JP2004094427A - Slip image processor and program for realizing the same device - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a pattern recognizing method and a device for realizing highly precise pattern recognition by fetching the surface of a document as a concentration image, and generating a binary image suitable for the recognition of characters, symbols, and marks even when the document image is constituted of various color frame lines, pre-print characters, backgrounds, entry characters, symbols and marks. <P>SOLUTION: A document such as a slip is inputted as an image, and a frame structure described in the document is extracted on the inputted document image, and the inside of the document image is divided into a plurality of areas based on the acquired frame structure, and a binary threshold suitable for recognizing characters, symbols, and marks described in the frame is calculated for each of the divided areas, and a binary image is generated for each of the divided areas by using the obtained threshold, and the recognition of the characters, symbols, and marks is executed on the basis of the generated binary image. <P>COPYRIGHT: (C)2004,JPO

Description

[0001]
TECHNICAL FIELD OF THE INVENTION
The present invention relates to a pattern recognition method and a pattern recognition device that captures the surface of a document as a gray-scale image and recognizes patterns such as characters, symbols, and marks from the image.
[0002]
[Prior art]
An example will be described in which characters are recognized in pattern recognition of characters, symbols, marks, and the like. The same applies when recognizing a symbol or mark pattern.
In general, when character recognition is performed from a gray image of a document, binarization processing is performed so that the character part is black and the background is white, and the character is cut out from the binary image generated by this processing, and the character is extracted. Characters are identified from the form There are various binarization methods, which can be classified from two viewpoints. One is a processing unit for determining a threshold, and typical examples are (1) the entire designated area (entire area) in the image, and (2) a minute area near the pixel of interest. (Local area), and (3) a method in which an image is divided into meshes (mesh area) as processing units. The other is a feature amount for determining a threshold value, and typical ones use (a) frequency distribution of gray values, (b) average, median, maximum, and minimum values of gray values. There is a way. {Circle around (1)} When the whole area is used as the processing unit, Taiichi Saito and Hirozo Yamada, "Threshold Selection Method Using Weighted Average of Two Mean Midpoints and Data for Binary Evaluation" Transactions of IEICE , D-2, Vol. J83-D-2, No. 2, pp. In the method described in 575-583, it is general to use a discriminant analysis criterion or a binarization method based on the k-means method based on the feature amount of (a). {Circle around (2)} When a local region is used as a processing unit, Ovid Deu Tier, Anbil K. et al. Jain, "Goal-Directed Evaluation of Binarization Methods", IEEE Trans. See Pattern Analysis and Machine Intelligence, vol. 17, no. 12, pp. 1191-1201, 1995. As described in (1), binarization is generally performed using the feature amount (average value) of (b) as a threshold value. FIG. 11 shows an example. {Circle around (3)} When the mesh area is used as a processing unit, as described in JP-A-6-4706, the entire surface is binarized once using the feature amount (average value) of (b), and characters are cut out from the binary image. Is performed, the size of one character is estimated from the character segmentation result, the character is divided into regions (mesh shape) for each character, and the characteristic amount (average value) of (b) is newly calculated for each divided region. There is a method of binarizing by using this.
In the case of recognizing all characters in a frame in a document in a document image including a lot of entry frames such as a form, the user inputs the document image as shown in FIG. Generates a binary image of the entire surface (302), extracts a frame structure from the generated binary image (303), divides the binary image into regions for each frame (304), and cuts characters for each frame region Then, character recognition is performed (305). Then, when reading only characters in a part of the plurality of frames in the document image, a document image as shown in FIG. 4 is input as disclosed in JP-A-2000-293629 ( 401), a form position in the input image is detected (402), form expansion / contraction at the time of scanning is detected, and a grayscale image is prepared based on frame position information of a form image of the same format prepared in advance. The area of the frame to be read within is estimated and cut out (403). Then, binarization is performed on the cut-out area (cut-out area) using any of the above methods (405), and a frame structure is extracted from the obtained binary image (406). Is divided into regions for each frame (407), and character recognition is performed (408). However, there are variations in the size of the form paper, errors in detecting the position of the form in the image, and errors in detecting expansion and contraction of the form when taking in the scanner, so that the entry frame and characters in the entry frame may protrude from the cutout area. For this reason, it is necessary to cut out a region that is slightly wider than the estimated frame position.
[0003]
[Problems to be solved by the invention]
An example will be described in which characters are recognized in pattern recognition of characters, symbols, marks, and the like. The same applies when recognizing a symbol or mark pattern.
Some document images such as forms are composed of various color frame lines, preprinted characters, backgrounds, and entered characters. When such a form is input as a multi-valued image and characters in some of the frames in the multi-valued image are recognized, the reading area (601) is, for example, as shown in FIG. ), (602) written characters (light), (603) frame lines (dark), (604) preprinted characters, (605) frame lines (light), (606) background portion (dark), It consists of a background part (thin) of (607). Consider a case where the reading area is binarized. When binarization is performed using the whole area as a processing unit as described in (1) of the prior art, the gray level distribution in the reading area is the gray level distribution of (601) in FIG. 6 as shown in FIG. 7 (701). 6, (703) which is the gray scale distribution of (602) in FIG. 6, (703) which is the gray scale distribution of (603) in FIG. 6, and (704) which is the gray scale distribution of (604) in FIG. 6 (705) which is the gray scale distribution of (605) in FIG. 6, (706) which is the gray scale distribution of (606) in FIG. 6, and (707) which is the gray scale distribution of (607) in FIG. Then, the binarization threshold values are the threshold value A (801) and the threshold value B (802) shown in FIG. However, the binarization result based on the threshold value A is shown in FIG. 9, and the character “630” to be read cannot be extracted. Since only the other characters are extracted, the character cannot be recognized. FIG. 10 shows a binarization result based on the threshold value B. The target character is painted over a dark background portion, and the character cannot be recognized. As described above, according to the method of the prior art (1), when there are various shades of frame lines, pre-printed characters, backgrounds, and entered characters in the reading area, the grayscale distribution is complicatedly overlapped. There is a problem that it becomes difficult to estimate a threshold value that can distinguish the surrounding background. When binarization is performed using the local region as a processing unit as described in (2) in the related art, the gray value (702) of the written character and the gray value (704) of the pre-printed character are low (dark), and the background color is low. When the gray value (706) is high (thin), as shown in FIG. 12, the input character and the pre-print character are both extracted, so that the input character and the pre-print character overlap (1201, 1202), and the character is correctly I can't recognize. Also, when the shade value (702) of the background color of the input character is low (dark), and the shade value (704) of the pre-print character and the shade value (706) of the background color are high (light) (however, the shade value of the pre-print character) Is lower than the density value of the background), as shown in FIG. 13, a part of the pre-printed character is missing, and the character cannot be correctly recognized. In addition, as shown in FIG. 14, when the gray value (703) of the frame line is low and the gray value (702) of the input character is high, the binarization result is as shown in FIG. I cannot recognize it correctly. When binarization is performed using the mesh area of (3) described in the prior art as a processing unit, the pitch and size of characters in a form are various, and written characters and preprinted characters may overlap. , It is difficult to divide each character (mesh shape). Therefore, in each mesh area, there may be cases where various shaded characters, preprinted characters, frame lines, and backgrounds are included, and the same problem as the binarization method of (1) and (2) occurs. The above problem also occurs when all the characters in the frame in the document are recognized.
Further, when a binary image of a recognition area is displayed on a display to confirm the result of character recognition, the above-described problem occurs, so that characters to be recognized are blurred or crushed, and a frame to be recognized is shifted. Therefore, it may be difficult to confirm the recognition result.
[0004]
The present invention has been made in view of such a problem of the related art, and includes a document image including various forms of frame lines, such as a form, a preprinted character, a background, and an input character. Even if there is, character recognition that generates a binary image suitable for recognizing characters entered in each frame, realizes highly accurate character recognition, and can easily confirm the recognition result visually. It is an object to provide a method and a character recognition device.
[0005]
[Means for Solving the Problems]
In order to solve the above-mentioned problem, a representative invention disclosed in the present application is to provide a character image from a grayscale image of a document image including a frame line of various colors, a preprinted character, a background, and an input character. This is a method for generating a binary image suitable for recognition. A grayscale image of the entire document or a part thereof is input (101), and a frame structure described in the document is extracted from the image (102). The input image is divided into a plurality of regions based on the obtained frame structure (103), a binarization threshold suitable for character recognition is calculated for each of the divided regions (104), and the obtained threshold is used. A binary image is generated for each divided region (105).
[0006]
Furthermore, in a method of recognizing a pattern of characters, symbols, marks, etc. from a document image such as a form, when recognizing a character in a specific frame in the document, input the surface of the document as a grayscale image, and By using the format (frame position) information prepared by detecting the position of a certain document and measuring the frame position in a document of the same format in advance, the characters to be recognized from the input image are described. Estimate the position where the frame exists, expand the estimated frame position up, down, left and right, cut out the enlarged area from the input image, extract the frame structure in the cut out partial area image, and fill in the frame for each frame It is characterized in that a binarization threshold suitable for recognizing a given character is calculated, a binary image is generated for each frame using the obtained threshold, and character recognition is performed for each frame.
[0007]
BEST MODE FOR CARRYING OUT THE INVENTION
Hereinafter, the present invention will be described with reference to the drawings. According to the invention disclosed in the present application, as shown in FIG. 1, a gray-scale image of the entire document or a part thereof is input (101), and the frame structure described in the document is extracted from the image (102). The input image is divided into a plurality of regions based on the frame structure (103), a binarization threshold suitable for character recognition is calculated for each of the divided regions (104), and the obtained threshold is used. A binary image is generated for each of the divided areas (105). According to the configuration of the present application, a binary image suitable for recognizing characters entered in each frame is generated even for a document image composed of various color frame lines, preprinted characters, backgrounds, and entered characters. And realizes highly accurate character recognition.
An overall flow of the present invention will be described in detail with reference to FIG. 26 which shows a processing flow of a method of reading characters in a form frame according to an embodiment of the present invention. First, in step 101, an image is input. The input image is a grayscale image, and is composed of input characters having various grayscale values, preprinted characters, a frame line, and a background as shown in FIG.
Next, in step 2601, a form position in the input image is detected. In this method, as shown in FIG. 28, the input image includes a form portion and a black background portion, and the position in the input image where the form exists is obtained. Specifically, one of the four corners of the input image is set as the origin, and the four corner coordinates of the form are obtained.
[0008]
Next, in step 2602, the area of the frame in which the character to be read is described is cut out from the input image. The position coordinates of the frame described in the form are measured in advance using a form in the same format as the form to be read, and stored as format information in the format information storage unit (209) in FIG. Then, based on the frame position coordinates of the format information, a frame region in which the character to be read is described is cut out. At this time, form expansion / contraction correction and inclination correction at the time of image capture are performed. However, due to errors in form position detection, form expansion / contraction correction, and inclination correction, and variations in the size of form paper, the four corners of the frame to be read cannot be accurately cut out, and the position of the held frame The position of the frame on the image to be read may be shifted. If the area is cut out using the held position information of the frame as it is, and the position is shifted, the character to be read may be missing and reading may not be possible. Therefore, in the frame area cutout processing, a range obtained by expanding the area estimated from the position of the form frame held in advance up, down, left, and right is cut out. This area is called a cutout area. FIG. 6 shows an example of the cutout area. Including the case (602), the frame at the center of the area is the frame to be read. The enlargement range is set in advance. For example, (A, B, C, D) is set in advance as a constant, and the area estimated from the position of the previously held form frame is Amm in the upward direction, Bmm in the downward direction, and Cmm in the leftward direction. , To the right to Dmm. With this configuration, the character to be read is completely included in the cutout area, and the character recognition accuracy is improved. Further, when displaying the image of the cut-out area for visual confirmation of the recognition result, the character to be read is completely included, so that the confirmation becomes easy.
Next, in step 102, the structure of the entry frame existing in the cutout area is extracted. To extract the structure of a frame means to calculate the position of each frame by detecting the position of a ruled line constituting the frame (FIG. 16) and detecting a closed area surrounded by the detected ruled line (FIG. 16). 17). Reference numerals 1701, 1702, 1703, 1704, 1705, 1706, 1707, and 1708 denote individual frame regions obtained by performing frame structure extraction. The position of the ruled line is detected by binarizing the entire surface once using the average value of the local region as a threshold as described in the prior art (2), and detecting horizontal and vertical black pixels from the obtained binary image. Do it by doing. However, as another method, a method of detecting from a multi-valued image using a Hough transform method as described in Shunji Mori and Tsugako Sakakura, “Basics of Image Recognition 2”, p3-11, Ohmsha, etc. may be used. it can. There are various methods for detecting the position of the frame, but the intersection of the detected horizontal ruled line and the vertical ruled line is detected, the detected intersection is traced, and a closed region having the intersection as a vertex is detected. A detection method can also be used.
[0009]
After the frame structure is extracted, in step 103, the inside of the shaded image cut out using the information on the frame structure obtained as shown in FIG. 18 is divided into regions for each frame. At this time, each frame area (single frame area) is divided into a frame area and an area surrounded by the frame area (an area where characters are written). This eliminates the need to consider the color of the frame line in the subsequent processing. Next, in step 104, a binarization threshold suitable for character recognition is calculated for each single frame area. A binary image suitable for character recognition is generated again without using the binary image generated for extracting the frame structure in step 102. The binarization threshold is calculated using the k-means method described in the related art. The k-means method is one of typical clustering methods, and is a method of dividing data having a numerical value or a numerical vector into a predetermined number of groups (clusters). In order to calculate a binarization threshold suitable for character recognition, each pixel in the region is clustered using a luminance value to distinguish a character, a preprinted character, and a background pixel. As described in the related art, there are various other methods for calculating the binarization threshold. For example, binarization can be performed by a method using discriminant analysis. However, the k-means method is a calculation method. There is an advantage that the amount is small and the processing time is short, and the accuracy is not inferior to the others. However, in a single frame area in a commonly used document image such as a form, there are two colors of a written character and a background and a preprinted character and a background like 1801 and 1807 as shown in FIG. In many cases, there are cases where there are three colors, that is, an input character, a preprinted character, and a background as in 1804. Therefore, a histogram of gray values in the frame area is created, and the obtained histogram is clustered into two clusters (hereinafter referred to as clustering (2)) and clustered into three clusters (hereinafter referred to as clustering (2)) by the k-means method. Clustering (3)) to calculate a threshold value. FIG. 20 shows the density histogram of 1801, FIG. 21 shows the density histogram of 1807, and FIG. 19 shows the density histogram of 1804. If it is known in advance how many colors are included in a single frame area, a predetermined number of clusterings may be performed. On the other hand, when handling document images of various formats, it is difficult to check the number of colors in each frame area in advance, and there are pre-printed characters even in the same position of a frame area of the same type of document In some cases, for example, the number of colors is not limited. For example, clustering (2) and clustering (3) are performed for each single frame region, and threshold values are calculated. The number of clusters is not limited to two or three depending on the form to be handled. Here, when the clustering (3) is performed, two thresholds are calculated as shown in FIG. 19, but the characters to be entered in the form, that is, the characters to be read, generally have lower grayscale values than the preprinted characters ( Therefore, the threshold value C (1904), which is the lower of the two gray levels, is used. Although only two colors or three colors exist in the single frame area of the form handled in this embodiment as described above, if N colors exist in the area to be binarized, the cluster is divided into N clusters. It is conceivable to divide each pixel by performing clustering.
[0010]
Next, in step 2607, the inside of the single frame area is binarized using the two threshold values calculated for each single frame area to generate two binary images. FIG. 22 shows a result of binarizing 1804 with a threshold C (1904), and FIG. 23 shows a result of binarizing 1804 with a threshold D (1905). As shown in FIG. 22, when binarized by the threshold C, only the entered characters can be black pixels except for the pre-printed characters, but binarized by the threshold D as shown in FIG. In this case, the entered characters and the pre-printed characters overlap, and the characters cannot be correctly recognized.
[0011]
Next, in step 2609, character recognition is performed for each single frame area. At this time, for each single frame area, there are two binary images with different binarization thresholds (threshold by clustering (2) and threshold by clustering (3)). Perform character recognition.
[0012]
Next, in step 2610, of the two sets of character recognition results with different binarization thresholds obtained for each single frame region, the character recognition results (e.g., amount, name, address, etc.) described in the frame match. The character recognition result to be output is output as a reading result. The knowledge of a character string refers to the attribute of each piece of information to be described in each form, and is a database of notation patterns of character strings described in the frame or all patterns. For example, if the amount is money, information such as a mark, a series of numbers, and commas is stored. If the name is an address, the notation patterns of all names and addresses to be entered are stored. These are stored in the format information recording unit 209 of FIG. 2 in advance.
[0013]
FIG. 27 is a configuration diagram of the above embodiment. In the figure, reference numeral 2701 denotes a means for inputting a gray image such as a scanner which reads a form as a gray image, 2702 denotes a read gray image, means for storing a binary image being processed, and 2704 denotes a form in the read gray image. Means for detecting the position (coordinates of the four corners), 2705 means for cutting out the portion where the characters to be read are described from the shaded image, 2706 means for extracting the frame structure of the form, and 2607 means the inside of the shaded image for each frame. Means 2708 for calculating a binarization threshold for generating a binary image suitable for character recognition for each partial area in the image, and 2709 means a partial area in the image based on the threshold obtained in 2708. Means for generating a binary image of 2710, means for cutting out characters from the binary image generated in 2709 and performing character recognition, and 2711 means for recognizing the character recognition result obtained in 2710. Chi is a means for outputting a match with the knowledge of the character string to be described in the reading target in the frame (amount, name, address, etc.). Reference numeral 2706 denotes a binarizing unit for extracting the frame structure of 2712, a unit for extracting the ruled line of 2713, and a unit for detecting the position of the frame of 2714. Reference numeral 2715 denotes a unit for holding information on a format (frame position) of a form used when cutting out a portion where characters are described in 2705, and 2716 denotes a unit for holding knowledge of a character string used in 2711. is there.
Reference numeral 2718 denotes a binary image display unit for visually confirming the recognition result.
The present invention disclosed above can be realized by a program and executed by an information device such as a computer.
[0014]
【The invention's effect】
According to the present invention, in a pattern recognition method for characters, symbols, marks, and the like in a document image such as a form, after extracting a frame structure in the document image, one information in the document image is extracted using information of the frame structure. It divides into one frame area, calculates a binarization threshold suitable for recognizing a pattern described in the frame for each frame area, and generates a binary image (binary image A) using this threshold. By doing so, it is possible to generate a binary image suitable for recognition even if the document image is composed of various color frame lines, preprinted characters, backgrounds, and entered characters, thereby improving recognition accuracy. be able to. Further, by changing the binary image displayed on the display for confirming the recognition result to the binary image A, it is possible to obtain a document image composed of various color frame lines, preprinted characters, backgrounds, and entered characters. Also, since the character to be recognized without blurring or crushing can be displayed, and the frame to be recognized can be accurately displayed, it is easy to visually confirm the recognition result.
[Brief description of the drawings]
FIG. 1 is a diagram showing a processing flow representing the present invention.
FIG. 2 is a configuration diagram of an embodiment.
FIG. 3 is a diagram showing a processing flow according to a conventional method when reading all characters in a form.
FIG. 4 is a diagram showing a processing flow according to a conventional method when reading a character at a specific position in a form.
FIG. 5 is a diagram showing an example of a processing flow of a frame structure extraction process.
FIG. 6 is a view showing an example of an image composed of various grayscale frame lines, written characters, preprinted characters, and a background.
FIG. 7 is a diagram showing a light and shade histogram in a reading area.
FIG. 8 is a diagram showing binarization thresholds (A, B) calculated based on a density histogram of a reading area.
FIG. 9 is a view showing a binary image based on a threshold value A (801).
FIG. 10 is a diagram showing a binary image based on a threshold value B (802).
FIG. 11 is a diagram showing a binarization method for each local region.
FIG. 12 is a diagram showing an example (1) of a binary image by a binarization method for each local region.
FIG. 13 is a diagram showing an example (2) of a binary image by a binarization method for each local region.
FIG. 14 is a diagram showing a multi-value image when a light character exists near a dark frame line.
FIG. 15 is a diagram showing a binary image by a binarization method for each local region when a light character exists near a dark frame line.
FIG. 16 is a diagram showing a ruled line detection result.
FIG. 17 is a view showing a frame detection result.
FIG. 18 is a diagram showing an area division image based on frame information.
FIG. 19 is a view showing a density histogram of a frame area (1804).
FIG. 20 is a view showing a density histogram (2) of a frame area (1801).
FIG. 21 is a view showing a density histogram (3) of a frame area (1807).
FIG. 22 is a view showing a binary image of a frame area (1804) based on a threshold value C (1904).
FIG. 23 is a view showing a binary image of a frame area (1804) based on a threshold value D (1905).
FIG. 24 is a view showing a binary image of a frame area (1801) based on a threshold value E (2003).
FIG. 25 is a view showing a binary image of a frame area (1807) based on a threshold value F (2103).
FIG. 26 is a diagram showing a processing flow of the embodiment.
FIG. 27 is a diagram illustrating a configuration of an input image according to the embodiment.
[Explanation of symbols]
601: Entry characters (dark), 602: Entry characters (light), 603: Border line (dark), 604: Pre-printed characters, 605: Border line (light), 606: Background part (dark), 607 ... Background part (Thin), 701: shading distribution of the entry character string “000257003” existing at the top in FIG. 6 and the entry character “1” existing at the right in FIG. 6, 702: entry existing near the center of FIG. Shade distribution of characters "6, 30", 703: shade distribution of thick frame line in FIG. 6, 704 ... shade distribution of pre-printed characters "post-date, month, day, business office" in FIG. 6, the density distribution of the thin frame line, 706... The density distribution of the dark background portion existing at the center in FIG. 6, 707 the density distribution of the thin background portion in FIG. 6, 801, 802. , Threshold values obtained by the k-means method from 02: a portion where the entered character overlaps the pre-printed character according to the binary image example by the local region-based binarization method; 1301, 1302 ... the pre-printed character is formed by the binary image example according to the local region-based binarization method. Missing portions, 1701, 1702, 1703, 1704, 1705, 1706, 1707, 1708 ... frame structure extraction is performed, and the obtained individual frame regions, 1801, 1802, 1803, 1804, 1805, 1806, 1807, 1808 ... The image is divided using the information on the frame structure, and the obtained individual frame areas are obtained, 1901... Distribution of the density of the characters existing in the frame area (1804), 1902... Preprinted characters existing in the frame area (1804) , Distribution of the background existing in the frame area (1804), 1904, 1905, density histogram of the frame area (1804) Binarization thresholds (C, D) calculated from the data using the k-means method, 2001: shading distribution of characters entered in the frame area (1801), 2002: shading of the background existing in the frame area (1801) Distribution, 2003: binarization threshold (E) calculated from the density histogram of the frame area (1801) using the k-means method, 2101: Density distribution of preprinted characters existing in the frame area (1807), 2102: frame The gray level distribution (F) of the background existing in the area (1807), 2103..., Calculated from the gray level histogram of the frame area (1807) using the k-means method.

Claims (9)

An image input unit, a form image processing apparatus having a processing unit,
The processing unit includes:
Detecting an entry frame from the form image input via the image input unit, and dividing the form image using information on the detected entry frame,
Calculating a threshold value for binarization for each of the divided areas;
Performing binarization using the threshold value for each of the regions;
A form image processing apparatus characterized by controlling 2. The form image processing apparatus according to claim 1, wherein the dividing step is performed for each of the detected entry frames. 3. The form image processing apparatus according to claim 1, wherein the dividing step further divides the line area of the entry frame and an area surrounded by the line area of the entry frame. The form image processing apparatus further includes a storage unit that stores format information of the form, and the control unit includes:
4. The form image processing apparatus according to claim 1, further comprising a step of recognizing a character entered in the binarized form image using the format information. 5. The form image processing apparatus according to claim 1, wherein the processing unit further controls a step of binarizing the form image prior to the step of detecting the entry frame. The step of calculating the threshold value for binarization includes:
A histogram of gray values in the region is created, and pixels in the region are clustered into a plurality of preset groups using the distribution of the histogram. Calculate thresholds from the results,
The step of performing the binarization is performed for each of the thresholds,
Recognizing the characters for each of the binarized images, comparing the character string knowledge stored in the storage unit with the respective recognition results, and outputting any of the recognition results based on the comparison results. The form image processing apparatus according to claim 4 or 5, wherein: Input means for a form image,
Storage means for storing format information of the form image input via the input means,
Means for detecting the coordinates of the reading area in the form image input from the format information,
Means for cutting out an area obtained by enlarging the area specified by the coordinates by a predetermined value;
Means for detecting a frame from the cut-out region, and dividing the cut-out region for each frame;
Means for calculating a binarization threshold for each frame,
Means for generating a binary image for each frame using the threshold value;
Means for recognizing characters in the binary image. 8. The pattern recognition apparatus according to claim 7, further comprising a display unit for displaying the cutout area. Obtaining a form image via the image input unit;
Detecting an entry frame from the form image;
Dividing the form image using the information of the detected entry frame,
Calculating a threshold value for binarization for each of the divided areas;
Performing binarization using the threshold value for each of the regions;
Performing character recognition for each of the binarized areas;
A program for causing a computer to execute a pattern recognition method characterized by having:

Images