JPH0521266B2 - - Google Patents

Description

[Detailed description of the invention]

(Field of Industrial Application) The present invention relates to a document image file registration and search device that can perform a search using images in document image file searches. (Prior Art and its Problems) In recent years, document image file devices have been put into practical use, and various search methods have been studied. Among these methods, the method of searching using the image itself as an index (heading) is very effective in searching redundant image data. At this time, it is desirable to display a plurality of images at the same time and to be able to select a desired image from among them, but it is necessary that each index image is easy to see. However, this was not sufficient with conventional devices. For example, Japanese Patent Application No. 186361/1986 filed on November 20, 1986 entitled "Image File Search Device" has a configuration as shown in FIG.
-A reduction device was used to reduce the entire original image and create a headline image. The conventional example shown in FIG. 10 is detailed in the above-mentioned document 1, so a brief explanation will be given below. Images input from the image input device 101 are stored in the image memory 102 and displayed on the monitor 105. Current monitor 175 from operation console 107
When you enter a command to register the image displayed in a file, the command processing device 10
8 receives the command and central processing unit 109
to interpret the meaning of the command. Central processing unit 10
9 writes the image data in the image memory 102 to the magnetic disk device 104 via the magnetic disk interface 103 when it learns that the registration command has been input. In addition, the image memory 102
The image data contained therein is reduced by the reduction device 106a, and the reduced image data is also written to the magnetic disk device 104. Here, for example, the image is reduced to 1/4 x 1/4 using a diagonal of every other pixel and 4-point OR. The correspondence between a registered image and its reduced image is one-to-one, and the addresses of both on the disk are written in the address area on the disk in the order of registration. In other words, two addresses are written in the address area, which indicate the addresses of the reduced image and the registered image, respectively. Next, when a command ``reduced table of contents'' is inputted from the console 107, the command processing device 108 receives the command, and the central processing unit 109 interprets its meaning. When the central processing unit 109 learns that the reduced image table of contents command has been input, it reads the reduced images from the magnetic disk device 104 in the order of the addresses written in the address area on the disk. Further, as for the first image to be read out at this time, the central processing unit 109 remembers which image is currently being displayed, and displays that image first. In this example, since the size is reduced to 1/4 x 1/4, 16 reduced images are read out, the 16 reduced images are combined in the composition device 76b, and the reduced image table of contents is stored in the image memory 100.
and display it on the monitor 15. The conventional example has been described above, but in the conventional example, the image reduction processing is uniform, and therefore, depending on the image, the reduced image is crushed and becomes extremely difficult to see. In particular, in images such as general documents where the font size of the heading and the main text are almost the same, the title of the document itself may become unreadable in the worst case, making it difficult to search using small images. Ta. (Object of the Invention) An object of the present invention is to provide an image file registration/search device that eliminates the drawbacks of the conventional search device using index images and is equipped with index images that are easy to view and effective for searching. . (Structure of the Invention) According to the present invention, means for inputting a document image;
means for temporarily storing an input image; means for specifying a position on the image data; means for inputting registration and search commands; file management information storage means for decoding the commands and performing registration processing and search processing; The image storage means scans the registered image block by block during the registration process, checks at least the number of black pixels and the connections between black runs, and scans at least the background, black,
A means for classifying blocks according to which of image patterns characteristic of document images represented by thick lines, thin lines, and halftone dots the blocks correspond to, a means for storing classification results, and a means for storing classification results. Based on the difference in the frequency of appearance of the characteristic image pattern between the main text and headings in the document, statistical characteristics regarding the size, shape, and position of the headings, and the general formatting rules of the document. , a means for automatically extracting a heading character string from an original image, an image storage means for storing the original image and the heading image, and a file management information storage for storing file management information of the original image and the heading image. It is possible to realize a document image file registration and retrieval device characterized by comprising a means for displaying an original image or a plurality of index images. (Summary of the Invention) The present invention solves the problems of the prior art by adopting the above-described configuration. When registering a document image, the user first inputs an image from the image input means and temporarily stores it in the buffer memory. Next, by inputting a registration command from the command input means, the input image is recorded in the storage means. After that, the index extraction means scans the input image in microblock units, examines the number of black pixels in the block and the connections between black runs, and determines whether the corresponding block is an image pattern characteristic of the document image, such as the background, black , classify whether it corresponds to bold text, thin text, or halftone dots. This result is stored in auxiliary memory. Next, the heading extracting means uses the previous classification results to extract information that is predicted to be a heading based on how frequently the previous image pattern appears statistically between the main text and the heading part of the document. Extract the area and record the area information in another auxiliary memory. Furthermore, the heading extraction means selects the recorded heading candidate areas based on the statistical size and shape of the heading part and the position thereof, Return to the original image and extract the character string that becomes the image's heading. Thereafter, the index image creation means enlarges or reduces the area containing the extracted index character string so that it fits within a certain frame, and records it in the storage means as a index image separately from the original image. At the same time, a file management table is created or updated that describes the correspondence between the original image file and the index image file, and the relationship with other index image files, and is recorded in the storage means. At the time of image search, by inputting a search command, a stored file management table is referred to and a plurality of index images related within the table are displayed on the display means. If the desired image is not found among the displayed index images, the user requests a search for another index image. When the user finds a desired headline image, the entire image is displayed on the display device by inputting a request to access the image from the command input means or the position specifying means. According to the present invention, when registering a document image, it is possible to use a high-speed algorithm to extract the character string that becomes the heading of the image, and by storing this separately, a heading image that displays important information in an easy-to-read manner can be created. This allows effective document image retrieval. (Example) Hereinafter, an example of the present invention will be described in detail with reference to the drawings. FIG. 1 shows a specific example of a document image file registration and search device according to the present invention. In the figure, 11 is an image input device, 12 is an image memory,
13 is an image display device for displaying image data in the image memory; a cursor or pointer is displayed on the screen of this device, and can be moved by a screen pointing device 16; The command input device 14 and the command processing device 15 are configured by, for example, a personal computer. 16
The pointing device consists of a joy stick or a mouse. The pointing device transmits coordinate data or key input data by inputting an accompanying function key. The central processing unit 17, which is composed of a microprocessor, ROM, and RAM, decodes commands and controls registration processing and search processing using an algorithm that will be explained later with reference to FIG. 18-1 is a heading extraction device that extracts a heading string from a registered image; 18-2 is an attribute memory that records intermediate results of the heading extraction process; and 18-3 is an area that is a candidate for a heading string. 18-4 is an image processing device used to create a headline image, 19-1 is a magnetic disk device for storing file management tables, and 19-2 and 19-3 are each for storing image data. These are an optical disk interface device and an optical disk device. FIG. 2a shows the registration processing algorithm of the central processing unit 17. The central processing unit 17 that decoded the registration command is an optical disk interface device 19-.
2, an original image write signal (to put the disk side in write mode) and an image data transfer start signal are sent. After data transfer, optical disk interface device 19
-2 status word and optical disk 19
-3 write address is stored. Subsequently, a start signal is sent to the heading extraction device 18-1 to cause it to perform extraction processing. After the extraction process is completed, the heading extraction device 18
-1 receives heading area information and stores the heading area. Subsequently, an activation signal is sent to the image processing device 18-4 to activate it, notify it of headline area information, and cause it to create a headline image. When the creation of the headline image is completed, a headline image write signal and an image data transfer signal are sent to the optical disk interface device 19-2.
After data transfer, read the status word and memorize the write address. Finally, referring to the file management table on the magnetic disk device 19-1,
In the new field, write the registration number and address of the image you just registered. Also, in order to associate the already registered headline image with the currently registered headline image, for example, write the address of the previously registered headline image. At the same time, the address of the currently registered headline image is written in the column of the previously registered headline image as "the one registered immediately after", and the registration process ends. FIG. 2b shows the search processing algorithm of the central processing unit 17. The central processing unit 17 that decoded the search command is sent to the optical disk interface device 19.
-2, a headline image readout signal is sent to set the readout mode, and the currently stored headline image readout address and image memory write address are notified. Subsequently, a data transfer signal is sent, and after the transfer is completed, the file management table is referred to and, for example, the address of the related index image registered immediately after the read index image is set as the new read address. Until one screen of image memory is completely written,
Sequentially change the write address of the image memory,
Repeat this process. Once the headline image for one screen has been written, the central processing unit 17 enters a state of waiting for input from the command processing unit 15. When the headline screen change command is received, the file management table is referred to, the headline image read address is changed, and the process returns to the previous one. When the coordinate data from the pointing device 16 is received, the current position of the cursor/pointer (indicated as a pointer in the figure) on the image data is updated, and the cursor/pointer is moved on the screen. When a function key input (key input is shown in the figure) is received from the pointing device 16, the coordinates of the current position of the cursor/pointer are referred to and it is determined which index image currently being displayed corresponds to. The address of the original image corresponding to the relevant headline image is read out by referring to the file management table. Furthermore, by sending an original image read signal and a read address to the optical disk interface device 19-2, and sending an image data transfer signal,
The original image is displayed and the series of search processes is completed. The reason why the pointing device 16 is used to select a headline image is because the operation can be performed without taking one's eyes off the screen, and there is no need to be conscious of the number of the image to be selected. When registering a document image, the user first inputs an image from the image input device 11. The input image is stored in the image memory 12 and displayed on the display device 13. When an image registration command is input from the command input device 14, the central processing unit 17 decodes it through the command processing section 15. The central processing unit 17 writes the image data in the image memory 12 to the optical disk device 19-3 via the optical disk interface device 19-2. After this, the central processing unit 17
-1 (which can be implemented using a microprocessor, ROM, and RAM) is activated to extract a header string. An example of a character string extraction algorithm is shown in FIG. This is a top-down algorithm that extracts features from image data and estimates the index area on a block-by-block basis, and refers to pixels as necessary.It is resistant to noise, fast, and flexibly compatible with document images of undefined formats. can. First, the image data in the image memory 12 is read out in rectangular blocks. For example, the size of this block is 16 pixels x 16 pixels (2 mm x 2 mm) for input data with a resolution of 8 lines/mm.
mm). The number of black pixels and the number of connected black runs are measured as feature quantities in the read block data.
The number of connected black runs is an evaluation value that allows you to easily distinguish between areas that are made up of thin strokes, such as characters in the main text, and other areas. The black pixels are extracted in units of runs in the horizontal direction, and the connection of these runs is evaluated. FIG. 4a schematically shows how the black pixel connection region (hatched area in the figure) is separated into horizontal run units. Specifically, in the case of the block size of this embodiment, the number of connected black runs is determined as follows, for example. (1) For one black run in the horizontal direction, determine the number of connections within the line according to the run length, for example, as follows.

[Table] Runs with a run length of 2 or less are considered too short and are not given the number of connections. For run lengths of 3 to 10, the number of connections is 1 as the basic unit.
If the run length is 11 or more, it is assumed that two runs giving the above-mentioned number of connections 1 are connected, and the number of connections is 2. The above values are given for all black runs in the block, and the total number of connections within the line is taken as the number of connections of black runs in the horizontal direction. (2) Next, in (1), the number of connections within the line is 1 or 2.
It is checked whether or not the black run is connected to a black run with an intra-line connection count of 1 or 2 one line before. This is done by calculating, for example, the logical product of the previous line and the current line, and if even one pixel in the run of interest is connected to the run of the previous line, it is considered that there is a connection between the runs between the lines. This is called line-to-line connection. Each time a connection between lines is detected, one is added to the obtained number of horizontal black run connections. Through the above procedure, for example, the number of connected black runs of the black pixel pattern shown in FIG. 4a is determined to be 17. Note that the black graph file can evaluate not only connections within blocks but also connections between blocks. For example, as shown in FIG. 4B, two registers A and B store information about the boundaries of blocks that touch on the left and top of the block of interest, respectively. Register A contains the number of intra-line connections (1) for each line of blocks that touch on the left.
Or, if run 2 touches the block boundary of interest, 1 is stored, otherwise 0 is stored bit by bit. If in a certain line of the block of interest, a run with an intra-line connection number of 1 or 2 touches the block boundary on the left, and the bit corresponding to the same line in register A is 1, it is assumed that black runs are connected between blocks. , the number of black run connections is one. On the other hand, register B stores the pixel pattern of the bottom line of the blocks that touch on the upper side, that is, the line that touches the top line of the current block at the block boundary. 0 in the diagram
indicates a white pixel, and 1 indicates a black pixel. Checks the line connection with the first line of the current block, and increments the number of connections by 1 each time a connection is detected. It is possible to check connections at the right and bottom boundaries of blocks, but if connections in all four directions are checked for each block, connections at the boundaries will be checked twice, which is wasteful. The number of connected black runs determined above has a positive correlation with the number of black pixels, but by ignoring the connections of runs with short run lengths, the correlation between patterns with concentrated black pixels and patterns with dispersed black pixels increases. Gender changes. As a result, the number of connected black runs in a character area made up of thin strokes is kept low compared to the number of black pixels, making it easier to distinguish the character area from a character area made up of thick strokes. In addition, in halftone photographs,
This evaluation value is also effective for identifying photographic areas because there are many areas where patterns with short run lengths are continuous and the number of connected black runs is lower than that of fine letters. The method for evaluating the number of connected black runs shown in this example is the simplest example, and there are various other methods such as evaluation by scanning in the vertical direction, changing the run length threshold, and changing the method of adding the number of connections. can be applied. After the feature values are obtained, the heading extraction device 18-1
Based on these, it is determined to which image pattern characteristic of the document image each block belongs.
An example of determination when the target document image is a newspaper article will be explained with reference to FIG. When the number of black pixels is s and the number of connected black runs is g, when Γs<s ₁ , the pattern is considered to be close to the background and is defined as the "background". (s ₁ is a constant) When Γs≧s ₂ , the pattern is considered to be close to pure black and is called "black." (s ₂ is a constant) When Γs ₁ ≦s<s ₂ and g≧r ₁ ·s, the number of connected black runs is large compared to the number of black pixels, so it is called a “bold character”. (r ₁ is a constant) When Γs ₁ ≦s<s ₂ and r ₁・s>g≧r ₁・s−g ₁ , the number of connected black runs is slightly smaller than the number of black pixels, so “fine letters” ”. (g ₁ is a constant) When Γs ₁ ≦s < s ₂ and r ₁・s−g ₁ > g, the number of connected black runs is very small compared to the number of black pixels, so a short representative black run The pattern is "halftone dots". Each constant depends on the size of the block and how to set the threshold value for the number of connected black runs.
10), for example, s ₁ = 10, s ₂ = 128, r ₁ =
0.3, and g ₁ =6. The reason for setting the threshold value _s2 for the number of black pixels to be judged as "black" to half the number of pixels in a block (256) is that pattern classification of blocks with a larger number of black pixels than this has little meaning. This is because the frequency of appearance of "black" blocks in the article section, headline section, and photo section differs greatly depending on the threshold value. Statistically, in newspaper images, article sections have thin lines and halftone dots, headings have thick lines, black, and thin lines, photographs have halftone dots and black, and figures and tables have thick lines. It mainly consists of blocks such as ”, ”thin lines”, and ”background”. Of course, the above classification items (hereinafter referred to as attributes) can also be applied to general documents. In addition, the longest run within a block and its adjacency relationship can be stored and given an attribute called "ruled line." Since halftone photographs have pitches, the "halftone dot" attribute may be determined by examining this. The obtained attributes are two-dimensionally stored in the attribute memory 18-2. The heading extraction device 18-1 scans this attribute image and cuts out a non-background block cluster surrounded by background blocks into a rectangular shape. (This can be done with one scan.) This is because the heading character string is not limited to newspapers.
This is an efficient method because it generally has a considerable amount of blank space around it and can be surrounded by a rectangle. With the block size shown above (2 mm/2 mm), the article section of the newspaper is not separated by line spacing and becomes a group of non-background areas. When cutting out, surrounding background blocks are also included in the cutting to prevent cutting errors due to block formation. In addition, when determining the attribute, estimate the size of the heading from the frequency of appearance of "thick line" or "black", assume the corresponding surrounding blank area, and use the surrounding background block width as a threshold for cutting out. I can do it. Further, when "ruled line" is included in the attribute, since the ruled line of the article part has a certain pitch, the pitch can be detected when determining the attribute and the article part can be estimated. Although the above method is fast,
It is difficult to cut out headings that are close to other areas or have complex shapes. To compensate for this, we track the connection of the "thick line" and "black" blocks in the uncut area and cut out the blocks larger than a certain size. If necessary, the boundaries are determined by scanning pixel by pixel. The cut out rectangular area is stored in the heading candidate memory 18-
3 as a heading candidate. Heading extraction device 18
-1 evaluates these positions, areas, shapes, and attribute values within the region and excludes those that are inappropriate as headings. Attribute values are statistically evaluated based on the ratio of the number of blocks for each attribute. Basically, articles with many "thin lines" and little "black" are excluded as articles, and the distinction between photographs and headlines is based on a comprehensive evaluation of "halftone dots" and rectangular shapes. Exclude items that are difficult to use as headings according to layout rules, especially items that are close to a square. When rectangles are cut out in a nested state, unnecessary ones are excluded based on size, attributes, etc. If one heading is considered to be divided into multiple rectangles, these rectangles are unified and recombined. The determination of connection is made based on whether the center lines or circumscribed lines of the rectangles match, whether the distance is within two blocks, whether the most common attributes within the rectangles match, and so on. Furthermore, the image processing device 18-4 is activated to remove noise. For example, a vertical heading may be extracted without being separated from the summary text, but the summary text is removed by pixel-by-pixel projection. Additionally, headings with background patterns are removed. The reconstructed heading candidate areas are prioritized based on arrangement rules, size, etc. In the case of newspapers, larger headings and headings located in the upper right corner have higher priority. These priorities are determined by examining the layout pattern of the headline candidate areas. Layout patterns can be classified into about 10 types, including vertical headings only, horizontal headings only, and hook-shaped (combination of vertical and horizontal headings). Finally, one or more heading candidate areas are selected from those with high priority, and these are set as heading areas. Although the extraction processing described above has been described for newspapers, it can easily be applied to general documents. For example, a ROM containing extraction rules for character strings that serve as headings for general document images is prepared, and when the user inputs a "general document" command from the command input terminal 14, the central processing unit 17 ROM referenced by the extraction device 18-1
Change. The headline area extraction algorithm is performed using attribute images in the same way as in the case of newspapers.
The thresholds for evaluating the number of connections, determining attributes, and cutting out rectangles are changed to suit general documents. When cutting out text, a character string with a thin stroke similar to that of the main text may serve as a heading, so even if such a character string is extracted, it is not excluded. In large headings, the spacing between characters may be large, so change the distance conditions for recombination. If the line spacing is wide and each line in the sentence is to be cut out, exclude rectangles with line spacing. Horizontal writing or vertical writing is determined by detecting the line pitch or projecting the cut rectangle, and a priority order determination rule is used depending on each. The simplest way is to have one or more blank spaces around it, and for horizontal writing, give priority to the horizontal rectangle at the top left, and for vertical writing, give priority to the vertical rectangle at the top right. If the corresponding rectangle is not cut out, it is determined that there is no heading area, and the beginning of the document image (upper left for horizontal writing, upper right for vertical writing) is forcibly cut out and used as the heading image. As soon as horizontal or vertical writing is determined, the area to be scanned can be limited to the top half or right half, and if there are no headings there, forced extraction can be performed to improve processing efficiency if it is not a problem for the search. It is effective in this respect. When the heading area is determined, the central processing unit 17 activates the image processing device 18-4 and normalizes the corresponding area so that it fits within the frame of the heading image. The size of the headline image is, for example, 1/4×1/4 of the original image. If the heading area exceeds this frame, reduction processing is performed; otherwise, enlargement processing is performed. Now, make sure that the heading is in the center of the heading image. The processed data is stored in the index image area of the optical disk device 19-3 via the optical disk interface 19-2. The heading extraction device 18-1 shows the heading area on the display device 13 with a frame. It is also possible to add interactive processing to this series of processing. For example, if the user is not satisfied with the result of determining the heading area, the user may use the command input device 1.
Enter the "Change" command from step 4. As a result, the central processing unit 17 displays the candidate area of the heading candidate memory 18-3 in a frame on the display device 13, and the user uses the pointing device 16 (or command input device 17)
Select the area to be used as the index image. Alternatively, the user himself/herself sets the area using the instruction device 16. The subsequent processing is exactly the same. The addresses of the original image and index image on the optical disk 19-3 are written into the file management table on the magnetic disk 19-1. The structure of the file management table is shown in FIG. n-1, n, and n+1 are numbers assigned in the order of registration, where a is the original image address, b is the headline image address, and c, d are the addresses of the front and rear headline images that are logically adjacent to the headline image. . This adjacency relationship is, for example, the order of registration. This situation is shown in FIG. The index image of image A is B, the index image registered immediately before is C, and the index image registered immediately after is D. a, b,
c and d are addresses, respectively. When a user searches for an image, the user inputs a "search" command from the command input device 14. Having received this, the central processing unit 17 sequentially reads out the 16 headline images based on the currently stored headline image reading address while referring to the file management table on the magnetic disk 19-1, and reads out the 16 headline images sequentially from the image memory 12. The write address is sequentially changed and written, and the data is displayed on the display device 13. This is shown in FIG. The display order is A, B, C, D, E, F, G, H, I,
J, K, L, M, N, O, P. After this, the headline image read address points to the next image after the headline image P. If the desired image is not among the 16 displayed images, the user inputs a "previous" or "back" command from the command input device 14. The central processing unit 17 that received the "previous" command is displayed.
Among the 16 images, the address of the index image 16 images before the first index image (A in Figure 8) is used as the readout address, and the 16 images are read from there and displayed. "Before"
If there are no 16 index images in the area specified by , the first 16 index images in the file management table will be displayed. The central processing unit 17, which receives the "after" command, reads out 16 images based on the index image read address at that time (indicating the image next to P in FIG. 8) and displays them. If there are no 16 images in the area specified by "After", processing ends when the last image is displayed. The user continues the search using the "Search", "Previous", and "Next" commands, and when the desired heading image is obtained, the user moves the pointing device 16 and moves the cursor/pointer on the display device 13 to the corresponding heading image. Fit within the frame of. By pressing the function key of the indicating device 16, the central processing unit 17 learns that there is a search request for the index image where the cursor/pointer is currently located, and transfers the corresponding original image data to the optical disk device. 19-3 and display it on the display device 13. In this embodiment, the index image is 1/4×1/4 of the original image, but a index image having a different shape may be adopted. For example, it is conceivable to use a horizontally written heading character string as a short fence-shaped heading image as shown in FIG. The above is just an example of the present invention, and the scope of the claims is not limited to the configuration of the embodiment. (Effects of the Invention) By using the document image file registration and retrieval device described above, when registering a document image, a character string that becomes the heading of the image is extracted using a high-speed algorithm, and this is stored separately. It is possible to obtain a heading image that displays information in an easy-to-read manner, and to perform an effective document image search.

[Brief explanation of the drawing]

FIG. 1 is a block diagram of an image file registration and search device according to the present invention, FIGS. 2a and 2b are diagrams showing the registration processing algorithm and search processing algorithm of the central processing unit, respectively, and FIG. 3 is a heading area determination Figure 4a is a diagram showing the algorithm. Figure 4a is a diagram showing how the black pixel connected area is extracted in horizontal run units. Figure 4b is a schematic diagram showing how the connection information at the block boundary is stored in registers A and B. 5 is a diagram showing the attribute determination algorithm, FIG. 6 is a diagram showing the structure of the file management table, FIG. 7 is a diagram showing the relationship between the headline image and the original image shown in FIG. 6, Figures 8 and 9 are diagrams showing the search heading screen;
FIG. 10 is a block diagram showing an example of an image file registration and search device according to the conventional invention. In the figure, 11 is an image input device, 12 is an image memory, 13 is an image display device, 14 is a command input device, 15 is a command processing device, 16 is a screen instruction device, 17 is a central processing unit, and 18-1 is a header. Extraction device, 18-2 attribute memory, 18-3 heading candidate memory, 18-4 image processing device, 19-1
1 is a magnetic disk device, 19-2 is an optical disk interface device, 19-3 is an optical disk device, 1
01 is an image input device, 102 is an image memory, 10
3 is a magnetic disk interface device, 104 is a magnetic disk device, 105 is an image monitor, 10
6a is an image reduction device, 106b is an image synthesis device,
107 is an operation console, 108 is a command processing device, 1
09 is a central processing unit.

Claims (1)

[Claims] 1 means for inputting a document image, means for storing the input image, means for specifying a position on the image data, means for inputting a registration and search command, and a means for decoding the command and performing registration processing and search processing. means for causing the file management information storage means and image storage means described below to perform the above-mentioned image storage means; and during the registration process, the registered image is scanned block by block, at least the number of black pixels and the connections between black runs are checked, and at least the background, black, thick line,
means for classifying blocks according to which of image patterns characteristic of document images represented by thin lines and halftone dots the blocks correspond to; means for storing the classification results; and means for storing the classification results. use,
Extracting a heading character string from the original image based on the difference in the appearance frequency of the characteristic image pattern between the main text and the heading in the document, and statistical characteristics regarding the size, shape, and position of the heading. means for creating a heading image from the region containing the extracted character string; image storage means for storing the original image and the heading image; and file management information for the original image and the heading image. A document image file registration and retrieval device comprising: a file management information storage means for storing; and a means for displaying the original image or a plurality of index images.