character string printed on the surface of a recording medium by an image scanner, cuts out a character image for each character to be recognized from the read character string image, and this character image Character recognition was performed by comparing and collating the feature vectors (feature points, feature quantities, etc.) extracted from the basic feature vectors (feature points, feature quantities, etc.) stored in the recognition dictionary. An optical character reader that outputs characters is used.

  In FIG. 1, a recording medium 2 is a general card conforming to JIS, for example, a plastic card having a size of 86 mm in width, 54 mm in height, and 0.76 mm in thickness.

  As shown in FIG. 1, the recording medium 2 is set with a character recording area 2A printed with optically readable characters and numerals. Specifically, a plurality of characters are regularly arranged in a certain direction to form a character string, and a sentence or the like is formed. In the character recording area 2A, for example, unique information for specifying the recording medium 2 is displayed. This unique information is, for example, the name of the user of the recording medium 2, the number that identifies the recording medium 2, the user number of the recording medium 2, and the like. Note that the recording medium 2 is not limited to this plastic card, and any recording medium 2 may be used as long as a character string is printed thereon. For example, the recording medium 2 may be an ID card, a passport book, or a driver's license.

  A character recognition device 1 shown in FIG. 1 is a device capable of optically reading a character string displayed on a recording medium 2, and includes a close contact type (one-dimensional) image sensor 11, an image memory 12, and a data processing unit. 13.

  As shown in FIG. 1, the contact type (one-dimensional) imaging device 11 is provided in a medium transport mechanism that transports the recording medium 2 in a predetermined transport direction T (from left to right in FIG. 1) by a transport unit (not shown). It has been. The image sensor 11 is a line sensor composed of a CCD or the like, and is arranged in a line in a direction substantially perpendicular to the transport direction T (width direction of the recording medium 2) and is a character recording area formed on the recording medium 2. It is formed longer than the width of 2A. The medium transport mechanism is formed with a transport guide (not numbered), and the recording medium 2 is transported with one end in contact with the transport guide. Note that the image sensor 11 is not limited to a one-dimensional image and may be a two-dimensional image sensor.

  The image memory 12 stores image data obtained by imaging. Specifically, the image data is regarded as a matrix of a finite number of pixels. For example, each pixel has 8-bit information amount, is represented by 256-level multi-value image data, and the luminance value as a pixel value is Expressed as an integer value from 0 to 255. The image memory 12 may be any memory, such as RAM, SDRAM, DDR SDRAM, or RDRAM, as long as it can store image data.

  The data processing unit 13 reads out the multivalued image data from the image memory 12, performs various processes in the above-described units, and finally recognizes the character string on the recording medium 2. In the present embodiment, as shown in FIG. 1, a binarization processing unit 13a, a character segmentation unit 13b, a similar feature extraction unit 13c, a character recognition unit 13d, a feature dictionary storage unit 13e, and a determination unit 13f A rebinarization processing unit 13g, a separation processing unit 13h, a character output unit 13i, and a separation condition storage unit 13j. Each of these units can be realized in hardware by an electrical element such as a CPU, MPU, RAM, ROM, or the like.

  The binarization processing unit 13a binarizes image data obtained by imaging the medium. In the present embodiment, 256-level multi-valued image data stored in the image memory 12 is read out, and the first binarization threshold value is obtained by an appropriate method such as a well-known discriminant analysis method. Conversion into binary binary image data is performed. The binary image data is stored as binary image data in the image memory 12 in a memory area different from the memory area in which the multi-value image data is stored.

  The character cutout unit 13b extracts a region including a recognition target character (hereinafter referred to as “character”) from the binary image data obtained by the binarization processing unit 13a, that is, a character image for each character. It is detected and cut out. In the present embodiment, a circumscribed rectangular area circumscribing one character (for example, a black pixel area forming one character) is cut out from the binary image data.

  The feature extraction unit 13c extracts a feature vector (for example, feature amount, feature point, etc.) of the recognition target character in the region cut out by the character cutout unit 13b. In the present embodiment, the circumscribed rectangular area supplied from the character cutout unit 13b is divided into a plurality of partial areas, and feature vectors for each partial area are extracted.

  The character recognition unit 13d performs character recognition for each character and outputs candidate characters. Here, the character recognition is performed by matching each of the character feature vectors cut out by the character cutout unit 13b with reference feature vectors of individual reference characters registered in advance in the recognition dictionary 315 one by one. Done. In the present embodiment, the feature vector extracted by the feature extraction unit is compared with each given reference feature vector for each of a plurality of reference characters, and the similarity with each reference feature vector is calculated.

  The feature dictionary storage unit 13e stores reference feature vectors obtained in advance for all characters used in the recording medium 2. When the feature comparison is performed, the feature vectors of candidate characters having a high similarity are read from the feature dictionary storage unit 13e, the features are compared, and the similarity is calculated.

  The determination unit 13f determines a recognition result from the similarity calculated by the character recognition unit 13d. In this embodiment, candidate characters having a calculated similarity higher than the similarity threshold are read.

  The re-binarization processing unit 13g, in the determination unit 13f, when the maximum similarity value does not clear the predetermined similarity threshold, that is, for the candidate character that does not clear the similarity threshold, the recognition target character Can not be recognized, re-binarization processing is executed. In this case, the above feature extraction unit 13c reprocesses using a second binarization threshold different from the first binarization threshold. In the separation processing unit 13h, if the separation evaluation value does not clear the predetermined separation evaluation threshold, that is, the recognition target character cannot be recognized for the candidate character that does not clear the separation evaluation threshold, rebinarization is performed. Processing is executed. In this case, the separation processing unit 13h performs reprocessing using a second binarization threshold different from the first binarization threshold.

  When the determination unit 13f clears a predetermined similarity threshold for a plurality of candidate characters, the separation processing unit 13h performs a separation process on the plurality of candidate characters. For example, when the similarity between the first and second character candidates clears the similarity threshold and both are close to each other (the number “0 (zero)” and the alphabet character “O (o)” , Etc.), etc., etc., etc., etc., etc., etc., etc.), etc., etc., etc., etc.), etc., etc., etc., etc., etc., etc., etc., etc., etc., etc., etc., etc., etc., etc., etc., etc., etc., etc., etc., etc., etc., etc., etc., etc., etc. In this process, an evaluation value that can best separate a plurality of candidate characters that have cleared the similarity threshold is calculated from an appropriate combination of one or more feature vectors and obtained as a separation evaluation value. Separation is performed by comparison with a separation evaluation threshold. Further, the separation processing unit 13h determines a recognition result from the calculated separation evaluation value. In the present embodiment, candidate characters having a calculated separation evaluation value lower than the separation evaluation threshold are read.

  The character output unit 13i outputs the recognition result from the determination unit 13f.

  The separation condition storage unit 13j has a plurality of candidate characters that are similar in similarity, for example, the number “0 (zero)” and the alphabet character “O (O)”, the numbers “6” and “8”, and the alphabet character “B”. A calculation formula for obtaining a separation evaluation value that makes it possible to determine the above, information on the height H of the candidate characters, and the like are stored.

[Character recognition method]
FIG. 2 is a flowchart showing a character recognition method according to the embodiment of the present invention.
FIG. 3 is a diagram showing binary image data when a character string printed in the character storage area 2A (see FIG. 1) of the recording medium 2 is imaged.

  As shown in FIG. 2, in the character recognition method according to the present embodiment, first, a recording medium 2 such as a card is transported along a transport guide of a medium transport mechanism and passes through an image sensor 11, thereby recording medium. 2 is recorded. Analog image data obtained by imaging is photoelectrically converted, and an analog / digital conversion circuit (not shown) is connected to convert the analog image data into digital image data. For example, it is converted into multi-value image data of 256 bits of 256 bits and stored in the image memory 12. Since such an analog / digital conversion circuit itself is already known, a detailed description thereof is omitted here.

  FIG. 3 is a diagram showing image data when a character string printed in the character storage area 2A (see FIG. 1) of the recording medium 2 is imaged. In the present embodiment, two character strings are printed. In order to simplify the description, the direction of the character string substantially coincides with the horizontal direction of the image memory 12, and the direction orthogonal to the direction of the character string is the vertical direction.

  In the present embodiment, the image memory 12 is configured by arranging M rows × N columns of pixels in a matrix, and each pixel stores multi-value image data in the character recording area 2A. Specifically, the luminance value (maximum value 255) is a relatively large value in the pixels constituting the white element, and the luminance value (minimum value 0) is a relatively small value in the pixels constituting the black element. ing.

  In FIG. 2, first, binarization processing is performed (step S1). Specifically, the binarization processing unit 13a reads 256-level multi-level image data from the image memory 12, obtains a first binarization threshold by an appropriate method such as a discriminant analysis method, and obtains a black and white binary value. Conversion to value image data. This step S1 corresponds to an example of a “binarization processing step” for binarizing image data obtained by imaging a medium.

  Next, a character cutout process is performed (step S2). The character cutout unit 13b cuts out a character image for each character of the recognition target character printed in the character recording area 2A formed on the recording medium 2. This step S2 corresponds to an example of a “character extraction step” for obtaining a region including a recognition target character from the binarized image data.

  More specifically, the character string is first cut out. Specifically, the character cutout unit 13b projects binary image data of two character strings in the horizontal direction (character string direction) and detects the upper and lower edges (upper and lower in FIG. 3) of each character string. To do. Then, two character strings are cut out using the center position of the upper and lower edges as the center line of each character string. Next, character segmentation is performed for each character string. First, character delimiter position detection is performed. Specifically, as described above, the character cutout unit 13b performs projection calculation in the vertical direction on the detected binary image data for each character string. Since the vertical projection data obtained by the vertical projection includes a blank portion between the character constituting the character string and the adjacent character, the location where the vertical projection data exceeds the predetermined threshold for detecting the delimiter position, It is detected as the delimiter position of the characters that make up the string. This process will be described in detail with reference to FIGS. In the character delimiter position detection, a position where the vertical projection data exceeds a predetermined delimiter position detection threshold (a line indicated by SLEV in FIG. 5) is detected as a delimiter position between the character constituting the character string and the adjacent character. .

  FIG. 4 is binary image data when attention is paid to characters (part) constituting the character string. FIG. 5 shows vertical projection data obtained by vertical projection of the binary image data shown in FIG. 4 shows the image data from “K” to “C” of the upper character string in FIG. 3 in FIG. 5, and the blank portion between each character exceeds SLEV (separation position detection threshold). It shows that.

  Next, a process for accurately determining the character position in the outer peripheral rectangular area of the temporarily determined character will be described with reference to FIG. FIG. 6 is a diagram showing binary image data of the character “<” constituting the character string.

  In FIG. 6, in order to accurately specify the position of the character “<”, vertical projection PX and horizontal projection PY are calculated in the binary image data in the outer peripheral rectangular area. Then, the left and right boundary positions of the characters are detected along the obtained PX, the PX is scanned rightward from the left end point of the outer peripheral rectangular area, and the level value ( When the pixel detection) continuously falls below the boundary position detection threshold value a certain number of times (for example, 3 times), the first point is set as the left edge of the character. Similarly, PX is scanned leftward starting from the right end point of the outer peripheral rectangular area, and when it falls below the boundary position detection threshold, the first point is set as the right edge of the character.

  Next, the same processing is performed for the horizontal projection PY. In FIG. 6, the range divided by the left edge position and the right edge position obtained by the boundary determination process described above is used as the definition area, and the upper and lower ends of the outer peripheral rectangular area are scanned from the outer side. If the value falls below the boundary position detection threshold for a certain number of consecutive times, the first point is taken as the upper and lower edges of the character. In this manner, the position of the character “<” in the character string is specified, and a circumscribed rectangular area circumscribing one character is cut out. Thus, by obtaining a circumscribed rectangular area circumscribing the character, it is possible to perform accurate character extraction, and reduce errors (due to margins) when performing the similarity calculation processing. Can do. When the character cut-out process in step S2 in FIG. 2 is completed, a circumscribed rectangular area of the recognition target character is obtained as shown in FIG.

  Next, feature extraction processing is performed (step S3). Specifically, the feature extraction unit 13c divides the circumscribed rectangular area into arbitrary partial areas. In the present embodiment, as shown in FIG. 7, for example, one circumscribed rectangular area is divided into 5 × 5 areas, and each area is a partial area. In each partial area, the ratio of the number of black pixels to the total number of pixels in the partial area is obtained, and a feature quantity as a feature vector having these as elements is generated. As shown in FIG. 7, the size of the area when the circumscribed rectangular area is divided into arbitrary partial areas may not be equal.

  FIG. 8 shows the circumscribed rectangular area shown in FIG. 7 (inside the area surrounded by the one-dot chain line in FIG. 6) divided into 5 × 5 areas shown by the one-dot chain line. The pixel value (binary) is shown. In FIG. 8, the numerical value 255 indicates binary “white”, and the numerical value 0 indicates binary “black”.

Next, a feature amount as a feature vector is calculated for each partial region. As the feature amount, a ratio of the number of black pixels (number of luminance values 0) to the total number of pixels in each partial region is obtained.
As a result, as shown in FIG. 9, 25 feature quantities can be generated. For example, 0.90909 in the upper right (symbol ax) in FIG. 9 is the number (30) of 0 (black pixels) in the upper right partial area ax in FIG. 8 corresponding to the partial area ax in FIG. The value is divided by the total number (33) of 0 (black pixels) and 255 (white pixels) (30 ÷ 33≈0.90909). This step S3 corresponds to an example of a “feature extraction step” for extracting feature vectors of recognition target characters in the region.

  Next, a character recognition process is performed (step S4). Specifically, the character recognizing unit 13d obtains the characteristic amount for each partial area obtained as described above for the reference characteristic amount (for example, FIG. 10) obtained in advance for all characters used in the recording medium 2. The similarity (for example, normalized correlation coefficient) is calculated, and the character having the highest similarity, that is, the one giving the maximum score value is set as the candidate character corresponding to the character. . In this embodiment, since the similarity to the character “<” is 0.902, which is the maximum value, “<” is the corresponding candidate character. In step S4, the extracted feature vector (feature amount) is compared with each given reference feature vector (reference feature amount) for each of a plurality of reference characters, and similar to each reference feature vector (reference feature amount). This corresponds to an example of a “similarity calculation step” for calculating the degree.

  The reference feature amount is stored in advance in the feature dictionary storage unit 13e. When feature comparison is performed, character data having a high degree of similarity is read from the feature dictionary storage unit 13e. A comparison is made.

  A candidate character determination process is performed (step S5). Specifically, the determination unit 13f determines whether a candidate character has been extracted. When one candidate character is extracted, it is output to the character output unit 13i, and the recognition result from the determination unit 13f is output (step S8). On the other hand, if the candidate character is not extracted, it is determined that re-binarization processing is necessary (step S5: NO), and the re-binarization processing unit 13g executes re-binarization processing on the (recognition target) character. (Step S6). When two or more candidate characters are extracted, they are output to the separation processing unit 13h, and separation processing is performed (step S7). This step S5 corresponds to an example of a “first character determination step” in which candidate characters are determined by comparing the calculated plurality of similarities with a similarity threshold.

(Separation process)
Separation processing by the separation processing unit 13h is performed (step S7). The character recognition unit 13d calculates the similarity with the (recognition target) character and the determination unit 13f determines the similarity. The plurality of candidate characters with the first and second similarities are similar to each other in the similarity threshold. If both are close to each other and the two are close to each other, the candidate characters cannot be identified and cannot be determined, and therefore, processing for separating them from the (recognition target) characters is performed. Examples of such candidate characters include the number “0 (zero)” and the alphabet “O (O)”, the alphabet lowercase “l (el)”, and the alphabet uppercase “I (eye)”. This step S7 corresponds to an example of a “separation process step” for obtaining a separation evaluation value for separating and evaluating each of the plurality of candidate characters and the recognition target character when there are a plurality of candidate characters.

  In this processing, an evaluation value that can best separate candidate characters having similar similarities with respect to the (recognition target) character is calculated from an appropriate combination of one or more elements of the feature amount, and the separation evaluation value Is compared with the separation evaluation threshold.

  The detailed information processing flow of this separation processing is as shown in FIG. A specific example is shown in FIG. As shown in FIG. 12, the number “0 (zero)” and the alphabet letter “O (o)” are typical candidate characters in the passport guide. On the font, it is possible to discriminate by the difference in the curvature of the stroke and the difference in the character height, but due to the printing condition, scanning condition, optical condition, etc., the distortion of the character and the change in the character line thickness are likely to occur. May become inseparable. The similarity in the feature amount in the character recognition process described above may take a close value, and the difference may not be significant. Therefore, in this embodiment, the separation is performed by comparing the separation evaluation value with the separation evaluation threshold.

For example, as shown in FIG. 12, the values DA, DB, DC, DD of the four corners of the partial divided area used when calculating the similarity, that is, the area divided by 5 × 5, and the median values E, F on the left and right are obtained. Using the calculation formula shown in FIG. 14A, a separation evaluation value is calculated for the characters that could not be distinguished from the first and second candidate characters (recognition target) (FIG. 11). Step S11). The separation evaluation value obtained by the calculation formula shown in FIG. 14A decreases with the number “0 (zero)”, but increases with the alphabet letter “O (O)”. Therefore, the discrimination logic is as shown in FIG. 14B (step S12 in FIG. 11). K1 and K2 in FIG. 14B are separation evaluation threshold values and can be obtained experimentally. When the obtained separation evaluation value fx is smaller than the separation evaluation threshold K1, the separation processing unit 13h determines that the (recognition target) character is the number “0 (zero)”. Further, when the determined separation evaluation value fx is greater than the separation evaluation threshold value K2, the character is determined as alphabetic letters "O (O)". Further, when the separation evaluation value fx is in a range between K1 and K2, another index described later is used as the non-determinable area. Note that although the 5 × 5 divided area is the partial divided area used in character recognition, a new partial divided area may be set to obtain a separation evaluation value.

  For example, if the character image (binary image) after the character recognition process is as shown in FIG. 12A, the separation evaluation value (the value of fx) cannot be determined. In the area. At this time, the character height H is used as another index. If H1 <H2 and the character height H (H ′) is H (H ′) <H1, “O” is set, and if H (H ′)> H2, “0 (zero)” is set. After the determination is made by the separation processing unit 13h and the result is confirmed (step S14 in FIG. 11), the character output is performed from the character output unit 13i (step S8 in FIG. 2). The character height H ′ indicates the character height of the character image formed by repeating the re-binarization process (step S13) described later (see FIG. 12B).

  If H1 ≦ H (H ′) ≦ H2, it is determined that the region is in an indeterminable region (step S12: NO in FIG. 11), and is output to the re-binarization processing unit 13g. The rebinarization processing unit 13g calculates a separation evaluation value fx for (recognition target) characters while repeating rebinarization by replacing the first binarization threshold with the second binarization threshold ( Step S13 in FIG. 11). The separation evaluation value fx obtained by recalculation is compared with the above-described separation threshold value K1 or K2, and it is determined whether the separation evaluation value fx obtained by recalculation is out of the indeterminable region or has reached the specified number of iterations. The final decision is made. When it is out of the indeterminable area, the corresponding character of “0 (zero)” or “O (O)” is output from the character output unit 13i, and when it is not out of the indeterminable area, the character output unit 13i. Is output as indeterminate.

  In the present embodiment, for example, when it is determined that the character image shown in FIG. 12A has a separation evaluation value (the value of fx) in an indeterminable region (step S12 in FIG. 11), a rebinarization process ( By repeating step S13), a character image as shown in FIG. 12B was formed. Separation processing is performed by the separation processing unit 13h using the binary image data of the character image, and the calculated separation evaluation value fx ′ <K1 (separation evaluation threshold value) is obtained, and “0 (zero)” is determined. In FIG. 12B, the re-binarized character image has the character height H ′, and the separation processing using the separation evaluation value fx ′ described above has not deviated from the undecidable area as before. As described above, the character height H ′ may be compared with H1 and H2. A detailed description of the rebinarization process will be described later. Step S11 described above is an example of a “separation process step” for obtaining a separation evaluation value (the above-described fx) for separating and evaluating each of the plurality of candidate characters and the recognition target character when there are a plurality of candidate characters. It corresponds to. Step S12 described above corresponds to an example of a “second character determination step” in which each of the separation evaluation values is compared with a separation evaluation threshold (K1 and K2 described above) to determine a candidate character.

(Rebinarization processing)
Returning to the flowchart of FIG. 2, when the candidate character is not extracted by the determination unit 13f, it is determined that the re-binarization process is necessary (step S5: NO), and the re-binarization process is performed on the (recognition target) character. The re-binarization process is executed by the unit 13g (step S6). As described above, in the character recognition process in step S4 of FIG. 2, the similarity to the reference character “<” is 0.902, which is the maximum value, and therefore “<” is determined as the corresponding character. Even if the character image is based on the first binarization threshold that is reasonably determined, the similarity between the recognition target character and the reference character may not reach the similarity threshold depending on the shape of the character (described above). In this case, a high degree of similarity such as 0.902 may not be given).
This is due to, for example, a decrease in resolution in the width direction due to the scanning speed in a manual scanning scanner, and the character shape is significantly distorted, or characters are blurred or missing during printing or during use. Because. In this case, in order to obtain the second binarization threshold value for re-binarization, for example, in the method using only the number of sunspots, the density is increased even if the density is apparently increased due to blurring / missing of characters. Threshold correction is performed in the direction of increasing darkness. When the character line other than the missing part is thick, it acts in the direction where the line becomes excessively thick, so that accurate recognition cannot be performed.

  Therefore, in the character recognition method according to the present embodiment, the similarity of feature amounts is used for determining whether or not re-binarization is necessary. Since deformations that occur in shapes such as missing characters or fading can be reliably reflected in the similarity, there is an advantage that a determination error is less likely to occur compared to the method using the number of black dots.

  FIG. 17 is an explanatory diagram for describing a specific example related to determination of necessity of re-binarization. FIG. 17 shows a character image (binary image data) of the capital letter “L” of the alphabet, and FIG. 17A shows the first binary used in the character recognition process (step S4 in FIG. 2). It is a figure at the time of applying a threshold value. In this case, since the binarization threshold is relatively low, the line becomes thin, and the right edge is detected on the inner side in the character cutting process. Therefore, the similarity to the reference character is a low value of 0.475. At this time, as described above, the determination unit 13f determines that the re-binarization process is necessary because it is impossible to determine because the collation is not correctly performed. Since the direction in which the binarization threshold is changed is unknown from the degree of increase or decrease, it is shifted in the direction of increasing first. As a character image, the number of black pixels increases. FIG. 17B illustrates a case where a luminance value increased by 8 from the luminance value on the 256-level multi-valued image data set as the first binarization threshold is set as the second binarization threshold. It is the character image which carried out the rebinarization process by this 2nd binarization threshold value. The similarity between this character image and the reference character is 0.572. Further, the luminance value is increased, the third binarization threshold is set, and the similarity between the re-binarized character image and the reference character is 0.886, which is a sufficiently large value (FIG. 17 (c)). )reference). As a result, it is determined that the original correct result has been reached. Note that, when the similarity is not sufficiently improved by changing the luminance value in the increasing direction, the luminance value is changed in a decreasing direction (for example, decreasing by 8 in the luminance value) to check whether the similarity is improved. If no improvement is observed, the result before the re-binarization process is adopted.

  Step S13 in FIG. 11 is a re-binarization process, and a process similar to that in step S6 in FIG. 2 is performed. Further, the second binarization threshold value, the third binarization threshold value, etc. in step S13 may be increased or decreased by 8 luminance values as described in step S6, or values other than 8. May be.

[Main effects of the embodiment]
As described above, according to the character recognition method according to the present embodiment, a character string composed of a known number of characters printed at a known position on the recording medium 2 is subjected to main scanning with the one-dimensional image sensor 11 and is manually or In the character recognition device 1 which converts the two-dimensional image information by performing sub-scanning by moving the recording medium 2 by machine drive, and recognizes the character string by appropriately processing the image data. A step of binarizing the multi-valued image scanned on the recording medium 2 (step S1 in FIG. 2), a step of obtaining a circumscribed rectangular region circumscribing the (recognition target) character, and extracting a feature amount in the circumscribed rectangular region ( 2 (step S2, step S3), a step of calculating the similarity with the reference feature quantity (step S4 of FIG. 2), and a step of determining the first character by comparing the similarity with the similarity threshold (FIG. 2 S5) and, based on the determination result, when there are a plurality of candidate characters, a separation evaluation value for separating and evaluating each of the plurality of candidate characters and the (recognition target) character is obtained. By configuring the candidate character by comparing with the separation evaluation threshold (step S7 in FIG. 2), another feature amount (separation) can be obtained for those that cannot be distinguished by the candidate character determination in the first character determination. It is possible to make a determination based on (evaluation value) and to improve the reliability of the recognition result.

Further, in the case where the determination cannot be made in step S5 of FIG. 2 described above, a step (step 6 of FIG. 2) for performing the re-binarization process on the (recognition target) character is configured, thereby performing the re-binarization. By configuring so that the (recognition target) character is output to the feature extraction step (step 4 in FIG. 2), the reading accuracy of the (recognition target) character that could not be determined can be further increased.
Similarly, in the separation processing step (step S7 in FIG. 2), the characters that could not be discriminated (recognition target) are output to the re-binarization processing unit 13g and re-binarization processing is performed, whereby re-binary processing is performed. The accuracy of separation between the (recognition target) character obtained by the conversion and the reference character with similar similarity can be increased, and as a result, the reliability of the recognition result can be increased.

[Other Embodiments]
Further, in the separation condition storage unit 13j described above, the separation evaluation is performed with the number “0 (zero)” and the alphabet character “O (o)”, and two candidate characters that are similar in degree of similarity. Also stored are calculation formulas of candidate characters, such as characters, in which the similarity between the numbers “6” and “8” and the alphabet character “B” is approximate. For example, as shown in FIG. 15, using the value DA ′ at the upper right of the region divided by 5 × 5, the value DB ′ below it, the value DC ′ at the center, and the value DD ′ at the lower left, FIG. The separation evaluation value is calculated using the calculation formula shown in FIG. 11 (step S11 in FIG. 11). The separation evaluation value decreases with the candidate character “6”, but increases with the candidate character “B”. Therefore, the discrimination logic is as shown in FIGS. 16B and 16C (step S12 in FIG. 11). K1 ′ and K2 ′ in FIG. 16B are separation evaluation threshold values and can be obtained experimentally. If it is within the range between K1 ′ and K2 ′, it is determined as a candidate character “8”.

  As described above, after the circumscribed rectangular area (in the area surrounded by the one-dot chain line in FIG. 6) is divided into 5 × 5 areas in the feature extraction step (step S3 in FIG. 2), the separation processing step (FIG. 2). In step S7), at least two or more other feature quantities (six A to F in FIG. 13, four A to D in FIG. 15) in the divided region are extracted (separation evaluation values are obtained). As a result, the accuracy can be adjusted by adjusting the number of areas to be divided. For example, in FIG. 13, the separation evaluation value can be obtained without E and F, but the numbers “0 (zero)” and the alphabet letter “O (o)” are present when E and F are present. The difference is clear (the value of fx is different), and the accuracy can be further improved.

  In the present embodiment, when the re-binarization process (step S6 in FIG. 2 and step S13 in FIG. 11) is performed, the binarization threshold for re-binarization is changed so that the similarity is improved. However, for example, a basic statistic of similarity regarding all the determined characters may be calculated, and the re-binarization process may be performed based on the calculated basic statistic. That is, the character recognition method according to the present embodiment includes a basic statistic calculating step for calculating a basic statistic of similarity for all determined characters based on the first character determining step, and the basic statistic calculating step. And re-binarization processing step for performing binarization processing based on the basic statistic calculated in step (2), and setting an optimal binarization threshold using the basic statistic Processing can be performed, and the reliability of the recognition result can be further improved. The basic statistics here are, for example, the total number of similarities, the maximum value, the minimum value, the average value, the standard deviation, and the like. When the minimum value is adopted as the basic statistic, the minimum value is changed. When the variance or standard deviation is adopted as the basic statistic, the values are changed so as to be reduced. . In this way, if the threshold value is changed so that the basic statistic is improved, in addition to improving the accuracy of each character to be recognized based on the separation evaluation value, it greatly contributes to improving the accuracy when considering the entire character string. be able to.

  Also, by repeating the re-binarization process based on the basic statistic of similarity, it is possible to appropriately set a binarization threshold for re-binarization, contributing to the efficiency of character recognition. be able to.

  In the present embodiment, the one-dimensional image sensor 11 and the linear transport mechanism are used. However, the present invention is not limited to these. For example, a combination of an area sensor such as a two-dimensional CCD or a CMOS imager and a subject support mechanism may be used. In addition, the recognition target can be applied not only to printed characters but also to handwritten characters. Further, it can be applied not only to character recognition but also to decoding 1D and 2D barcodes.

  The character recognition method and the character recognition device according to the present invention are useful for enhancing the reliability of character recognition.

It is a block diagram which shows the electric constitution of the character recognition apparatus which concerns on embodiment of this invention. It is a flowchart which shows the character recognition method which concerns on embodiment of this invention. It is a figure which shows image data when the character printed on the OCR character storage area of the recording medium is imaged. This is image data when attention is paid to characters (part) constituting the character string. It is the vertical projection data obtained by carrying out the vertical projection of the image data shown in FIG. It is a figure which shows the image data of the character "<" which comprises a character string. It is a figure which shows an example of a circumscribed rectangle area | region. It is a figure which shows the brightness | luminance in the partial area when dividing | segmenting into a 5 * 5 area | region. It is a figure which shows the ratio of the number of black pixels which occupies for the total number of pixels in each partial area. It is a figure which shows an example of a reference | standard feature vector. It is a flowchart for demonstrating the flow of the detailed information processing of a separation process. It is a figure which shows the specific example of the image used as the object of a separation process. It is explanatory drawing for demonstrating calculation of a separation evaluation value. It is a figure which shows the calculation formula for calculating a separation evaluation value. It is explanatory drawing for demonstrating calculation of another isolation | separation evaluation value. It is a figure which shows the calculation formula for calculating another isolation | separation evaluation value. It is explanatory drawing for demonstrating the specific example regarding the necessity determination of rebinarization.

DESCRIPTION OF SYMBOLS 1 Character recognition apparatus 2 Recording medium 11 Image pick-up element 12 Image memory 13 Data processing part 13a Binarization process part 13b Character extraction part 13c Feature extraction part 13d Character recognition part 13e Feature dictionary storage part 13f Judgment part 13g Rebinarization process part 13h Separation processing unit 13i Character output unit 13j Separation condition storage unit

Claims (2)

A binarization processing step for binarizing image data obtained by imaging a medium;
A character extraction step for obtaining an area including a recognition target character from the binarized image data;
A feature extraction step of extracting a feature vector of a recognition target character in the region;
Comparing the feature vector extracted in the feature extraction step with each given reference feature vector for each of a plurality of reference characters, and calculating a similarity with each reference feature vector;
A first character determination step of determining candidate characters by comparing a plurality of similarities calculated in the similarity calculation step with a similarity threshold;
Based on the first character determination step, when each similarity of a plurality of candidate characters clears the similarity threshold, in order to separate and evaluate each of the plurality of candidate characters and the recognition target character A separation processing step for obtaining a separation evaluation value by a combination of feature vectors in at least two or more divided regions in the region divided into n × m divided regions ;
Wherein comparing the separate evaluation value respectively given separation evaluation threshold, to include a second character judgment step of judging candidate characters by the magnitude relationship between the separation evaluation value and the predetermined separation evaluation threshold Character recognition method. A binarization processing unit that binarizes image data obtained by imaging a medium;
A character cutout unit for obtaining an area including a recognition target character from the binarized image data;
A feature extraction unit for extracting a feature vector of a recognition target character in the region;
A character recognition unit that compares the feature vector extracted by the feature extraction unit with each given reference feature vector for each of a plurality of reference characters, and calculates a similarity to each reference feature vector;
A determination unit that determines a candidate character by comparing a plurality of similarities calculated by the character recognition unit with a similarity threshold;
Based on the determination unit, when each similarity of a plurality of candidate characters clears the similarity threshold, in order to separate and evaluate each of the plurality of candidate characters and the recognition target character , n × m The separation evaluation value is obtained by a combination of feature vectors in at least two or more divided regions in the divided region, and each of the separation evaluation values is compared with a predetermined separation evaluation threshold, and the separation evaluation value and character recognition apparatus characterized by comprising a separation processing unit that determines candidate characters by the magnitude relationship between the predetermined separation evaluation threshold.

Images