JP4885112B2 - Document processing apparatus, document processing method, and document processing program - Google Patents

Description

  The present invention relates to a document processing apparatus, a document processing method, and a document processing program that perform collation between document images.

  Conventionally, various techniques have been proposed as a method for extracting a character string from a document image in which the character string is recorded as an image (character line). For example, a technique capable of recognizing a character line as a character string by applying a plurality of restrictions on features (size, spacing, etc.) related to the shape and position of a rectangle circumscribing the character line included in the document image Has been proposed (see, for example, Patent Documents 1 and 2).

JP 11-219407 A International Publication No. 00/62243 Pamphlet

  However, in the techniques described in Patent Documents 1 and 2, it is necessary to manually adjust a plurality of constraints that function in the circumscribed rectangle to an optimum value in order to recognize a character line with high accuracy. In addition, although the character-likeness can be determined, the characteristics relating to the contents of the character line cannot be recognized, so that there is a possibility that sufficient accuracy cannot be obtained even when used for collation between document images. Further, since there is no mention of the use of the relative positional relationship between a plurality of lines, there is a problem that it is not possible to deal with a partial image that is a part of a document image as a collation target.

  The present invention has been made in view of the above, and provides a document processing apparatus, a document processing method, and a document processing program capable of determining similarity between document images more efficiently and with high accuracy. With the goal.

  In order to solve the above-described problems and achieve the object, the invention according to claim 1 is a document processing apparatus that performs collation between document images, based on a circumscribed rectangle for each character image included in the document image. A character line cutting unit that cuts out character lines that connect the circumscribed rectangles, a quantizing unit that quantizes arrangement information representing the characteristics of the circumscribed rectangles in the character lines in a fixed stage, and the quantized arrangement information A symbol generation means for symbolizing each of the symbols into a fixed type of symbol, an appearance frequency calculation means for calculating an appearance frequency of a symbol series composed of a predetermined number of the symbols, a document image to be collated, and the document image The appearance frequency calculated by the appearance frequency calculating means is checked against a plurality of document images to be checked, and a predetermined number of document images to be checked having higher correlation are selected. On the basis of the arrangement information corresponding to the symbol series that coincides with the document image to be collated and each document image to be collated selected by the collation target selecting unit. A distribution state deriving unit for deriving, for each document image, a distribution state of appearance positions of circumscribed rectangles represented by any or all of the arrangement information, and the document image to be collated derived by the distribution state deriving unit. Collation result selection means for determining a similarity between the distribution state and the distribution state of the document image to be collated, and selecting a document image to be collated having the highest similarity as a collation result; It is characterized by that.

  The invention according to claim 2 is the invention according to claim 1, wherein the distribution state deriving means derives the distribution state of the appearance position of the circumscribed rectangle in the horizontal direction and / or the vertical direction of the document image. It is characterized by that.

  The invention according to claim 3 is the invention according to claim 1 or 2, wherein the distribution state deriving means derives the distribution state of the appearance position of the circumscribed rectangle as a frequency distribution histogram.

  The invention according to claim 4 is the invention according to claim 1 or 2, wherein the distribution state deriving means regards the distribution state of the appearance position of the circumscribed rectangle as a normal distribution, and calculates an average and a standard deviation of the normal distribution. The method is characterized by deriving skewness and kurtosis.

  The invention according to claim 5 is the invention according to claim 4, wherein the distribution state deriving means includes each of the character lines included in the document image to be collated and the document image to be collated. The size of the circumscribed rectangle of the character image is totaled, and a numerical value defining the normal distribution is normalized by an average value or a mode value of the size.

  Further, the invention according to claim 6 is the invention according to claim 5, wherein the distribution state deriving means corresponds to the symbol series matched in the document image to be collated and the document image to be collated. The size of the circumscribed rectangle represented by the arrangement information is totaled.

  The invention according to claim 7 is a document processing method executed by a document processing apparatus that performs collation between document images, wherein the character line cut-out means is connected to each character image included in the document image. A character line cutting step for cutting out a character line connecting the circumscribed rectangles based on the rectangle, and a quantization step for quantizing the arrangement information representing the characteristics of the circumscribed rectangle in the character line to a fixed stage. A symbol generation step in which the symbol sequence generation means converts each of the quantized arrangement information into symbols of a fixed type, and an appearance frequency calculation means generates an appearance of a symbol sequence consisting of a combination of the predetermined number of symbols The appearance frequency calculating step for calculating the frequency, and the matching target selecting means, for the document image to be verified and the plurality of document images to be compared with the document image, A collation target selection step for collating the appearance frequencies calculated by the degree calculation means and selecting a predetermined number of document images to be collated with higher correlation; and a distribution state deriving means for comparing the document images to be collated The circumscribed rectangle represented by any or all of the arrangement information based on the arrangement information corresponding to the symbol series matched with each of the document images to be checked selected in the check target selection step A distribution state deriving step for deriving a distribution state of appearance positions for each document image, and a collation result selection unit, the distribution state for the document image to be collated derived in the distribution state deriving step; A collation result selection step of determining a similarity between the document image and the distribution state, and selecting a collation target document image having the highest similarity as a collation result. And wherein the door.

  The invention according to claim 8 is the invention according to claim 7, wherein the distribution state deriving means derives the distribution state of the appearance position of the circumscribed rectangle in the horizontal direction and / or the vertical direction of the document image. It is characterized by that.

  The invention according to claim 9 is the invention according to claim 7 or 8, wherein the distribution state deriving means derives the distribution state of the appearance position of the circumscribed rectangle as a frequency distribution histogram.

  The invention according to claim 10 is the invention according to claim 7 or 8, wherein the distribution state deriving means regards the distribution state of the appearance position of the circumscribed rectangle as a normal distribution, and calculates an average and a standard deviation of the normal distribution. The method is characterized by deriving skewness and kurtosis.

  The invention according to claim 11 is the invention according to claim 10, wherein the distribution state deriving means includes each of the character lines included in the document image to be collated and the document image to be collated. The size of the circumscribed rectangle of the character image is totaled, and a numerical value defining the normal distribution is normalized by an average value or a mode value of the size.

  The invention according to a twelfth aspect is the invention according to the eleventh aspect, wherein the distribution state deriving means corresponds to the symbol series matched in the document image to be collated and the document image to be collated. The size of the circumscribed rectangle represented by the arrangement information is totaled.

  According to a thirteenth aspect of the present invention, a computer that performs collation between document images, based on a circumscribed rectangle for each character image included in the document image, cuts out a character line that connects the circumscribed rectangles. Output means, quantization means for quantizing the arrangement information representing the characteristics of the circumscribed rectangle in the character line in a fixed stage, and symbol generation means for symbolizing each of the quantized arrangement information into a fixed type of symbol And an appearance frequency calculating means for calculating an appearance frequency of a symbol series composed of a combination of a predetermined number of symbols in the symbol series, a document image to be collated, and a plurality of document images to be collated with the document image The matching target selection is performed by collating the appearance frequency calculated by the appearance frequency calculating unit and selecting a predetermined number of document images to be compared having higher correlation. Each piece of arrangement information based on the arrangement information corresponding to the symbol series matched in the stage, the document image to be collated, and each document image to be collated selected by the collation target selection unit A distribution state deriving unit for deriving the distribution state of the appearance position of the circumscribed rectangle represented by any or all of each document image, a distribution state for the document image to be collated derived by the distribution state deriving unit, and It is characterized by functioning as a collation result selecting means for determining a similarity with a distribution state of a collation target document image and selecting a collation target document image having the highest similarity as a collation result. And

  The invention according to claim 14 is the invention according to claim 13, wherein the distribution state deriving means derives the distribution state of the appearance position of the circumscribed rectangle in the horizontal direction and / or the vertical direction of the document image. It is characterized by that.

  The invention according to claim 15 is the invention according to claim 13 or 14, wherein the distribution state deriving means derives the distribution state of the appearance position of the circumscribed rectangle as a frequency distribution histogram.

  The invention according to claim 16 is the invention according to claim 13 or 14, wherein the distribution state deriving means regards the distribution state of the appearance position of the circumscribed rectangle as a normal distribution, and calculates an average and a standard deviation of the normal distribution. The method is characterized by deriving skewness and kurtosis.

  The invention according to claim 17 is the invention according to claim 16, wherein the distribution state deriving means includes each of the character lines included in the document image to be collated and the document image to be collated. The size of the circumscribed rectangle of the character image is totaled, and a numerical value defining the normal distribution is normalized by an average value or a mode value of the size.

  The invention according to claim 18 is the invention according to claim 17, wherein the distribution state deriving means corresponds to the symbol series that matches in the document image to be collated and the document image to be collated. The size of the circumscribed rectangle represented by the arrangement information is totaled.

  According to the present invention, with respect to a document image to be collated and a document image to be collated, arrangement information representing features of a circumscribed rectangle in a character line is extracted, and these are quantized to a fixed stage to generate a symbol. Thus, it is possible to extract the characteristics of the character line without recognizing characters, and it is possible to select a predetermined number of document images to be verified that have a high correlation with the document image to be verified from the document images to be verified. . Further, by comparing the distribution state of the appearance positions of the matching symbol series between the document image to be verified and the selected document image to be verified, the similarity of the relative positional relationship of the symbol series can be obtained. Since it can be determined, the similarity between the document image to be verified and the document image to be verified can be determined with high accuracy. As a result, even if a partial image in the document image is a document image to be collated, a document similar to the partial image based on the positional relationship of the circumscribed rectangle of the character image included in the partial image Images can be searched with high accuracy.

  Exemplary embodiments of a document processing apparatus, a document processing method, and a document processing program according to the present invention are explained in detail below with reference to the accompanying drawings.

(Hardware configuration of document processing device)
FIG. 1 is a block diagram showing a hardware configuration of a document processing apparatus 100 according to the first embodiment of the present invention. As shown in FIG. 1, the document processing apparatus 100 is a computer such as a PC (Personal Computer), and a CPU (Central Processing Unit) 1 that controls each part of the document processing apparatus 100 and a program for starting the CPU 1 are provided. A ROM (Read Only Memory) 2 to be stored, a hard disk 3 for storing a document image and an operating system input by an image input unit 21 (to be described later), various programs, etc., and a RAM (Random Access Memory) functioning as a work area for the CPU 1 4, a keyboard 5 for receiving various inputs from an operator, a display device 6 for displaying input conditions, an optical disk drive 7 for reading programs stored in various optical information recording media (not shown) such as a CD-ROM, the Internet Document images via telecommunication lines such as LAN and Local Area Network (LAN) Communication device 8 to transmit and receive, are composed of a scanner 9, etc. for performing optical reading of a document image, a data input and output between these units is a bus controller 10 operates to arbitrate.

  In the document processing apparatus 100, when the operator turns on the power, the CPU 1 activates a program called a loader in the ROM 2, loads a program for managing the computer hardware and software called the operating system from the hard disk 3 into the RAM 4, and loads this operating system. Start. Such an operating system starts a program, reads information, and stores information in response to an operator's operation. As typical operating systems, Windows (registered trademark), UNIX (registered trademark), and the like are known. An operation program running on these operating systems is called an application program.

  Here, the document processing apparatus 100 stores in the hard disk 3 a document processing program related to document collation processing (to be described later) as a program executed by the CPU 1. In this sense, the hard disk 3 functions as a storage medium that stores the document processing program.

  In general, a program installed in the hard disk 3 of the document processing apparatus 100 is recorded on a storage medium such as various optical information recording media such as a CD-ROM or a magnetic medium such as an FD, and is recorded on the storage medium. The installed program is installed in the hard disk 3. For this reason, portable storage media such as various optical information recording media such as CD-ROM and magnetic media such as FD can be storage media for storing document processing programs. Further, the document processing program may be imported from the outside via the communication device 8 and installed in the hard disk 3, for example.

  When the document processing program operating on the operating system is activated, the CPU 1 realizes each functional unit described later in cooperation with the document processing program. Hereinafter, the functional configuration of the document processing apparatus 100 will be described.

(Functional configuration of document processing device)
FIG. 2 is a block diagram illustrating a functional configuration of the document processing apparatus 100. As shown in FIG. 2, the document processing apparatus 100 includes, as function units, an image input unit 21, a collation image selection unit 22, a rectangle extraction unit 23, a line cutout unit 24, a quantization unit 25, a symbol generation unit 26, an appearance, and the like. A frequency totaling unit 27, a candidate image selecting unit 28, an appearance position distribution deriving unit 29, a matching result selecting unit 30, and a display unit 31 are included.

  The image input unit 21 receives a document image input from the outside and stores it in the hard disk 3. Specifically, the function of the image input unit 21 can be realized by the optical disc drive 7, the communication device 8, and the scanner 9 shown in FIG.

  The collation image selection unit 22 selects a document image input from the image input unit 21 or a document image stored in the hard disk 3 designated via the keyboard 5 as a document image to be collated. Hereinafter, the document image to be collated is referred to as “collation image”. When a specific area in the document image is designated via the keyboard 5, the collation image selection unit 22 selects a partial document image (partial image) included in this area as a collation image. Shall.

  The collation image selection unit 22 selects a document image to be collated as a collation destination of the collation image. Here, the document image to be collated may be, for example, a part or all of the document images stored in advance in the hard disk 3, or the document image designated via the keyboard 5 is used as the document image to be collated. Also good. Hereinafter, the document image to be collated is referred to as “collated image”.

  The rectangle extraction unit 23 extracts a circumscribed rectangle of each character image included in the document image. Here, “character image” means an image in which characters in a predetermined language are represented as an image. The line cutout unit 24 cuts out character lines by connecting the circumscribed rectangles extracted by the rectangle extraction unit 23. Hereinafter, a circumscribed rectangle included in a character line is referred to as an “in-line rectangle”.

  The quantization unit 25 quantizes the arrangement information representing the characteristics of the in-line rectangles included in the character line extracted by the line extraction unit 24 at a fixed stage. Here, the characteristics of the in-line rectangle are parameter groups such as the black pixel density of the character image corresponding to each in-line rectangle, the height of the in-line rectangle in the character line, the start point position, and the like. It represents. The quantization of the arrangement information will be described later.

  The symbol generation unit 26 converts each of the arrangement information quantized by the quantization unit 25 into a fixed type symbol, and generates a series of symbol sequences corresponding to each character line constituting the document image. Hereinafter, a symbol series for the entire document image is referred to as an entire symbol series.

  The appearance frequency totaling unit 27 calculates a frequency (appearance frequency) at which a symbol series composed of a predetermined number of symbols appears in the entire symbol series. The candidate image selection unit 28 collates the appearance frequency calculated by the appearance frequency totaling unit 27 for the collation image and the collation image to be collated with the collation image, and selects the collation image having a higher correlation. Select a predetermined number. Hereinafter, the image to be verified selected by the candidate image selection unit 28 is referred to as “candidate image”.

  The appearance position distribution deriving unit 29, in each document image of the collation image and the candidate image, based on the in-line rectangle represented by any or all of the arrangement information corresponding to each symbol series matched in both document images, The appearance state distribution state is derived for each document image. Further, the appearance position distribution deriving unit 29 calculates the similarity between the distribution state for the collation image and the distribution state for the candidate image, and stores the calculated similarity in association with the corresponding candidate image in the RAM 4 or the like. .

  The matching result selection unit 30 selects a candidate image having the highest similarity as a matching result based on the similarity calculated by the appearance position distribution deriving unit 29.

  The display unit 31 displays the document image input from the image input unit 21, the progress of each process, and the like, and displays the candidate image selected by the matching result selection unit 30. The function of the display unit 31 can be realized by the display device 6 shown in FIG.

  Hereinafter, a document collation process which is a characteristic process of the present embodiment among various processes executed by the document processing apparatus 100 will be described below.

  FIG. 3 is a flowchart showing the procedure of the document matching process. First, the collation image selection unit 22 selects a document image input from the image input unit 21 or a document image designated via the keyboard 5 as a collation image (step S1). Next, the collation image selection unit 22 selects a collation image as a collation destination of the collation image selected in step S1 (step S2).

  Subsequently, the rectangle extracting unit 23, the line extracting unit 24, the quantizing unit 25, the symbol generating unit 26, and the appearance frequency totaling unit 27 execute an appearance frequency totaling process for each document image selected in steps S1 and S2. (Step S3). Hereinafter, with reference to FIG. 4, the appearance frequency totaling process in step S3 will be described. It should be noted that the appearance frequency totaling process is performed for each of the collation image and the collation image, but in the following description, it will be collectively referred to as “document image”.

  FIG. 4 is a flowchart showing the procedure of the appearance frequency counting process. First, the rectangle extraction unit 23 extracts a circumscribed rectangle circumscribing the black pixel of each character image included in the document image (step S31). Subsequently, the line cutout unit 24 connects the circumscribed rectangles adjacent in the horizontal direction to grow into character lines, and then cuts out each of these character lines (step S32).

  Here, extraction of a line of a document image will be described with reference to FIGS. The rectangle extraction unit 23 obtains a black pixel connected component for the document image (FIG. 5-1) and obtains circumscribed rectangles A, B, C,. Then, the line cutout unit 24 grows the circumscribed rectangle obtained by the rectangle extracting unit 23 into a character line Z by connecting the circumscribed rectangles adjacent in the horizontal direction (FIG. 5-3). Since the processing itself related to the generation of the in-line rectangle and the cut-out of the character line can use a known method, a detailed description thereof will be omitted.

  It should be noted that it is preferable to cut out a single character line from a document image in line units, paragraph units, chapter units, or the like. In general, the size of a character image included in a document image is uniform for each line, paragraph, or chapter. Therefore, by cutting out a character line in such a group, the size of the character image included in the character line is as follows. (Character size) can be made uniform. In this embodiment, the circumscribed rectangle is grown in the horizontal direction. However, the present invention is not limited to this. As an aspect in which the growth is performed in the vertical direction or in both the horizontal direction and the vertical direction according to the character direction. Also good.

  Returning to FIG. 4, the quantization unit 25 and the symbol generation unit 26 execute symbol generation processing for each character line cut out in step S <b> 32. Hereinafter, the symbol generation processing in step S33 will be described with reference to FIG.

  FIG. 6 is a flowchart showing a procedure of symbol generation processing. First, the quantization unit 25 measures the height of each character line cut out in step S32 (step S331).

  Next, the quantization unit 25 sorts the in-line rectangles in ascending order based on the horizontal start point (Xs) of the in-line rectangles included in each character line (step S332). Subsequently, the quantization unit 25 obtains arrangement information indicating the arrangement state of the aligned in-row rectangles, and quantizes the arrangement information in a fixed stage (step S333). Hereinafter, with reference to FIG. 7A, FIG. 7B, FIG. 8, and FIG. 9, the processing in steps S332 and S333 will be described.

  FIG. 7A and FIG. 7B are explanatory diagrams illustrating an example of arrangement of in-row rectangles. As shown in Fig. 7-1, Western character lines have mixed uppercase and lowercase letters, and there are symbols such as apostrophe, axis, umlaut, etc. It is clear that the heights are concentrated at two positions, a position and b position in FIG. That is, the rectangular arrangement positions are not symmetrical vertically. Asian character lines, on the other hand, have a complicated character structure such as kanji, hiragana, katakana, and hangul, as shown in Fig. 7-2. There is no clear concentration in two places. However, the arrangement position of the rectangle is not symmetrical in the up / down / left / right direction, which is the same as the Western line.

  Comparing the in-line rectangle of the European characters in Fig. 7-1 with the in-line rectangle of the Asian characters in Fig. 7-2, the arrangement of the in-line rectangles depends on the contents of the character line, regardless of the language type. It can be seen that it has changed. Therefore, by extracting the circumscribed rectangle of the character, it is possible to capture a rough feature of the character. That is, even if the character itself is not specified, for example, as shown in FIG. 8, the start point (Xs, Ys) and end point (Xe, Ye) of the rectangular coordinates are obtained, and the arrangement state of the circumscribed rectangle of the character image using this The image feature of each character line can be captured simply by acquiring the feature representing

  Even if the arrangement positions of the in-line rectangles are the same, Western characters have a simple structure, so the density of black pixels in the rectangles is lower than that of Asian characters. Of course, even in Asian characters, the black pixel density of hiragana and katakana with a simple structure is low, and the black pixel density of kanji with a complicated structure is high.

  Thus, the arrangement state of one rectangle in a character line can be uniquely defined by measuring the height of the start point of the in-line rectangle, the rectangle size (width, height), the black pixel density in the in-line rectangle, and the like. it can. In the process of step S333, these measurement results are acquired as arrangement information for each in-line rectangle of each character line, and quantized to a fixed stage.

In the following, an example of defining the arrangement state of the in-line rectangle with reference to the height of the start point of the in-line rectangle is shown. FIG. 9 is an explanatory diagram showing a method of quantizing the feature indicating the arrangement state of the in-line rectangle. Under the situation where the document is not specified, the line height is variable, and the height of the in-line rectangle is normalized by the following expression so that the processing does not depend on the value of the line height. In addition, ys means the height of the in-line rectangle start point, and H means the line height acquired in step S332.
YsRate = ys / H (1)

Here, since 0 <YsRate ≦ 1, it is easy to quantize YsRate in a fixed stage. For example, if you quantize to N stages,
YsVal = INT (YsRate * (N−1)) (2)
(However, INT (): rounded down to the nearest decimal point)
And it is sufficient. Each stage is labeled 0- (N-1). The rectangular width w and the rectangular height h are also quantized in the same procedure.

  By the way, in order to save storage capacity and reduce the amount of calculation, in image processing, a compressed image is often used as a processing target instead of an original image itself. Since the compressed image has a reduced number of pixels, information on the details of the character image is lost. As shown in FIG. 9, the present invention focuses on the circumscribed rectangle of the character image, and is not based on the detailed features of the image itself. Therefore, it can function effectively not only for the original image but also for the compressed image.

In the above description, the character line image is characterized by the height of the start point of the in-line rectangle, but the present invention is not limited to this. For example, in the case where the height of the in-line rectangle is used as the feature of the character line image, it is as follows in FIG.
HeightRate = h / H (3)
HeightVal
= INT (HeightRate * (N-1)) + 0.5 (4)
(However, INT (): rounded down to the nearest decimal point)
Each stage is labeled 0- (N-1).

Further, when the width of the in-line rectangle is used as a feature of the character line image, it is as follows.
WidthRate = w / H (5)
WidthVal
= INT (WidthRate * (N-1)) + 0.5 (6)
(However, INT (): rounded down to the nearest decimal point)
Each stage is labeled 0- (N-1).

  Returning to FIG. 5, subsequently, the symbol generation unit 26 converts each of the arrangement information quantized in step S <b> 333 into a fixed type symbol (step S <b> 334), and then proceeds to the process of step S <b> 34 in FIG. 4. .

  Hereinafter, the processing in step S334 will be described with reference to FIGS. As described above, the arrangement information acquired in step S333 characterizes the arrangement state of the corresponding in-line rectangle. In the process of step S334, a plurality of types of measurement results included in the quantized arrangement information are grouped into one symbol to make one in-line rectangle correspond to one symbol.

For example, three kinds of information of the height of the start point of the rectangle, the rectangle height, and the rectangle width are collected. Suppose that the height of the rectangular starting point (ys / H) is quantized to 15 levels, the rectangular height (h / H) is quantized to 8 levels, and the rectangular width (w / H) is quantized to 2 levels. As a result, as shown in FIG. 10, each piece of information has a rectangular start point height (ys / H) of 15 levels, 4 bits, and a rectangular height (h / H) of 8 levels, 3 bits, a rectangle. Since the width (w / H) has two stages, it can be expressed by 1 bit. Also,
4bits + 3bits + 1bit = 8bits
Therefore, all information can be stored in each bit of 1 byte. And the type of symbol that combines these three types of information into one,
15 stages × 8 stages × 2 stages = 240 types.

  By the way, if a plurality of features representing a rectangular arrangement state are regarded as each dimension of a multidimensional vector, the rectangle can be converted into one vector data (vector quantization) using each feature. As is well known, vector quantization is to obtain a small number of vector data representing them from a large variety of vector data. By labeling the obtained representative vectors in order, the vector data series can be converted into a simple one-dimensional symbol data series. For vector quantization, see “Vector quantization and information compression” (Corona) Allen Gersho, Robert M. et al. Familiar with Gray, Saburo Tazaki et al.

  Note that the type of information to be collected and the storage area for storing the information are not fixed in storage size, and it is possible to appropriately select and determine information suitable for specifying the character line to be identified. Needless to say. In addition, in FIG. 10, an example in which the height of the start point of the rectangle, the height of the rectangle, and the rectangle width are symbolized is shown. It is good.

  Through the above operation, the symbol generation unit 26 can convert the in-line rectangle included in each character line into a fixed number of symbols (labels). Therefore, the actual arrangement of the in-line rectangles can be regarded as a simple symbol arrangement as shown in FIG. Thus, the arrangement tendency of the symbol series can be recorded, which is equivalent to recording the arrangement tendency of the in-line rectangles.

  Returning to FIG. 4, the appearance frequency totaling unit 27 calculates the appearance frequency of a symbol series composed of a combination of a predetermined number of symbols for each of the collation image and the collation image for each arrangement information symbolized in step S <b> 34. Then, the calculation is performed (step S34), and the process proceeds to step S4 in FIG.

  Hereinafter, the process of step S34 will be described. After the arrangement information is symbolized, the search can be performed by a general search method as in the text search. That is, it is only necessary to obtain complete matching between symbol sequences for the collation image and the collation image. However, since the measurement result of the feature of the character rectangle differs depending on the reading error of the character line image, the symbol conversion result may not be the same even if the character lines are the same. Therefore, there is a possibility that the same character line image cannot be searched only by obtaining a complete match of the symbol series.

Therefore, in the process of step S34, not the complete matching of the symbol series but the correlation of the arrangement tendency of the symbol series is obtained. Specifically, the appearance frequency of a symbol series composed of a predetermined number of symbols in each of all the symbol series generated for the collation image and the collation image is calculated and aggregated.
Details will be described below.

  There is an n-gram model as a means for recording the tendency of arrangement. The n-gram model is a language model proposed by Claude Elwood Shannon. In this model, the appearance of a symbol in a sequence is influenced by the immediately preceding n symbols (n is a natural number). A stochastic process whose current state is determined depending on the n-th previous input is called an n-fold Markov process, and the n-gram model is also called an (n-1) -fold Markov model. In particular, the case of n = 3 is called trigram and is widely used.

Specifically, it is a model represented by the following formula (7). Furthermore, according to the equation (8), the appearance frequency of a symbol series (trigram) composed of a set of three symbols is calculated from all the symbol series of the collation image and the collation image.

On the other hand, the appearance frequency rank of the trigram is obtained, and the trigrams are aggregated in descending order of the appearance frequency. Table 1 shows an example of the trigram count result.

  In Table 1, the appearance frequency represents the frequency at which arrangement information representing the triple symbol series shown in the trigram, that is, three in-line rectangles, appears in this order in the entire symbol series. For example, in trigram [s013, s045, s032], the frequency of s032 appearing after s013, s045 is 324, and in trigram [s013, s064, s033], the frequency of s033 appearing after s013, s064 312 is shown. As described above, obtaining the tiger count result as shown in Table 1 for all the symbol sequences of the document image corresponds to obtaining (learning) the characteristics of each document image.

  Through the above operation, the appearance frequency totaling unit 27 derives the total result of the occurrence probability of the trigram as shown in Table 1 for each document image of the collation image and the collated image.

  Subsequently, the candidate image selection unit 28 obtains a totaling result (matching image totaling result) corresponding to the collation image derived in the process of step S3 and a totaling result (matching image totaling result) corresponding to the collated image. The top n checked images having higher correlation are selected as candidate images (step S4). Here, “n” is an integer of 1 or more, and an arbitrary value can be set.

When the collation image aggregation result and the collated image aggregation result are collated, since the number of in-line rectangles included in one character line often does not have the same value, it is not significant to compare the appearance frequencies themselves. Therefore, in step S4, the correlation between the collation image aggregation result and the collated image aggregation result is determined by using the rank correlation coefficient expressed by the following equation (9).
Rxy = 1− (6 * Σ (Rxi−Ryi) ^ 2) / (n * (n ^ 2-1)) (9)

  Here, n means the number of data, Rxi means the appearance frequency for each rank of the collated image aggregation results, and Ryi means the appearance frequency for each rank of the collated image aggregation results, and the difference between Rxi and Ryi for each rank. The sum of the squared values is calculated by Σ. The rank correlation coefficient is detailed in “Non-parametric method” (Baifukan) by Yanagawa.

  The candidate image selection unit 28 calculates the rank correlation coefficient Rxy for each appearance frequency included in the collation image aggregation result and the collation image aggregation result, and among the collation images, the value of Rxy is “1”. N images to be compared are selected as candidate images from the closest. Note that the rank correlation coefficient may be statistically tested, and if the maximum rank correlation coefficient does not show a significant value, it may be determined that there is no image to be collated similar to the collation image.

  Here, with reference to FIG. 12, the outline | summary of the process to step S1-S4 mentioned above is demonstrated. When a collation image X and a plurality of collated images Y are selected in the processes of steps S1 and S2, in the process of step S3, each of the in-line rectangles included in each character line constituting these document images is displayed. The symbols are converted into symbols based on the arrangement information, and all symbol sequences X1 for the matching image X and all symbol sequences Y1 for each matching image Y are generated. Then, by summing up the appearance frequency of the trigram in all the symbol sequences, a collation image aggregation result X2 corresponding to the collation image X and a collation image aggregation result Y2 corresponding to each of the collation images Y are derived. The Subsequently, the rank correlation coefficient Rxy is calculated based on the appearance frequency for each rank included in the collation image aggregation result A2 and the appearance frequency for each rank included in the collated image aggregation result B2. .

  In subsequent step S4, based on the value of the rank correlation coefficient Rxy for each image to be verified Y calculated in step S3, n images to be verified Y having the value close to “1” are used as candidate images. Will be selected.

  Returning to FIG. 3, the appearance position distribution deriving unit 29 executes an appearance position distribution matching process (step S5). Hereinafter, the appearance position distribution matching process in step S5 will be described.

  FIG. 13 is a flowchart showing the procedure of the appearance position distribution matching process. First, the appearance position distribution deriving unit 29 selects one candidate image as a target of the present process from the n candidate images selected in the process of Step S4 (Step S51).

  Subsequently, the appearance position distribution deriving unit 29 matches the trigrams between the two document images based on the collated image aggregation result of the candidate image to be processed in step S51 and the collation image aggregation result of the collation image. That is, a symbol series composed of a set of three symbols is selected (step S52). Here, the number of trigrams to be selected is not particularly limited, but it is preferable to select a trigram having a higher appearance frequency. Moreover, it is good also as an aspect which selects any symbol among the three symbols which comprise trigram.

  Next, the appearance position distribution deriving unit 29 determines the distribution state of the appearance position of the in-line rectangle corresponding to the symbol series selected in step S52 in the horizontal direction and the vertical direction of the document image for the collation image and the candidate image to be processed. Derived as a histogram (frequency distribution histogram) (step S53).

  As shown in FIG. 8, the in-line rectangle is represented by two points, a start point (Xs, Ys) and an end point (Xe, Ye). Therefore, when taking a distribution in the horizontal direction (X axis), a histogram may be generated for the start point Xs, and when taking a distribution in the vertical direction (Y axis), a histogram may be generated for the start point Ys.

  FIG. 14 is a diagram showing an example in which the distribution state of the existence positions of the in-line rectangles is represented by a histogram. As shown in the figure, the histograms in the horizontal direction and the vertical direction of the document image are derived for the in-line rectangles (K in the figure) that coincide between the document images of the collation image and the collation image. The total width for the histogram total is not particularly limited, but may be, for example, about the average value of the height of each character line cut out in the process of step S3.

  Returning to FIG. 13, next, the appearance position distribution deriving unit 29 collates both histograms obtained in step S53, and calculates the similarity (step S54). In this embodiment, as a method for comparing both histograms, the histogram values of the data section to which each of median (median value), mode (mode), and average belong are compared between the two histograms.

Specifically, the data sections of the histogram are sequentially numbered from the smallest coordinate, and the number of the data section to which the median, mode, and average belong is obtained. Here, if the data section number to which the median, mode, and average belong is expressed as (MedianClassNo, ModeClassNo, AvClassNo), the following four types of sets are obtained.
(MedianClassNoXaxQuery, ModeClassNoXaxQuery, AvClassNoXaxQuery) (10)
(MedianClassNoYaxQuery, ModeClassNoYaxQuery, AvClassNoYaxQuery) (11)
(MedianClassNoXaxDB, ModeClassNoXaxDB, AvClassNoXaxDB) (12)
(MedianClassNoYaxDB, ModeClassNoYaxDB, AvClassNoYaxDB) (13)

  “XaxQuery” means a horizontal histogram of the collation image, and the above equation (10) means a histogram value of the corresponding data section number in the horizontal histogram of the collation image. “YaxQuery” means a vertical histogram of the collation image, and the above equation (11) means a histogram value of the corresponding data section number in the vertical histogram of the collation image. “XaxDB” means a horizontal histogram of the image to be verified, and the above equation (12) means a histogram value of the corresponding data section number in the horizontal histogram of the image to be verified. To do. “YaxDB” means the vertical histogram of the image to be verified, and the above equation (13) means the histogram value of the corresponding data section number in the vertical histogram of the image to be verified. To do.

After calculating the above four types of values, the appearance position distribution deriving unit 29 uses the following equations (14) to (16) to calculate the matching image histogram and the matching image histogram in the vertical direction. Calculate the similarity of shapes.
MedianClassNoXaxDB + CA = MedianClassNoXaxQuery (14)
ModeClassNoXaxDB + CA = ModeClassNoXaxQuery (15)
AvClassNoXaxDB + CA = AvClassNoXaxQuery (16)

  In the above formulas (14) to (16), “CA” is a constant and is a value obtained from one formula (for example, formula (14)) to be processed first. The appearance position distribution deriving unit 29 determines whether or not the value of the constant CA is satisfied by the remaining two expressions, that is, the degree of deviation from the constant CA in the remaining two expressions, the histogram of the collation image, and the collation target It is calculated as the similarity of the shape with the histogram of the image. The degree of deviation from the constant CA can be derived, for example, by calculating CA ′ / CA. Here, CA ′ is CA + α (α is a deviation value from the constant CA), and the degree of deviation when completely matching, that is, the similarity is “1”.

Similarly, the appearance position distribution deriving unit 29 calculates the similarity between the shapes of the matching image histogram and the matching image histogram in the vertical direction using the following equations (17) to (19). .
MedianClassNoYaxDB + CB = MedianClassNoYaxQuery (17)
ModeClassNoYaxDB + CB = ModeClassNoYaxQuery (18)
AvClassNoYaxDB + CB = AvClassNoYaxQuery (19)

  In the above formulas (17) to (19), “CB” is a constant, and is a value obtained from the first formula (for example, formula (17)) to be processed first, similarly to the CA described above. The appearance position distribution deriving unit 29 determines whether or not the value of the constant CB is satisfied by the remaining two expressions, that is, the degree of deviation from the constant CB in the remaining two expressions, the histogram of the collation image, and the collation target It is calculated as the similarity of the shape with the histogram of the image. The degree of deviation from the constant CB can be derived in the same manner as for the constant CA described above.

  The appearance position distribution deriving unit 29 stores the similarity in the horizontal direction and the vertical direction calculated by the above procedure in the RAM 4 or the like in association with the verification target image. Here, the number of degrees of deviation derived with respect to the horizontal direction (or vertical direction) is for two formulas (or three formulas), but these may be individually retained as similarities, It is good also as an aspect which hold | maintains these average values as similarity.

  In the present embodiment, the similarity of the histogram shape is calculated for both the horizontal direction and the vertical direction of the document image, but it may be calculated only for one direction. In the present embodiment, the distribution state of the appearance position of the in-line rectangle is represented by a histogram. However, the present invention is not limited to this, and may be represented using, for example, a normal distribution.

  FIG. 15 is a diagram illustrating an example in which the distribution state of the position of the in-line rectangle is expressed by a normal distribution. As shown in the figure, when the normal distribution is used, the average μx, the standard deviation σx, the skewness, the kurtosis in the horizontal direction (X axis) are obtained from the aggregation result based on the start point of each in-line rectangle. Similarly, average μy, standard deviation σy, skewness, and kurtosis may be calculated in the vertical direction (Y-axis).

  In this case, as for the average value, there is a possibility that the image size may be different between the collation image and the collation image, so that direct comparison is not significant. In order to determine whether the shapes of the normal distribution match, it is only necessary to determine whether the standard deviation, skewness, and kurtosis are similar. For example, the standard deviation, skewness, and kurtosis of the search image are compared with the standard deviation, skewness, and kurtosis of the image to be searched. Judgment can be made.

  In addition, when the resolution of the collation image and the resolution of the collated image match, the number of dots constituting the same character is the same, but when the resolution is different, the number of dots is not the same. In other words, when evaluating the coincidence of the shape of the histogram or normal distribution, both numerical values may be used as they are when the resolution is the same, but when the resolutions are different, the numerical value based on the number of dots is used as it is. It cannot be used.

  Therefore, when the document images have different resolutions or when the resolution itself is unknown, it is necessary to normalize numerical values. In a general document image, the character size is the same for each paragraph, so that the character lines belonging to the same paragraph have the same line height. In addition, if the collation image is a part of the collation image, it is clear that there is a high possibility that the row height will be the same. Therefore, in the search target image and the search image, the line height of each character line is totaled, and the numerical value (average, mode, median) that defines the histogram is divided for the line height that occurs most frequently. The same applies to the normal distribution. Moreover, you may divide by the average value of the line height of each character line instead of the most frequent line height. Which to select is a design matter and may be determined according to the environment to be used.

  Further, even if the collation image is a part of the image to be collated, when only a specific part of the collation image becomes the collation image, the line height that appears most frequently in the entire image becomes the most frequent in the partial image. It is possible that the line height does not match. For example, in a document image in which the text line and the headline line have greatly different line heights, it is expected that the text line is overwhelmingly large as the number of lines of the entire image. If the partial image of the document contains only the heading line, the most frequent line becomes the heading line and does not match the line height estimated from the most frequent line of the whole image. It is clear that correct comparison results cannot be obtained even if normalized.

  In such a case, in both document images of the collation image and the collation image, the rectangular size is aggregated only for the matching in-line rectangle (symbol series), and based on the number of rectangular size dots that appear most frequently, This can be dealt with by normalizing numerical values (average, mode, median).

  Returning to FIG. 13, the appearance position distribution deriving unit 29 stores the similarity obtained in the process of step S54 in the hard disk 3 or the RAM 4 in association with the candidate image to be processed (step S55). Subsequently, the appearance position distribution deriving unit 29 determines whether or not all n candidate images selected in the process of step S4 have been processed in this process (step S56). Here, when it is determined that there is an unprocessed candidate image that is not a target of the main processing (step S56; No), the process returns to step S51, and one of the unprocessed candidate images is selected as a processing target. To do.

  On the other hand, if it is determined in step S56 that all candidate images have been processed (step S56; Yes), the process proceeds to step S6 in FIG.

  Returning to FIG. 3, the collation result selection unit 30 determines the candidate image having the highest similarity, that is, the value of the similarity based on the similarity of the n candidate images held in the RAM 4 or the like by the process of step S5. The candidate image that is closest to “1” is selected as the collation result (step S6).

  Subsequently, the display unit 31 displays the document image selected as the collation result in the process of step S6 on the display device 6 as the collation result for the collation image (step S7), and ends this process.

  FIG. 16 is a diagram for explaining the operation of the document collation process. In the drawing, D11 is a collation image, and shows a case where a partial image that is a part of a specific document image is selected as the collation image. D21 to D24 indicate four candidate images selected by the processing up to step S4. The collation image D11 is a partial image of the candidate image D24. That is, the candidate image D24 is a document image that is most similar to the collation image D11.

  As described above, in the process of step S4, it is possible to narrow down candidate images D21 to D24 as document images correlated with the collation image D11 by collating the symbol series corresponding to the in-line rectangle arrangement information. is there. In addition, the part represented by the rectangle K in the collation image D11 and the candidate images D21 to D24 means the in-line rectangle of the symbol series (or symbol) that matches in each document image.

  However, since the degree of similarity is only determined based on the appearance frequency of the symbol series in the process of step S4, the candidate image D24 is the document image most similar to the collation image D11, that is, the collation image D11 is the candidate image. It cannot be determined that it is a part of D24. Therefore, in the process of step S5, the candidate image D24 is the document image that is most similar to the collation image D11 by collating the relative positional relationship of the symbol series that matched in each document image, that is, the distribution state of the appearance positions. It becomes possible to specify.

  As described above, according to the present embodiment, the arrangement information representing the characteristics of the circumscribed rectangle in the character line is extracted from the collation image and the collation image, and these are quantized to a fixed stage to generate a symbol. Thus, it is possible to extract the characteristics of the character line without recognizing characters, and it is possible to select, as a candidate image, a predetermined number of images to be verified that have a high correlation with the verification image. Further, by comparing the distribution state of the appearance positions of the matching symbol series for the matching image and the candidate image, the similarity of the relative positional relationship of the symbol series can be determined. Similarity with a candidate image can be determined with high accuracy. As a result, even if a partial image in the document image is a document image to be collated, a document similar to the partial image based on the positional relationship of the circumscribed rectangle of the character image included in the partial image Images can be searched with high accuracy.

  It should be noted that the present invention is not limited to the above-described embodiment as it is, and can be embodied by modifying the constituent elements without departing from the scope of the invention in the implementation stage. In addition, various inventions can be formed by appropriately combining a plurality of constituent elements disclosed in the above embodiments. For example, some components may be deleted from all the components shown in the embodiment. Furthermore, constituent elements over different embodiments may be appropriately combined.

  For example, the program relating to the document matching process executed in the present embodiment may be stored on a computer connected to a network such as the Internet and provided by being downloaded via the network. In addition, a program related to document collation processing executed by the document processing apparatus 100 according to the present embodiment may be provided or distributed via a network such as the Internet.

  Further, the program relating to the document matching process executed in the present embodiment may be provided by being incorporated in advance in a storage medium such as a ROM.

  In the above-described embodiment, each functional unit illustrated in FIG. 2 is realized by cooperation with the CPU 1 and a predetermined program stored in the ROM 2. However, the present invention is not limited to this and is realized by a hardware configuration. It is good. Specifically, when the real-time property is regarded as important, it is necessary to speed up the processing. Therefore, a logic circuit (not shown) is separately provided, and various arithmetic processes are executed by the operation of the logic circuit. It is preferable to make it.

  In the above embodiment, the in-line rectangle is focused on as a unit smaller than the character line. However, the present invention is not limited to this, and other units are applicable. For example, by characterizing and quantizing image characteristics in units of characters (character images) or words, it is possible to symbolize and collate as described above. In this case, it is possible to cope by cutting out the character image based on the black image pixels and then using the circumscribed rectangle of the character image in character units or word units. In addition, what is necessary is just to use the well-known character extraction method used by OCR (Optical Character Recognition) etc. for the division | segmentation in a character unit or a word unit.

  As a typical character segmentation method, there is a method using projection. In this method, the number of black pixels in the horizontal direction is counted in the vertical direction, the distribution is obtained, and a portion where a certain number of black pixels is equal to or smaller than a threshold value is determined as a division position candidate. In addition, for the division position candidates, a compromise point is evaluated from the viewpoint of the character width estimated from the line height, the distance from the adjacent division position, the periodicity of the division position over the entire line, and the like. Is selected (the same applies to the vertical row).

  In addition, as another method of dividing into words, it is possible to easily realize a language that has a habit of dividing equally into European sentences based on a space between words. In this way, even when divided in units of characters, words, etc., it is possible to obtain a rectangle that circumscribes the image in that range, and use the start position and end point position of the circumscribed rectangle Thus, quantization can be performed in the same procedure as for the in-line rectangle.

  As described above, the document processing apparatus, the document processing method, and the document processing program according to the present invention are useful for a character processing apparatus that collates between document images, and particularly, a partial image that is a part of a document image is a collation target. It is suitable for a document processing apparatus that searches for a document image similar to this partial image.

It is the block diagram which showed the hardware constitutions of the document processing apparatus. It is the block diagram which showed the functional structure of the document processing apparatus. It is the flowchart which showed the procedure of document collation processing. It is the flowchart which showed the procedure of appearance frequency totaling process. It is a figure for demonstrating extraction of a character line. It is a figure for demonstrating extraction of a character line. It is a figure for demonstrating extraction of a character line. It is the flowchart which showed the procedure of the symbol production | generation process. It is the figure which showed the example of arrangement | positioning of the rectangle in a line. It is the figure which showed the example of arrangement | positioning of the rectangle in a line. It is a figure for demonstrating the example of a setting of the coordinate with respect to the rectangle in a line. It is a figure for demonstrating the arrangement | positioning state of the rectangle in a line. It is a figure for demonstrating the quantization of arrangement | positioning information. It is the figure which showed an example which symbolized the arrangement | positioning information quantized. It is a figure for demonstrating the outline | summary of a document collation process. It is the flowchart which showed the procedure of appearance position distribution collation processing. It is the figure which showed an example which expressed the distribution state of the presence position of the rectangle in a line with the histogram. It is the figure which showed an example which expressed the distribution state of the existing position of the rectangle in a line by normal distribution. It is a figure for demonstrating the outline | summary of a document collation process.

Explanation of symbols

100 document processing apparatus 1 CPU
2 ROM
3 Hard disk 4 RAM
DESCRIPTION OF SYMBOLS 5 Keyboard 6 Display apparatus 7 Optical disk drive 8 Communication apparatus 9 Scanner 10 Bus controller 21 Image input part 22 Collation image selection part 23 Rectangle extraction part 24 Line extraction part 25 Quantization part 26 Symbol generation part 27 Appearance frequency totaling part 28 Candidate image Selection unit 29 Appearance position distribution deriving unit 30 Verification result selection unit 31 Display unit

Claims (18)

In a document processing apparatus that performs collation between document images,
Based on a circumscribed rectangle for each character image included in the document image, character line cutting means for cutting out a character line connecting the circumscribed rectangles;
Quantization means for quantizing the arrangement information representing the characteristics of the circumscribed rectangle in the character line in a fixed stage;
Symbol generating means for symbolizing each of the quantized arrangement information into a fixed type of symbol;
Appearance frequency calculating means for calculating the appearance frequency of a symbol series composed of a predetermined number of symbol combinations;
The document image to be collated has a higher correlation by collating the appearance frequency calculated by the appearance frequency calculating means for the document image to be collated and a plurality of document images to be collated with the document image. A check target selection means for selecting a predetermined number,
Based on each piece of arrangement information corresponding to the symbol series matched between the document image to be collated and each document image to be collated selected by the collation target selection unit, any one of the pieces of arrangement information Or a distribution state deriving means for deriving the distribution state of the appearance position of the circumscribed rectangle that all represents, for each document image;
A collation target having the highest similarity by determining a similarity between the distribution state of the collation target document image derived by the distribution state deriving unit and the distribution state of the collation target document image; Collation result selection means for selecting a document image of
A document processing apparatus comprising:   The document processing apparatus according to claim 1, wherein the distribution state deriving unit derives a distribution state of appearance positions of the circumscribed rectangle in a horizontal direction and / or a vertical direction of the document image.   The document processing apparatus according to claim 1, wherein the distribution state deriving unit derives a distribution state of appearance positions of the circumscribed rectangle as a frequency distribution histogram.   The distribution state deriving means regards the distribution state of the appearance position of the circumscribed rectangle as a normal distribution, and derives an average, standard deviation, skewness, and kurtosis of the normal distribution. The document processing apparatus described.   The distribution state deriving means totals the sizes of the circumscribed rectangles of the character images included in the character lines in the document image to be collated and the document image to be collated, and an average value of the sizes or The document processing apparatus according to claim 4, wherein a numerical value defining the normal distribution is normalized by a mode value.   The distribution state deriving unit aggregates the sizes of circumscribed rectangles represented by the arrangement information corresponding to the symbol series that match between the document image to be collated and the document image to be collated. 5. The document processing apparatus according to 5. A document processing method executed by a document processing apparatus that performs collation between document images,
A character line cutting step for cutting out a character line obtained by connecting the circumscribed rectangles based on a circumscribed rectangle for each character image included in the document image;
A quantization step, wherein the quantization means quantizes the arrangement information representing the characteristic of the circumscribed rectangle in the character line in a fixed stage;
A symbol sequence generating means for symbolizing each of the quantized arrangement information into a fixed type of symbol; and
An appearance frequency calculating means for calculating an appearance frequency of a symbol series comprising a combination of a predetermined number of the symbols;
The collation target selecting means collates the appearance frequencies calculated by the appearance frequency calculation means for the document image to be collated and the plurality of document images to be collated with the document image, and has a higher correlation. A verification target selection step for selecting a predetermined number of document images to be verified;
The distribution state deriving means, based on each arrangement information corresponding to the symbol series matched between the document image to be collated and each document image to be collated selected in the collation target selection step, A distribution state deriving step for deriving a distribution state of appearance positions of circumscribed rectangles represented by any or all of the respective pieces of arrangement information for each document image;
The collation result selecting means determines the similarity between the distribution state for the document image to be collated derived in the distribution state deriving step and the distribution state for the document image to be collated, and the highest similarity A matching result selection step for selecting a document image to be verified as a matching result,
A document processing method comprising:   The distribution state deriving means derives a distribution state of appearance positions of the circumscribed rectangles in the horizontal direction and / or the vertical direction of the document image in the distribution state deriving step. Document processing method.   9. The document processing method according to claim 7, wherein the distribution state deriving unit derives the distribution state of the appearance position of the circumscribed rectangle as a frequency distribution histogram in the distribution state deriving step.   The distribution state deriving means regards the distribution state of the appearance position of the circumscribed rectangle as a normal distribution in the distribution state deriving step, and derives an average, standard deviation, skewness, and kurtosis of the normal distribution. The document processing method according to claim 7 or 8.   In the distribution state deriving step, the distribution state deriving unit totalizes the sizes of the circumscribed rectangles of the character images included in the character lines in the document image to be collated and the document image to be collated. The document processing method according to claim 10, wherein a numerical value defining the normal distribution is normalized by an average value or a mode value of the size.   In the distribution state deriving step, the distribution state deriving unit totalizes the sizes of circumscribed rectangles represented by the arrangement information corresponding to the symbol sequences that coincide in the document image to be collated and the document image to be collated. The document processing method according to claim 11. A computer that performs collation between document images
Based on a circumscribed rectangle for each character image included in the document image, character line cutting means for cutting out a character line connecting the circumscribed rectangles;
Quantization means for quantizing the arrangement information representing the characteristics of the circumscribed rectangle in the character line in a fixed stage;
Symbol generating means for symbolizing each of the quantized arrangement information into a fixed type of symbol;
Appearance frequency calculating means for calculating the appearance frequency of a symbol sequence consisting of a combination of a predetermined number of symbols in the symbol sequence;
The document image to be collated has a higher correlation by collating the appearance frequency calculated by the appearance frequency calculating means for the document image to be collated and a plurality of document images to be collated with the document image. A check target selection means for selecting a predetermined number,
Based on each piece of arrangement information corresponding to the symbol series matched between the document image to be collated and each document image to be collated selected by the collation target selection unit, any one of the pieces of arrangement information Or a distribution state deriving means for deriving the distribution state of the appearance position of the circumscribed rectangle that all represents, for each document image;
A collation target having the highest similarity by determining a similarity between the distribution state of the collation target document image derived by the distribution state deriving unit and the distribution state of the collation target document image; Collation result selection means for selecting a document image of
Document processing program characterized in that it functions as a computer program.   The document processing program according to claim 13, wherein the distribution state deriving unit derives a distribution state of appearance positions of the circumscribed rectangles in a horizontal direction and / or a vertical direction of the document image.   15. The document processing program according to claim 13, wherein the distribution state deriving unit derives the distribution state of the appearance position of the circumscribed rectangle as a frequency distribution histogram.   The distribution state deriving means regards the distribution state of the appearance position of the circumscribed rectangle as a normal distribution, and derives an average, standard deviation, skewness, and kurtosis of the normal distribution. The document processing program described.   The distribution state deriving means totals the sizes of the circumscribed rectangles of the character images included in the character lines in the document image to be collated and the document image to be collated, and an average value of the sizes or 17. The document processing program according to claim 16, wherein a numerical value defining the normal distribution is normalized by a mode value.   The distribution state deriving unit aggregates the sizes of circumscribed rectangles represented by the arrangement information corresponding to the symbol series that match between the document image to be collated and the document image to be collated. The document processing program according to 17.

Images