JP2000032247A - Image recognition device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide an image recognition device capable of speeding up a processing without lowering the reliability of a judged result in the top/bottom recognition processing of an original. SOLUTION: An area division part 230 divides image data outputted from an image signal processing part 120 and binarized into plural areas. Then, for the respective divided areas, a reliability judgement part 240 obtains the reliability in the case of using them for top/bottom recognition. A top/bottom recognition part 250 segments character data from the area whose reliability is a highest value and executes the top/bottom recognition processing.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an image recognition apparatus for recognizing the direction of a document read by an image forming apparatus such as a copying machine. Hereinafter, recognizing the orientation of a document is referred to as “top and bottom recognition”.

[0002]

2. Description of the Related Art In a copying machine, in particular, in a digital copying machine, when a large number of originals are successively copied, research and development of a technology for enabling copying in the same direction regardless of the orientation of the originals has been progressed. JP-A-6-103410). If the orientation of the copy result is not constant unless the orientation of the original is constant, there is an inconvenience that the user must rearrange the original or the copy result before or after copying.

In order to align the direction of the copy result, it is necessary to perform the process of recognizing the top and bottom of the document and rotating the image. In many of the upside-down recognition processing methods, the direction of a character cut out from image data of a document is determined, and the direction of the character is used as the direction of the document. The image rotation process is a process of rotating the image data by a required angle when the direction of the document obtained by the top / bottom recognition process does not match the predetermined direction, so that the image data matches the predetermined direction. If a copy image is formed from the image data after the rotation processing, the direction of the copy result becomes constant.

Various methods have been devised for the top-bottom recognition process in order to increase the efficiency of the process and improve the reliability of the judgment result. Among them, a method for improving reliability is described in Japanese Patent Application Laid-Open No. 9-69136. The method disclosed here takes into account the existence of characters that are exceptions to the basic premise of the orientation recognition process “text direction = original direction”, and An attempt is made to reduce erroneous recognition caused by the operation.

FIG. 8 shows an example of a character whose orientation does not match that of a document. A character string 801 in FIG.
This is a caption character added for explanation in a chart or the like. In the upward document 810, the document is oriented leftward. A character string 802 in FIG. 8B is a character in the table. Since the table 821 is placed horizontally (leftward), the character string 802 is leftward even though the document 820 is upward.

It is easy to understand that if the top-bottom recognition is performed based on these caption characters and characters in the table, the orientation of the document is erroneously recognized as a result. Therefore, Japanese Patent Application Laid-Open
According to the method described in Japanese Patent No. 69136, the top and bottom recognition using caption characters and characters in the table is performed as little as possible in the following procedure. First, the character portion of the document is divided into a plurality of areas. Next, the attribute of each area is determined. The attributes are
"Text attribute", which corresponds to the text, "Title attribute", which indicates the title, "Character attribute in the table", which indicates the description in the table, "Caption attribute", which indicates the description character accompanying figures and graphs, etc. is there. Further, a priority is set for each area based on the attribute. The priority is generally high for "text" and "title", and low for "characters in table" and "caption". Then, a plurality of characters are cut out from the high priority area, and top and bottom recognition is performed on each character. If the result of the recognition of the plurality of characters matches, the result is adopted. If the result of the recognition does not match, the character is cut out from the area having the next highest priority and the top and bottom recognition processing is performed.

[0007]

However, in the above-mentioned prior art, the attribute is determined for each area, and the top / bottom recognition processing is performed for each area in accordance with the priority while considering the priority of the attribute. Of course, such processing is intended to improve the accuracy of the top-bottom recognition result, and is not useless, but in reality, the priority is fixed, and even if the attribute is "title" or "text", In most cases, top and bottom recognition is performed. Performing top / bottom recognition using “characters in table” or “captions” is when only these characters are present in the manuscript, and the reliability of the top / bottom recognition result is low regardless of the top / bottom recognition method used for such a manuscript. . It is difficult to imagine a case where a document in which “text” and “caption” are mixed and the recognition of the top and bottom is performed by intentionally increasing the priority of “caption”. Therefore, in the upside-down recognition processing performed by setting up the priorities in the areas divided for each attribute, the load on the effect is often excessive. SUMMARY OF THE INVENTION The present invention has been made in view of the above circumstances, and has as its object to provide an image recognition apparatus capable of performing top and bottom recognition of a document with a smaller load and without reducing the reliability of the result.

[0008]

In order to solve the above-mentioned problems, an image recognition apparatus according to the present invention comprises an image reading means for reading an original and generating image data, and a dividing means for dividing the image data into a plurality of areas. Means, for each of the plurality of areas, a reliability calculation means for calculating the reliability when used in the top and bottom recognition processing of the document, and the top and bottom of the document to be read from the image data of the area with the highest reliability It is characterized by including a top and bottom recognition means for determining, and this configuration makes it possible to improve the top and bottom recognition processing speed without lowering the accuracy of the top and bottom recognition result.

As for the reliability, the reliability calculating means creates a histogram of the image data for each of the divided areas, and determines the reliability of the area based on the difference between the maximum value and the minimum value of the frequency in the scanning direction. Calculate the degree. Regarding the reliability, the reliability calculating means creates a histogram of the image data for each of the divided areas, and obtains the number of change points whose frequency increases in the scanning direction and the number of change points whose frequency decreases. It can be said that the reliability of the area is obtained from the two values.

[0010] Even when there are a plurality of areas having the highest reliability in the plurality of divided areas, the top-and-bottom recognizing means adds the top / bottom recognition result of the plurality of areas to the area having the next highest reliability. Since the top and bottom of the document is determined with reference to the top and bottom recognition results, the accuracy of the recognition results is high.

[0011]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Embodiments of the present invention will be described below, taking a digital copying machine as an example, with reference to the drawings. (1) Overall Configuration of Digital Copying Machine First, a digital copying machine 1 (hereinafter, referred to as a digital copying machine) according to the present embodiment.
Simply referred to as “copier 1”. 1) will be described with reference to FIG. As shown in FIG. 1, the copying machine 1 includes an automatic document feeder 10, an image reading section 30, a printer section 50,
And a paper feeding unit 70.

The automatic document feeder 10 is a device for automatically feeding a document to the image reading section 30. The document placed on the document feed tray 11 is fed by a feed roller 12 and a separating roller 13. Each sheet is separated and sent downward, and is transported by the transport belt 14 to a document reading position on the platen glass 31. The document transported to the document reading position
After being scanned by the scanner 32 of the image reading section 30, the sheet is again sent rightward in the drawing by the transport belt 14,
The document is discharged onto a document discharge tray 16 via a discharge roller 15.

The image reading section 30 includes the platen glass 3
The scanner optically reads an image of a document conveyed to one document reading position, and includes a scanner 32, a CCD image sensor (hereinafter, referred to as a “CCD sensor”) 38, and the like. The scanner 32 is provided with an exposure lamp 33 and a mirror 34 for changing the optical path of light reflected from the original by irradiation of the exposure lamp 33 in a direction parallel to the platen glass 31. The original on 31 is scanned. The reflected light from the original is reflected by a mirror 34 and then further transmitted through mirrors 35 and 36 and a condenser lens 37 to a CCD image sensor 38.
, Where it is converted to an electrical signal to generate image data.

The image data is A / D-converted into a digital signal by the control unit 100, subjected to shading correction, density conversion, and the like, subjected to a known error diffusion process, and then temporarily stored in a memory. Is stored. Then, rotation processing is performed according to the result of the top and bottom recognition, and the rotation signal is used as a drive signal for the laser diode 51 of the printer unit 50.

The printer section 50 forms an image on a recording sheet by a known electrophotographic method, and upon receiving the drive signal, drives the laser diode 51 to emit laser light. The laser light is reflected by a mirror surface on the side of the polygon mirror 52 rotating at a predetermined angular velocity,
The surface of the photosensitive drum 56 is exposed and scanned through the fθ lens 53 and the mirrors 54 and 55. This photosensitive drum 56
Before the exposure, the cleaning unit 57 removes the residual toner on the surface of the photoreceptor, and after being irradiated with an eraser lamp (not shown), is discharged, and then uniformly charged by the charging charger 58. When the above-described exposure is performed in a state of being uniformly charged, the photosensitive drum 5
An electrostatic latent image is formed on the six surfaces. The developing device 59 develops the electrostatic latent image formed on the surface of the photosensitive drum 56.

On the other hand, the paper supply section 70 is provided with two paper cassettes 71 and 72,
In synchronism with the exposure and development operations in 6, a recording sheet of a required size is fed from one of the paper cassettes 71 and 72 by driving the paper feed roller 711 or 721. The fed recording sheet is a photosensitive drum 5
6, the surface of the photosensitive drum 56 contacts the surface of the photosensitive drum 56. At this time, the toner image formed on the surface of the photosensitive drum 56 is transferred to the surface of the recording sheet by the electrostatic force of the transfer charger 60.

Thereafter, the recording sheet is separated from the separation charger 6.
The photosensitive drum 56 is separated from the surface of the photosensitive drum 56 by the electrostatic force of 1, and is transported to the fixing unit 63 by the transport belt 62. The toner image transferred to the recording sheet is fixed by being pressed while being heated by a fixing roller 64 having a heater inside in a fixing unit 63. The recording sheet after fixing is discharged onto a discharge tray 66 by a discharge roller 65.

An operation panel 90 is provided at an easy-to-operate position on the front of the image reading section 30. The operation panel 90 has ten keys for inputting the number of copies, a start key for instructing the start of copying, and various copy modes. There are provided a setting key, a display unit for displaying a mode set by the setting key or the like by a message, and the like.

(2) Configuration of Control Unit 100 Next, the configuration of the control unit 100 installed inside the copying machine 1 will be described with reference to the drawings. FIG. 2 is a block diagram illustrating a configuration of the control unit 100. The control unit 100 includes an image reading control unit 110, an image signal processing unit 120, a memory control unit 130, a printer control unit 140, a main control unit 150,
The document recognition unit 200 is included. Each of the components described above is configured around a CPU, and transmits information and commands via a command line (indicated by a dotted line in the figure) to image data via an image data bus (indicated by a solid line in the figure). Interact with each other.

The image reading control section 110 controls the operations of the automatic document feeder 10 and the image reading section 30. That is, it is started in response to an execution instruction from the main control unit 150, and first causes the automatic document feeder 10 to sequentially feed a document. Then, it instructs the image reading unit 30 to read the conveyed document, and causes the image signal processing unit 120 to output the read image data.

The image signal processing unit 120 includes the CCD sensor 3
8 is converted into a digital multi-value signal by an A / D converter, and the shading correction unit causes uneven illuminance of the exposure lamp 33 and the CCD sensor 38.
To correct the sensitivity unevenness. After that, the MTF correction unit performs processing such as improving image quality such as edge enhancement, and outputs the processed image to the document recognition unit 200 and the memory control unit 130.

The document recognizing unit 200 performs the top and bottom recognition of the document based on the image data, and if the orientation of the document needs to be adjusted as a result of the top and bottom recognition, the memory control unit 13
Instruct 0 to rotate the image data. Details of the configuration and processing of the document recognition unit 200 will be described later.

The memory control unit 130 includes the image signal processing unit 1
The image data output from 20 is binarized, and if necessary, compressed and stored in the image memory 131 once. Then, upon receiving an instruction from the main control unit 150, the image data is read from the image memory 131, multi-valued, and if compressed, decompressed to return to the image data before storage in the image memory 131. Further, when an instruction for image rotation processing has been received from the document recognizing unit 200, the image data is rotated by an angle corresponding to the instruction and output to the print control unit 140 for image forming processing. In addition,
The image rotation processing is performed using a known technique (for example, Japanese Patent Application Laid-Open No. 60-126679).

The printer control section 140 converts the image data output from the memory control section 130 into laser diode drive signals for each reproduced color, outputs each to the laser diode 51, and performs exposure scanning. Let it do. When the main control unit 150 receives a user's designation (copy number, single / double-side designation, copy start instruction, etc.) from an operation panel (not shown), the main control unit 150 notifies the components of the control unit 100 of the designated contents. In addition, the processing timing of each component is uniformly controlled to realize a smooth copying operation.

(3) Configuration of Document Recognition Unit 200 Next, the configuration and processing of the document recognition unit 200 that performs the top-bottom recognition process in the control unit 100 will be described. FIG. 3 is a block diagram illustrating the configuration of the document recognition unit 200. The document recognition unit 200 includes a recognition control unit 210, a binarization unit 220, an area division unit 230, a reliability determination unit 240,
It comprises a top and bottom recognition unit 250, a work memory 260, and the like.

The binarizing section 220 includes an image signal processing section 120
Is converted to binary data by comparing the multi-tone image data output from the FD with a threshold level of a predetermined tone level. Then, the binarized image data is stored in the working memory 26.
0, and notifies the recognition control unit 210 of the end of the process.

The area dividing section 230 receives an instruction from the recognition control section 210 and divides the binarized image data in the working memory 260 into a plurality of areas. FIG. 4 is a schematic diagram illustrating an example of area division. Here, the original is divided into two equal parts in the main scanning direction and the sub-scanning direction, and divided into four areas A, B, C, and D. The region dividing unit 230 identifies the image data of the divided region with the identification information (work memory 26).
(Address at 0) to the reliability determination unit 240. Then, it notifies the recognition control unit 210 of the end of the process.

The reliability judging section 240 includes the area dividing section 230
A histogram is created for each area in the working memory 260 based on the address notified from the user, and the direction (line direction) of the character string in the image data is determined from the histogram in either the main scanning direction or the sub-scanning direction. judge. Then, based on the histogram in the row direction, the reliability when the image data of each area is used for the top and bottom recognition is determined. The reliability here is specifically an MTF value calculated from a histogram of image data. The MTF value is obtained by taking the maximum value (max) and the minimum value (min) of the histogram values in the histogram in the row direction and applying them to the following (Equation 1). MTF value = (max−min) / (max + min) (Equation 1) The reliability calculation unit 240 divides the row-direction histogram of the region into several sections to obtain MTF values, and calculates the MTF value of each section. The average value is used as the reliability of the area.

FIG. 5 shows an example of image data in which the reliability (MTF value) increases. The histogram 510 is divided into six sections 511 to 516, and in each section,
The maximum value of the histogram value varies, but the minimum value is 0. As a result, the MTF value for all sections is: max−0 / max + 0 = max / max = 1. “1” is the maximum value of the MTF value. The MTF has the maximum value as described above because the histogram 510 has a valley (portion where frequency = 0).
The image data 520 that is the basis of the histogram 510 includes
This means that a plurality of character data arranged in a line are arranged at intervals.

The MTF value becomes low in the case of image data which includes information other than characters, such as table ruled lines and figures, and in which a valley cannot be formed, in addition to the case where the original is skewed. Character data that is not suitable for use in top and bottom recognition, such as caption characters and characters in tables, often includes information other than characters such as graphs and table ruled lines. It can be considered that the characters included in the (contained area) are inappropriate to be used for the top and bottom recognition. Conversely, in an area having a large MTF value, it can be considered that character data is arranged in a line without inclination as described above, and the reliability in top-bottom recognition is high. The above is the basis for setting the MTF value as the reliability of the top-bottom recognition result.

The reliability determination section 240 creates a histogram from the binary image data in this way, and obtains the average of the MTF values of the created histogram as the reliability. When the calculation of the reliability is completed, the address of the binary image data of the area, the address of the corresponding histogram, and the numerical value of the reliability are output to the recognition control unit 210 as a pair. In the example of FIG. 4, among the four areas, the areas C,
Since D includes a graph, the reliability is low. The region B does not include a graphic such as a graph, but has many blank portions. The highest reliability is the area A in which all data is a character string.

The top / bottom recognition unit 250 includes an area from which the address is output from the recognition control unit 210 (the reliability determination unit 240).
Is determined by using a known method. A variety of methods for recognizing the top and bottom are disclosed (JP-A-4-229763, JP-A-7-65120, etc.), so a detailed description will be omitted, but the basic procedure is as follows. First, data of one character is cut out from the image data of the processing target area in accordance with the histogram, and character data (characters for comparison) corresponding to the cut data is found from a pattern dictionary in a memory (not shown). Then, the character for comparison is rotated by 90 degrees and compared with the cut-out data. Then, the angle (0, 90, 180 or 270 degrees) at the time of matching is output to the recognition control unit 210 as information indicating the direction of the cut-out character.

Next, the recognition control unit 210 will be described. Since the recognition control unit 210 also controls the entire processing of the document recognition unit 200, the operation of the document recognition unit 200 is also described. FIG. 6 is a flowchart illustrating the flow of the document top / bottom determination process performed by the document recognition unit 200. The processing by the document recognition unit 200 is started at the timing when the corrected image data is output from the image signal processing unit 120.

First, the recognition control unit 210 includes a binarization unit 220
, The image data is binarized and stored in the temporary work storage (S601), and the area dividing unit 230 is instructed to divide the image data. The area dividing unit 230 divides the binarized image data (S602).

When the recognition control unit 210 receives the number of divided regions and the address of each region from the region dividing unit 230, it outputs these information to the reliability determination unit 240 to determine the reliability of each region. The reliability determination unit 240 generates a pixel histogram for each region (S604), calculates the MTF value, and notifies the recognition control unit 210 (S605).
This is repeated until there is no unprocessed area (S603).

The recognition control section 210 includes a reliability determination section 240
The reliability information for each area output from the above is held, and when the reliability information for all the areas is available, an area to be used for top-bottom recognition is selected based on the value of the reliability. The recognition control unit 210 selects an area where the value of the reliability is maximum (S606). Then, the recognition control unit 210 compares the value of the reliability with a predetermined threshold. When the maximum value of the reliability is lower than the threshold value (S607: No), the memory control unit 13 does not issue a process execution instruction to the top and bottom recognition unit 250.
For 0, information that the image data does not require rotation correction (rotation angle = 0 degrees) is output (S615, S61).
8). This is only a relative maximum,
This is because, for example, even in an area having a reliability of about 20%, the maximum value is obtained if the reliability of another area is a low value such as 10%. If the maximum value of the reliability is low, it is considered that reliable results cannot be obtained even if any area is used for top-bottom recognition. And copy it.

On the other hand, when the maximum value of the reliability is equal to or larger than the predetermined threshold (S607: Yes), the recognition control section 210 outputs the address of the binary image data and the histogram of the area to the top and bottom recognition section 250. Then, an instruction is given to perform upside down recognition using these (S608). Then, in response to this instruction, the top / bottom recognition unit 250 outputs the top / bottom recognition result of the area, and based on the result, determines the rotation angle required to orient the copy of the original in a predetermined direction. Calculation and outputs this to the memory control unit 130 (S61).
7).

When there are a plurality of areas having the highest reliability values (S609: Yes), the recognition control section 210 causes the above-described top / bottom recognition processing to be performed on all the areas (S609).
608), if the result matches in a plurality of areas (S610: Y
es), adopt the result. If the result does not match between the regions (S610: No), the other region having the second highest reliability is further subjected to the top-bottom recognition process (S61).
1). At this time, the recognition control unit 210 also compares the second highest reliability with the threshold, and performs the top / bottom recognition only when the reliability is equal to or higher than the threshold (S612: Yes) (S612).
613). If the difference is smaller than the threshold value (S612: No), information indicating that rotation correction of image data is unnecessary (rotation angle = 0 degrees) is output to the memory control unit 130, and the process ends (S615). , S618).

When the top / bottom recognition processing is performed on the area having the second highest reliability, the recognition control unit 210 compares this result with the two kinds of results obtained earlier, and if it matches one of the results (S
614: Yes), calculate the required rotation angle based on the matching result, and output this to the memory control unit 130 (S61).
6, S618). If none of the results of the top and bottom recognition match the results of the top and bottom recognition performed later (S614: N
o), the recognition control unit 210 determines that the orientation cannot be recognized, and outputs information (rotation angle = 0 degrees) to the memory control unit 130 that the rotation correction of the image data is unnecessary (S615,
S618).

As described above, in this embodiment,
The document recognizing unit 200 divides the image data into regions and recognizes the top and bottom of the document using data cut out from the region with the highest reliability. At this time, the region is divided according to a simple rule, and the attribute (text , Captions, characters in the table, etc.). In addition, since the determination of the reliability is determined by the MTF value of the histogram of the image data, the load of the top and bottom recognition processing is greatly reduced as compared with the related art. In addition, by using the MTF value as a reference for reliability, data that causes erroneous recognition, such as caption characters and characters in a table, can be eliminated, so that the reliability of the recognition result does not decrease as compared with the related art.

In the present embodiment, an area having a high degree of reliability (an area in which character data is arranged in a line) is determined based on the MTF value when used for upside-down recognition. By using this, it is also possible to determine a region having high reliability.

FIG. 7 shows the relationship between the number of edges and image data. FIG. 11A shows image data of horizontally written left-justified text without inclination, and a histogram 710 of the image data in the scanning direction that does not coincide with the row direction. In the histogram 710, the number of change points in the direction in which the histogram value increases (increasing edge: indicated by white circles in the figure) is the number of change points in the direction in which the histogram value decreases (decreasing edge: indicated by black circles in the figure). Less than the number of. (In FIG. 7, there are two increasing edges and four decreasing edges.
Pieces. The start position of the character string matches on the left except for the line feed, while the end of the character string is unspecified. In image data including an inclined character string or a chart, the number of edges increases, and it is difficult to make a difference between the numbers of increasing edges and decreasing edges. Therefore, by paying attention to the number of increasing edges and the number of decreasing edges, and the difference between the two, it is possible to find image data containing many character strings. It can be determined that the smaller the sum of the number of increasing edges and the number of decreasing edges, and the greater the difference between the two, the higher the possibility of character string image data such as image data 720.

In the above-described embodiment, an example in which the image recognition apparatus according to the present invention is applied to a monochrome copying machine has been described. However, in other apparatuses requiring document recognition, for example, a color copying machine or a facsimile machine. It is also applied as an image recognition device. However, in this case, it is necessary to incorporate a circuit for canceling chromatic data in the image data in advance. Since the chromatic data cancel circuit is a known technique, a detailed description is omitted.

[0044]

As is apparent from the above description, according to the image recognition apparatus of the present invention, there are provided image reading means for reading an original to generate image data, and dividing means for dividing the image data into a plurality of areas. A reliability calculating unit that calculates a reliability of each of the plurality of regions when the plurality of regions are used in the top and bottom recognition process of the document, and determines a top and bottom of the document to be read from the image data of the region having the highest reliability. Since the top / bottom recognition processing is performed by the top / bottom recognition means, it is not necessary to determine the attribute of the area as in the related art, and the top / bottom recognition processing can be executed quickly. In addition, the reliability is calculated based on the values appearing in the histogram of the image data, and the higher the area containing a large amount of character data arranged in a row, the higher the reliability. Therefore, the accuracy of the recognition result performed using the area selected based on the reliability is Is also expensive.

[Brief description of the drawings]

FIG. 1 is a cross-sectional view showing the overall configuration of a copying machine to which an image recognition device according to the present invention is applied.

FIG. 2 is a block diagram showing a configuration of a control unit in the copying machine.

FIG. 3 is a block diagram illustrating a configuration of a document recognition unit in the control unit.

FIG. 4 is a diagram showing an example of area division of image data by the document recognition unit.

FIG. 5 is a diagram illustrating an example of image data of a type in which an MTF value, which is a specific measure of reliability, is increased, and a histogram thereof;

FIG. 6 is a flowchart illustrating a flow of a top / bottom recognition process performed by the document recognition unit.

FIG. 7 is a diagram for explaining edge counting, which is another measure of reliability.

FIG. 8 is a diagram illustrating an example of character data that causes erroneous recognition in conventional top-down recognition processing.

[Explanation of symbols]

 REFERENCE SIGNS LIST 1 copier 100 control unit 120 image signal processing unit 130 memory control unit 131 image memory 150 main control unit 200 document recognition unit 210 recognition control unit 230 area division unit 240 reliability determination unit 250 top and bottom recognition unit

Continued on the front page (72) Inventor Kazuhiro Ueda 2-3-13 Azuchicho, Chuo-ku, Osaka-shi, Osaka F-term in Osaka International Building Minolta Co., Ltd. 5C072 AA05 BA03 BA10 UA07 UA11 UA13 UA20 XA01 5C076 AA01 AA24 BA05 CA10

Claims (4)

[Claims] An image reading unit that reads an original to generate image data; a dividing unit that divides the image data into a plurality of regions; and a reliability when each of the plurality of regions is used in a document direction recognition process. An image recognition apparatus comprising: a reliability calculation unit that calculates a degree; and a top and bottom recognition unit that determines the orientation of a document to be read based on image data of an area having the highest reliability. 2. The reliability calculating means creates a histogram of image data for each of the plurality of regions, and calculates the reliability of the region based on a difference between a maximum value and a minimum value of a frequency in a scanning direction. 2. The method according to claim 1, wherein
An image recognition device according to claim 1. 3. The reliability calculation means creates a histogram of image data for each of the plurality of regions, and obtains the number of change points whose frequency increases in the scanning direction and the number of change points whose frequency decreases. The image recognition device according to claim 1, wherein the reliability of the area is obtained from two values. 4. When there are a plurality of areas having the highest reliability among the plurality of areas and the results of the top and bottom recognition of these areas do not match, the top and bottom recognition means adds the reliability to the top and bottom recognition results of these areas. 4. The image recognition apparatus according to claim 1, wherein the direction of the document is determined by referring to a top and bottom recognition result of an area having the next highest degree.