JP5107100B2 - Character recognition method, character recognition device, and character recognition program - Google Patents

Description

  The present invention relates to a character recognition method, a character recognition device, and a character recognition program for recognizing characters included in an image.

  A general-purpose optical device, such as a camera-equipped mobile phone, is close to the environment where images can be taken anywhere, and there is a need for character recognition using these. In addition, mobile phone features include screen display, computation, and network functions, and software that uses them can be built in independently. It is possible to take actions such as.

On the other hand, as a disadvantage compared to the existing methods of these devices,
・ Image quality is not good (distortion and resolution is low)
・ Image quality varies greatly depending on the model. ・ Shooting environment is sparse, shooting magnification, light source, shooting direction cannot be fixed, and there is no or insufficient camera shake correction. It can be seen that the recognition is very difficult compared to capturing.
JP 2003-256429 A

  In general, the worse the state of the image, the more difficult it is to recognize characters, and the more calculation is required to recognize characters.

  As a conventional method of character recognition using a mobile phone, there is a method in which an image taken by a mobile phone is transmitted from the mobile phone to a server, and character recognition is performed by a server having a higher computing power than the mobile phone. The image tends to increase in size, so that the transmission time is long and the communication cost increases. Also, the method of performing character recognition with a mobile phone has a problem that the calculation capability is lower than that of the server, so that accuracy is significantly reduced when speed is prioritized, and speed is significantly decreased when accuracy is prioritized.

  Therefore, the present invention can recognize a character with a small amount of calculation and high accuracy even for an image in a bad state, and can further reduce the amount of data to be transmitted to the server, a character recognition device, and a character The purpose is to provide a recognition program.

According to the present invention, a pixel group that is determined to include a single character and that is selected one by one from among pixels that are included in a pixel group that includes binarized pixels, the pixel group A statistical processing step for performing statistical processing on the pixels included in the pixel, and in the statistical processing step, the attention is made from the a × b pixel in the region of vertical a pixel × horizontal b pixel centered on the currently focused pixel. The (1 + c) pixel including the selected pixel and another arbitrary c pixel is selected. From the _{a × b−1} C _c combinations, d combinations whose patterns overlap with the case where the other pixels are the center are selected. The _{a × b−1} C _c −d combinations excluding the first dimension, and the values of all pixels from 2 ^{(1 + c)} combinations of possible values of the (1 + c) pixels including the pixel of interest. 2 but excluding the combination of two types are identical ^{( + C)} providing a two-dimensional array of combinations of types -2 and the second dimension _{(a × b-1 C c} -d) × (2 (1 + c) -2), each pixel included in the pixel group The distribution of the element values of the two-dimensional array corresponding to the pixel group is obtained by increasing the value of the corresponding element of the two-dimensional array by 1 for all the pixels included in the pixel group, and Determining that a character having a distribution of element values of the two-dimensional array closest to the distribution of element values of the two-dimensional array obtained for the pixel group is a character included in the pixel group A character recognition method comprising the steps of:

Further, according to the present invention, the pixel group determined to include one character, and the pixel of interest is selected one by one from among the pixels included in the pixel group composed of binarized pixels, Statistical processing means for performing statistical processing on the pixels included in the pixel group, wherein the statistical processing means includes a × b pixels in a region of vertical a pixel × horizontal b pixel centered on the pixel of interest. From the _{a × b−1} C _c combinations that select (1 + c) pixels including the pixel of interest and other arbitrary c pixels, there are d patterns whose patterns overlap with the case where the other pixels are the center. The _{a × b−1} C _c −d combinations excluding the combinations are the first dimension, and all the pixels from the 2 ^{(1 + c)} combinations of possible values of the (1 + c) pixel including the currently focused pixel are included. values excluding the combination of two types are the same 2 ^{(1 c)} providing a two-dimensional array of combinations of types -2 and the second dimension _{(a × b-1 C c} -d) × (2 (1 + c) -2), each pixel included in the pixel group The distribution of the element values of the two-dimensional array corresponding to the pixel group is obtained by increasing the value of the corresponding element of the two-dimensional array by 1 for all the pixels included in the pixel group, and Determining that a character having a distribution of element values of the two-dimensional array closest to the distribution of element values of the two-dimensional array obtained for the pixel group is a character included in the pixel group And a character recognizing device.

  Furthermore, according to the present invention, there is provided a character recognition program for causing a computer to execute the above character recognition method.

  According to the statistical processing step of the present invention, characters can be recognized with a small amount of calculation and high accuracy from the element values of the two-dimensional array, and the distribution of the element values of the two-dimensional array is transmitted from the mobile phone to the server. When the determination step is performed by the server, the amount of data transmitted from the mobile phone to the server can be reduced.

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

  Character recognition in the present invention is to select an image most similar to a photographed image from images registered in the server (this is called search), and the character written in the registered image is the recognition result. This is a kind of image recognition.

  It is assumed that the character string to be searched is surrounded by a rectangular frame. By recognizing the frame, the rotation angle and the enlargement / reduction ratio can be determined, and the position of the character in the frame can be made constant. As a result, the degree of freedom of shooting increases, such as enabling recognition even if the camera rotation, distance to the object, and position are not fixed. Unlike optical scanners, it is particularly useful to have a degree of freedom in photographing in optical devices that can be photographed in various postures (distance from the subject, presence / absence of rotation, blurring, etc.) like a mobile phone.

  The search is performed by comparing image feature amounts, which are the results of analyzing the captured image and the registered image. The comparison result of the image feature amount between the captured image and the registered image is called a distance. The smaller the distance, the more similar the two images are. Therefore, the image having the minimum distance value is the search result image. What will be used as the image feature amount will be described later.

  The process up to the image feature extraction includes a first half of extracting a character string image having a certain size from the image and a second half of calculating the image feature.

  The extraction of the character string image of the first half part will be described.

  First, the photographed color image is converted into a gray scale image. Since this conversion is a general process, this algorithm will not be described.

  Next, the frame area is recognized from the gray scale image. The frame area is an area surrounded by a rectangular frame. A character string is inserted in the frame. Since this recognition method uses an invention different from the present invention, a specific description of the recognition method is omitted.

  As a result, the x and y coordinates of the four vertices are obtained.

  If this coordinate can be taken at the outermost point or the inner point of the frame area, and if a certain coordinate can be taken, the process of expanding the next frame can be omitted, but here the frame is roughly recognized and A slightly larger range is used as the frame area.

  Regarding the process of expanding the frame, first, the coordinates of the vertex are corrected so as to point outside the frame as much as possible. Look at the two sides that touch each point, and decide the direction to spread from here. The frame is expanded toward the direction of the vector obtained by adding the vector in the vertex direction of each side, and this is represented by (+, 0, −) of the x and y components. In the conversion, like the correction direction in the (x, y) coordinates for each vector, the previous vector direction is expressed by the positive and negative directions of the x and y components. If the current vertex is (X, Y) and the vector is x + and y is-, it is checked whether (X + 1, Y) is a frame for the x component, and then (X, Y-) for the y component. Check 1). If either is a frame pixel, the vertex coordinates are moved to that coordinate. The same processing is performed for the vertex coordinates until the coordinates stop moving. Do this for all vertices.

  Finally, a constant value is added / subtracted in the vector direction for each vertex. By doing so, the quadrilateral connecting the vertices becomes larger than the frame.

  Next, an image inside the frame (vertex) is extracted. If the frame is rotating, there will be white space outside the frame. The margin is filled with gray (160 here).

  At this point, it is relatively easy to separate the black part with the frame and characters from the white part as the background. Therefore, the binarization threshold value is calculated from the image by a discriminant analysis method (Otsu's method) or the like. At this point, the image is not converted into a binarized image and remains a grayscale image.

  Next, a rectangular shape is recognized from the image, and it is examined how many times the rectangle is rotated.

  As the rotation angle, 0 to 90 degrees is found. The reason why the angle is not 0 to 360 degrees is that the top and bottom of the frame is not known and the determination of the vertical and horizontal sides cannot be made. However, when writing characters in a normal frame, the width is longer than the height. By making this a rule for all search target character strings, the length and breadth are found from the comparison of the lengths of the sides, and the rotation from 0 degrees to 180 degrees is found.

  If the rotation angle is known, the image is rotated in the reverse direction by the angle. As a result, the rotation direction of the frame is constant. However, the possibility of upside down remains for the reason mentioned above.

  Next, the image is enlarged or reduced so that the size of the frame becomes a predetermined size.

  The size of the frame after enlarging or reducing the image is fixed or selected from a plurality of candidates. In the latter case, a candidate having the closest frame aspect ratio is selected. It should be noted that the size of the search image and the size of the registered image must be the same.

  Next, the frame is removed from the image. For this purpose, several pixels are removed from the top, bottom, left, and right outside the image. The number of pixels to be hollowed out is determined by the size selected in the previous scaling.

  Next, separate the characters one by one. If there is a space of 1 pixel or more between characters in the binarized image, check whether it overlaps the character by scanning in the vertical axis direction for all pixels in the vertical line If a vertical line that overlaps with a character continues, the character is divided by a method of recognizing the entire continuous vertical line as an area of one character. One character area extends from a certain vertical line to the next vertical line.

  However, depending on the photographed image, characters may be connected correctly when binarized because the performance of the camera lens is low, the subject is out of focus, camera shake, or the character spacing of the subject is narrow. May not be separated. Therefore, an area of one character is detected by another method described below.

  First, for all the vertical axes, the minimum value of the pixels on each axis is obtained. Assign this value to the array of image width lengths.

  This array has a feature that a value is small in a portion where there is a character and a value is large in a portion where there is no character. If the values of the array are viewed continuously, it can be seen that the waveform from the top of the wave to the next top corresponds to one character. The sequence of peaks (tops), valleys (bottoms), and peaks (tops) of this waveform is a single character. Therefore, the character delimiter can be determined by finding the valley and the mountain from the array.

  First, explain how to find the valley. First, determine the starting point of the sequence. If the search is for the first character, the start of the array is taken as the starting point. Otherwise, the starting point is the mountain that delimits the previous character. The value of the starting point array is held as a provisional minimum value. Next, the values of the array of points are referred to in order from this starting point. Looking at the value of the array, if this is smaller than the provisional minimum value, this value becomes the provisional minimum value. If the array reference is advanced for a while, the value becomes larger. This is because the area around the minimum value is a character area, and the value increases when the character area ends and the character breaks. If the difference between the array value and the minimum value exceeds a certain threshold value, it can be determined that the character and the character delimiter have been confirmed, and therefore the minimum value is determined to be a valley. With this method, even when an array of a region without characters is referred to, the valley cannot be determined because there is almost no difference from the minimum value, and it can be assumed that characters do not exist in this region.

  Next, explore the mountains. The search for the mountain is almost the opposite of the search for the valley. The starting point is the valley point found earlier. This value is the provisional maximum value. Refer to the array in order and find the maximum value. In addition, the value decreases rapidly after a while. This is because the area around the maximum value is a break of characters, and the reason for the decrease is that the character area starts. If the difference between the maximum value and the value of this array exceeds a certain threshold value, it can be determined that the presence of the character and the character delimiter has been confirmed, and thus the maximum value is determined as a peak.

  The search for valleys and mountains as described above is performed alternately, and from the result, it can be determined that the area from one mountain to the next is an area of one character.

  This character area may include a background before and after the character. Therefore, when the value of the previous array is larger than the binarization threshold, the area is excluded from the character area as a background. As a result, the character region extends from the left end to the right end.

  When the characters are separated, binarization is performed using a binarization threshold.

  As a result of the processing so far, only one character can be extracted from the image. From here, image processing is performed on each character image.

  Although the character image is a binary image, it is difficult to cut out the outline of the character, and when the character and the background are separated by the binarization threshold, the outline often becomes rattled. This causes an image feature amount related to the relationship between the three pixels, which will be described later, to vary from shooting to shooting, and adversely affects the accuracy of character detection. Therefore, I want to make the boundary as gentle as possible.

  As a method for this, a median filter is applied or a thick line is applied. Thereby, the rattling of the outline is somewhat reduced, and the accuracy of character detection can be increased.

  Next, the enclosed area is filled. The enclosed area is a space in characters such as “0” and “9”, for example. As a result, the difference in image between characters with similar outlines “6” and “9” becomes clear, and the vertical and horizontal histogram values of the image feature amount are stabilized, and the accuracy is improved.

  From here, an image feature amount acquisition method, which is the latter half of character recognition, will be described. There are two types of image feature amounts for each character: a histogram of the relationship between three pixels and a histogram in the vertical direction and the horizontal direction.

  A histogram of the relationship between the three pixels will be described.

Attention is paid to a total of 3 pixels, that is, a certain pixel in the image and 2 pixels in the surrounding 8 directions. There are 28 ways of ₈ C _{2 to} take 2 pixels in 8 directions, but 8 of them refer to the same 3 pixels as when centering on adjacent pixels, so this overlapping 8 ways are omitted A total of 20 patterns are seen (see FIG. 1).

  There are eight possible combinations of values between these three pixels, but pay attention to six patterns excluding two patterns in which all three pixels have the same value. The reasons for excluding the two patterns all represent the background and the pattern that becomes the character, that is, it can be said that the relation to the character area is strong. However, the character area changes greatly every time an image is taken. For example, the character width “l” does not become constant every time shooting is performed. It is assumed that the character width in a certain shooting is 4. When the height is the same in another shooting but the character width is 5, the character area is greatly different from 4: 5 even though the same object is shot. Such a difference in area frequently occurs, and it is difficult to process an image so as to be constant for each photographing. Especially when the image rotates, it is not constant. Therefore, two patterns related to the character area are excluded.

  The relationship between the three pixels captures the feature of the outline of the character. By excluding the above two patterns, 0 and 1 are included between the three pixels. That is, an edge is included in the surrounding 3 × 3 pixels. By associating this with the positional relationship of the three pixels and recording the feature, it can be said that it is closely related to the contour shape.

  This is performed for all pixels in the image. Twenty patterns of the relationship between the three pixels are distinguished from each other, and six patterns of possible values are also distinguished. An array having a length of 120 per character is an image feature amount having a relationship between three pixels. If there are 5 characters, the image feature amount is an array of length 600. For each character, counting is performed every time a pattern appears for all pixels, and a 20 × 6 array histogram is created.

  The image feature extraction will be described in order from the beginning. All pixels of one character are scanned. Regarding a certain pixel, three pixels in the first pattern in the 20 patterns between the three pixels are seen. If all have the same pixel value, ignore it, otherwise count the corresponding histogram array. The sequences counted at this time are 1 to 6. In the second pattern, the array is numbered 7 to 12. Since this is performed for all 20 patterns, the arrangement of one character has a length of 120 (= 20 × 6). The second character array is numbered 121-240.

  A method of creating an array histogram when focusing on one character will be described with reference to FIG.

  The processes in and after step S103 are repeated for all the pixels included in one character area.

  In step S103, the processes in and after step S105 are repeated for all combinations i of the three pixels including the pixel of interest. However, from all the combinations, eight combinations that overlap with the same pattern of another pixel are excluded. Therefore, this repetition is performed 20 times.

  In step S105, it is determined which of the six patterns the pattern formed by the three pixels is.

  When the pattern formed by the three pixels is pattern 1, the value of array A (i, 1) is increased by 1 (step S107).

  When the pattern formed by the three pixels is pattern 2, the value of array A (i, 2) is increased by 1 (step S109).

  When the pattern formed by the three pixels is pattern 3, the value of array A (i, 3) is incremented by 1 (step S111).

  When the pattern formed by the three pixels is pattern 4, the value of array A (i, 4) is increased by 1 (step S113).

  When the pattern formed by the three pixels is pattern 5, the value of array A (i, 5) is increased by 1 (step S115).

  When the pattern formed by the three pixels is the pattern 6, the value of the array A (i, 6) is increased by 1 (step S117).

  The above is how to create an array histogram when focusing on one character.

  In addition to alphanumeric characters such as 0 to 9, A to Z, and a to z, which are used as a reference for character recognition, a histogram of an array is created in the same manner as described above. Therefore, the character having the histogram closest to the histogram of the character to be recognized (characters such as 0 to 9, A to Z, a to z, or any one of symbols, etc.) after the character recognition It becomes a character.

  Next, histograms in the vertical and horizontal directions are created. In other words, an array of lengths in the vertical direction of a character image and an array of lengths in the horizontal direction are prepared, and if there are black pixels that are characters, the corresponding array in both directions is counted, and the histogram Create two.

  Thereby, the image feature amount reflecting the character position can be acquired.

  In addition to this, the number of characters and the center x-coordinate of each character of the image before dividing the character are also acquired as image feature amounts. These are useful for improving the search accuracy and speeding up the search.

  As a postscript regarding search, a function that does not give an answer when a registered image does not exist is implemented. As a method, it is confirmed whether the distance to the registered image exceeds a certain threshold for each character. If there is a character exceeding the threshold, even if the distance is minimum, the image is not answered and re-photographed.

As another check function, the minimum distance and the second distance are:
(Second distance) / (First distance) <D
When the threshold value D is smaller than a certain threshold value D, there is a high possibility that it is mistaken for another image, so that the re-photographing is performed without outputting the answer.

  The comparison of histograms uses the Euclidean distance as in the relationship between the three pixels. However, a city block distance or Mahalanobis distance may be used instead of the Euclidean distance. The distance between the vertical and horizontal histograms is calculated separately for each character. When the distance between the registered image and the captured image is obtained, the histogram at the same array position is referred to. In addition, the distance obtained by shifting the array position on one side one by one is calculated. The smallest value among these three distances is selected as the formal distance. This takes into account a small image shift of less than one pixel.

  As the final distance, the total sum of the distance between the three pixels, the vertical histogram, and the horizontal histogram of all characters is used. Alternatively, the distance between the three points, the distance of the vertical histogram, and the distance of the horizontal histogram may be weighted and then summed. Character recognition is performed using only the distance between the three pixels, and when a plurality of candidates having a close distance between the three pixels appear, a vertical histogram and a horizontal histogram may be used.

  The character string to be recognized may be limited, for example, such that the first n characters are uppercase alphabets and the next m characters are numbers. In such a case, character recognition may be performed sequentially from the first character, and if the number is recognized by the i-th character smaller than n, the error may be terminated.

  Further, in order to reduce the number of comparisons at the time of search, the search by the print definition that will be described later when explaining image registration is also omitted. In the print definition, information on characters in the same format is written. For example, frame candidate 1 is used between 5 and 10 characters. At the time of image search, the comparison with the printed matter definition is performed before the character comparison and the character position comparison. Compare the number of characters with the frame candidates. If either of them is different, you can see that there is no answer in the registration data using this print definition, so you do not have to compare it with the registered image using this print definition. .

  Next, image registration will be described.

  There are two types of images: a method using a digital image used for printing and a method using an actual captured image.

  Registered images are sorted by printed matter. The features for each printed matter are registered in advance in the DB. The registered features include a minimum maximum number of characters and a vertical and horizontal size of the frame. In addition, the name of the printed matter, information associated therewith, font type, character spacing, and character pattern (including upper and lower case letters of the alphabet) can be mentioned.

  When registering an image, one definition of the printed material is selected. A frame candidate having the closest aspect ratio is selected from the definition value. When extracting the image feature amount, enlargement / reduction is performed using the value of the frame candidate. Image feature extraction is basically the same as described above, but depending on the type of digital image to be registered, there is no need to rotate the image, and the position of the frame is fixed without performing frame extraction. Can be obtained from These pieces of information are also determined in advance in the print definition.

  By these processes, the number of characters of the image can be obtained. If this is compared with the definition, if it is different, it is found that there is an error in the definition of the image or the printed matter. This can also be said to compare the vertical and horizontal lengths of the frame and the print definition when performing frame recognition. That is, the print definition can also be used as an error check during registration.

  In the image registration, the normal position direction and two types of image feature values rotated by 180 degrees are registered. This is a measure for making it impossible to distinguish the upper and lower sides of the captured image, and it is possible to search regardless of the orientation of the captured image.

  The image feature amount data registered in the server includes a name for identifying an image and an image feature amount. Two types of rotation, 0 degree and 180 degree rotation, are registered. When providing a service, an event is specified when each registered image is selected. In the case of a mobile phone, you can access WEB to access digital contents such as music and videos, receive emails, perform operations such as sound and vibration functions, and generate new contacts and schedules. Or make a phone call.

  Up to this point, the description has been made with reference to the schematic diagram, but another character recognition mode will be described.

  This form is the same up to the point where the image feature amount is obtained, but the search is also carried out in the mobile phone. Here, an image feature amount for each character is provided in the mobile phone. One character of the image feature value of the image analysis result is compared with the image feature value in the mobile phone, and the character with the smallest distance is applied to the character string of the photographed image. This is performed for all analysis result characters, and as a result, recognition of the character string of the photographed image is completed.

  The character recognition method described above is realized by hardware, software, or a combination thereof.

  In the above description, the relationship between three pixels is an image feature amount, but a relationship between four pixels or a relationship between more pixels may be an image feature amount. Also, instead of selecting a pixel for calculating the image feature amount from a 3 pixel × 3 pixel region centered on the target pixel, the image feature amount is calculated from a 5 pixel × 5 pixel region or a larger region. You may make it select the pixel for doing.

  In the above description, a mobile phone is taken as an example, but the present invention is not limited to this.

It is the figure which enumerated the combination which selects three pixels including a center pixel from the 3 * 3 area | region by embodiment of this invention. It is a figure which shows the method for calculating | requiring the feature-value of the image containing the character by embodiment of this invention.

Claims (9)

A pixel group that is determined to include a single character, and selects one pixel of interest from among pixels included in the pixel group that includes the binarized pixels, and the pixel group includes the pixel group. A statistical processing step for statistical processing;
The statistical processing step includes the pixel of interest and other arbitrary c pixels from a × b pixels in a vertical a pixel × horizontal b pixel region centered on the pixel of interest currently (1 + c) ) a combination of _{a × b-1 C} c _-d Street when the other pixels with the center of a combination of street _{a × b-1} C _c and pattern excluding the combination of d as duplicate for selecting the pixel first 2 ^{(1 + c)} − which is one dimension and excludes 2 ^{(1 + c)} combinations of values that can be taken by (1 + c) pixels including the currently focused pixel from 2 ^{(1 + c)} combinations where the values of all pixels are the same. A two-dimensional array of ( _{a × b−1} C _c −d) × (2 ^{(1 + c)} −2) having two combinations as the second dimension is provided, and each pixel included in the pixel group corresponds. Increasing the value of an element of a two-dimensional array by 1 is included in the pixel group By performing the pixel of Te, we obtain a distribution of the values of the elements of the two-dimensional array corresponding to the pixel groups,
Furthermore,
A determination step of determining that a character having a distribution of element values of the two-dimensional array closest to the distribution of element values of the two-dimensional array obtained for the pixel group is a character included in the pixel group; A character recognition method comprising: The character recognition method according to claim 1,
A vertical distribution calculation step for obtaining a vertical distribution of values obtained by adding the values of the pixels included in the pixel group in the horizontal direction for each vertical position;
A horizontal distribution calculation step of obtaining a horizontal distribution of values obtained by adding the values of the pixels included in the pixel group in the vertical direction for each horizontal position;
Further comprising
In the determining step, the distribution of element values of the two-dimensional array obtained for the pixel group, the vertical distribution, the horizontal distribution, and the distribution of element values of the two-dimensional array of each candidate character, the vertical A character recognition method comprising: determining a character included in the pixel group based on a distance between the distribution and the horizontal distribution. The character recognition method according to claim 1 or 2,
Character string candidates obtained by performing character recognition for each character are limited to a plurality of predetermined character strings,
An error that terminates with an error if a part of a character string that is not included in any of a plurality of predetermined character strings is obtained during character recognition for each character sequentially. A character recognition method further comprising an end step. The character recognition method according to any one of claims 1 to 3,
A character recognition method, wherein a = 3, b = 3, c = 2, and d = 8. A pixel group that is determined to include a single character, and selects one pixel of interest from among pixels included in the pixel group that includes the binarized pixels, and the pixel group includes the pixel group. Provide statistical processing means for statistical processing,
The statistical processing means includes the pixel of interest and another arbitrary c pixel from a × b pixels in a region of vertical a pixel × horizontal b pixel centered on the pixel of interest at present (1 + c ) a combination of _{a × b-1 C} c _-d Street when the other pixels with the center of a combination of street _{a × b-1} C _c and pattern excluding the combination of d as duplicate for selecting the pixel first 2 ^{(1 + c)} − which is one dimension and excludes 2 ^{(1 + c)} combinations of values that can be taken by (1 + c) pixels including the currently focused pixel from 2 ^{(1 + c)} combinations where the values of all pixels are the same. A two-dimensional array of ( _{a × b−1} C _c −d) × (2 ^{(1 + c)} −2) having two combinations as the second dimension is provided, and each pixel included in the pixel group corresponds. Everything contained in the pixel group to increase the value of an element of a two-dimensional array by 1 By performing the pixel, we obtain a distribution of the values of the elements of the two-dimensional array corresponding to the pixel groups,
Furthermore,
Determination means for determining that a character having a distribution of element values of the two-dimensional array closest to the distribution of element values of the two-dimensional array obtained for the pixel group is a character included in the pixel group A character recognition device comprising: The character recognition device according to claim 5,
Vertical direction distribution calculating means for obtaining a vertical distribution of values obtained by adding the values of the pixels included in the pixel group in the horizontal direction for each vertical position;
Horizontal direction distribution calculating means for obtaining a horizontal direction distribution of values obtained by adding the values of the pixels included in the pixel group in the vertical direction for each horizontal position;
Further comprising
The determination means includes a distribution of element values of the two-dimensional array obtained for the pixel group, a distribution of element values of the two-dimensional array of the vertical distribution and the horizontal distribution and each candidate character, A character recognition apparatus, wherein a character included in the pixel group is determined based on a distance between a vertical distribution and the horizontal distribution. In the character recognition device according to claim 5 or 6,
Character string candidates obtained by performing character recognition for each character are limited to a plurality of predetermined character strings,
An error that terminates with an error if a part of a character string that is not included in any of a plurality of predetermined character strings is obtained during character recognition for each character sequentially. A character recognition apparatus, further comprising an ending unit. The character recognition device according to any one of claims 5 to 7,
A character recognition device, wherein a = 3, b = 3, c = 2, and d = 8.   A character recognition program for causing a computer to execute the character recognition method according to claim 1.

Images