JP2008204184A - Image processor, image processing method, program and recording medium - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To accurately identify and automatically classify the document type of an image. <P>SOLUTION: A document type identification processing part 115 identifies an attribute (form, drawing, etc.) of a registration image 112, registers the attribute in an image information DB 117, and registers the registration image 112 in an image DB 114. At the time of document retrieval, a display screen 119 classified by the document type is formed and displayed on a display device 101, and a user selects a category such as "form" from the display screen using an input device 103. A thumbnail list of the selected "form" is displayed. <P>COPYRIGHT: (C)2008,JPO&INPIT

Description

  The present invention relates to an image processing apparatus, an image processing method, a program, and a recording medium for recognizing and / or automatically classifying a document type of a document image. For example, a multifunction peripheral (MFP (Multi Function Printer)), a file server, The present invention relates to a technique suitable for an image processing program.

  There are devices such as electronic filing for digitizing paper documents using an input device such as a scanner, but they have been used exclusively for business purposes dealing with a large amount of paper documents. In recent years, due to the low price of scanners, the widespread use of MFPs equipped with scanning functions and the legislation such as the e-document method, the handling and convenience of offices have been recognized in offices, and paper documents have been digitized. On the other hand, the use of an image DB that integrates electronic document image data, photographic image data, document data created by an application such as a PC into a database (hereinafter referred to as DB), and the like, is also increasing. For example, even if it is necessary to save an original paper document, an image DB may be constructed for ease of management and retrieval.

  The above-mentioned image DB varies from a large-scale one installed by a server device to be accessed by a large number of users to a personal use for constructing a DB in a personal PC. For example, recent MFPs have built-in HDDs. A function for storing documents is provided, and an image DB based on the MFP is constructed.

Some document image DBs have a search function for searching a desired document image from a large number of document images. The current mainstream search function for document images generally performs full-text search or concept search using a character recognition result obtained by OCR (Optical Character Reader) processing as a keyword. However, in such text-based searches,
(1) Depends on OCR accuracy (2) Need search keyword (3) There is a problem that it is difficult to narrow down when the number of hits is large.

  With regard to (1), since it is currently impossible to obtain 100% correct answer by OCR, there is a problem that if there is an OCR mistake in the input search keyword portion, it will not hit. As for (2) above, text-based search is highly effective when searching for unknown things such as searching for HPs on the Internet, or when the keyword is clear, but for example several years ago When searching for a document with an ambiguous memory input, the search cannot be performed unless an appropriate keyword is conceived. In addition, if the entire page is a photograph or graphics and there is no text, it cannot be retrieved naturally. With regard to (3) above, it is difficult to rank the text-based search, so that a keyword hit is treated as equivalent. Therefore, when the number of hits is large, it is necessary to check a large number of hit document images one by one, resulting in poor usability.

  Therefore, as a technique for improving the problem of the text-based search, there is a search technique based on an approach different from the text-based search. This search method is a method of classifying document images into a plurality of categories and gradually narrowing down the document images, or a document image classification (document that searches for images by cooperation between a given category attribute and text-based search) Image attribute assignment).

  In the document image classification, for example, when a user registers a document image, a desired document category attribute is given, and the document image is classified based on a user classification system. This classification method is a method for obtaining an ideal classification category for each user. However, since a process for registering a large number of document images is complicated and a user's work load is large, a specialized operator is required. It is not common except when performing work.

  For example, when a category attribute is assigned to a scanned document using the scanning function of the MFP, it is performed on the operation panel of the MFP, but the scanning operation must be interrupted for each document and a category name must be input. Work efficiency is poor. As another method, there is a method of assigning a category attribute to each document after scanning all the documents. However, it is necessary to check each document, and the work burden on the user is large.

  As means for solving such a problem, there is a technique for automatically classifying document images. Although automatic classification of document images is difficult to obtain the ideal classification category for users, it is an effective means of searching for document images because it greatly reduces the burden on users and solves the problem of text-based search. is there.

  As a technique for automatic classification of document images (or automatic attribute assignment to document images), for example, there is Patent Document 1. In Patent Document 1, a scanned document image is divided into rectangular areas having the same attributes, layout analysis processing for determining the attributes of each area is performed, and layout characteristics (attributes of each block, their sizes, etc.) are recognized. As a result, the document image is automatically classified into any one of a table, a form document, a photo document, and other documents.

  Another example is Patent Document 2. In Patent Document 2, an input document image is compressed into a binary compressed image, a rectangle circumscribing a black pixel connected component of the binary compressed image is extracted, and the rectangle is classified into a character rectangle and other rectangles. Character areas and other areas are extracted by integrating each rectangle, and areas with the same attributes are divided into rectangular areas such as text areas, table areas, ruled lines, figure areas, and photo areas, and layout analysis is performed. Document images are classified using information on the type and number of areas as attributes.

JP 2001-101213 A JP 2003-178071 A

However, the document image automatic classification method based on the layout analysis process described above has the following problems.
(1) It is difficult to extract a region with high accuracy when the shape of the region is complicated, such as a key shape, or when the regions overlap each other.
(2) There is no recovery means when an area attribute is mistakenly identified (classification, attribute determination).

  That is, as in the prior art described above, the method of automatically classifying document images based on rectangular area information obtained by layout analysis cannot avoid the erroneous identification (classification determination) as described above.

  For example, when a document image to be classified into the “form” category is input and the table area of the document image is mistakenly identified as a figure area, the document image cannot be correctly classified as “form”. In addition, when the area cannot be correctly determined for a complicated layout such as a flyer, not only the area shape is erroneously identified, but also the area attribute is erroneously determined. It will be classified incorrectly. Even if a document image that has been misclassified in this way is searched, a correct search result cannot be obtained, and the search takes time and the search efficiency deteriorates.

  In the method described above, an area mainly composed of a line drawing used for a table or a figure is erroneously determined. Generally, the ruled line is determined by using a method of determining by the length of a black pixel connected component (black run) of a binarized image as in Patent Document 2 described above. When the line is binarized, the line is interrupted, causing erroneous determination and causing a problem in accuracy. Further, in a general method for detecting image areas having the same attribute as a single area, the same image is scanned a plurality of times in order to detect a circumscribed rectangle of the area as in Patent Document 2 described above. The process must be performed once, that is, for example, the process proceeds as the image is raster scanned from the upper left, and the process ends when the lower right of the image is reached. Requires processing. In addition, since the processing time differs for each document image, it is difficult to predict the processing time.

The present invention has been made in view of the above problems,
An object of the present invention is to provide an image processing apparatus, an image processing method, a program, and a recording medium that can identify a document type of an image with high accuracy and automatically classify it while shortening the processing time and simplifying the processing. There is.

  The present invention is an image processing apparatus for identifying a document type of an image, comprising: referring to pixels in a predetermined region of the image; a local region identifying unit that identifies whether or not the image is a line drawing region; and the local region identifying unit The main features include a feature amount calculating unit that calculates a feature amount from the identified line drawing identification result and a document type identifying unit that identifies the document type of the image according to the feature amount.

(1) In an image processing apparatus and method for identifying a document type of a bitmap image such as a scanned image, whether or not the image is a line drawing for each local area, and a feature amount is calculated from the local area identification result Since document types are identified and document classification is performed based on feature quantities, it is possible to identify document types with high accuracy even for document images with complex layout shapes. Since there is no need to scan, it can be realized by a single scan, so it is possible to reduce processing costs (processing time is shortened and complex processing is not required), and processing time is easy to predict and easy to use. realizable.
(2) Since the line drawing / character is identified for each local area and the dimension of the feature amount is increased, the identification accuracy is improved.
(3) Since the line drawing / character / photo is identified for each local area and the dimension of the feature amount is increased, the accuracy of identification is improved and the category of the document type including the photo can be identified.
(4) By performing frequency conversion, it is possible to concentrate power on a specific frequency coefficient when identifying a local area such as a line drawing, and the identification accuracy of the local area is improved.
(5) When recognizing document type attributes, a plurality of identifying means are used in parallel to allow duplication of identification results, thereby improving identification accuracy and searching for an image with different user subjectivity. In addition, the search can be performed by the classification display using the document type identification.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

Example 1:
FIG. 1 shows a system configuration of Embodiment 1 of the present invention. In FIG. 1, reference numeral 100 denotes a client device such as a personal computer (hereinafter referred to as a PC), a mobile terminal such as a PDA or a mobile phone. 101 is a display device such as a monitor, 102 is interpretation of a user instruction, communication with the server apparatus 110, an application program for controlling the display device 101, 103 is an input device such as a keyboard or mouse that is an instruction input means from the user, Reference numeral 104 denotes an external communication path such as a LAN or the Internet.

  110 has a database (hereinafter referred to as DB) for storing image data, identifies the document type of the input image data, registers the document image and its attribute information in the DB, and displays them in accordance with commands from the client device 100. A server device that generates a screen and outputs it to the client device 100, 111 is an interface (hereinafter referred to as I / F) with the external communication path 104, 112 is registered image data to be registered in the image DB 114, and 113 is a registered image 112 having a predetermined size or less. The thumbnail generation processing unit 114 generates a thumbnail image by scaling, the image database 114 stores the image data of the registered image 112, the image DB stores the thumbnail image data of the registered image 112, and 115 identifies the document type for the image data of the registered image 112. Document type identification processing unit 117 to be registered in the image DB 114 An image information DB for storing information for each image data. Here, the information is, for example, the file name of the registered image data, the creation date, the association information with the image data (the association information is, for example, an ID given to each image data when stored in the image DB 114) Or file name), document type (attribute), etc.

  A display screen control processing unit 118 generates a display screen for display on the client device 100 and controls the display screen according to the content of the screen control data 120, and 119 displays the display screen on the display device 101 of the client device 100. Display screen data 120 for this is screen control data designated and input by the client device 100. The dotted line in the figure represents the data flow at the time of image registration, and the solid line represents the data flow at the time of generating the display screen.

  FIG. 2 shows the configuration of the server apparatus 110 / client 100 apparatus. In FIG. 2, 201 is a CPU that performs calculations and processing according to a program, 202 is a volatile memory used as a work area for temporarily storing and holding data such as program code and image code data, 203 Is a hard disk (hereinafter referred to as HDD) for storing and accumulating image data and programs, and holds an image DB 114 and an image information DB 117. A video memory 204 is a data buffer for displaying on the monitor 205. The image data written in the video memory 204 is periodically displayed on the monitor 205. Reference numeral 206 denotes an input device such as a mouse or a keyboard, 207 denotes an external I / F that transmits and receives data via the external communication path 104 such as the Internet or a LAN, and 208 denotes a bus that connects each component.

  In the present embodiment, an example in which the server device 110 is configured by a server computer and processing such as display screen generation is realized by software. That is, the processing in the server is realized by an application program (not shown). The embodiment of the present invention is not limited to this, and may be configured to perform processing by hardware in an apparatus such as an MFP. Also, for example, a single PC or MFP can be used without adopting a server or client configuration. 1 may be configured in such a device.

  Hereinafter, an outline of the operation of this embodiment will be described. The system according to the first embodiment is roughly divided into two operations. One is an operation for registering a document image, and the other is an operation of “using a document image in a DB” for searching, browsing, and acquiring (downloading from a server) a desired document image. In using a document image, a desired document image is first searched, and then the image is viewed using an application viewer and stored in the user's PC. Hereinafter, an operation at the time of document image registration and a document image search operation according to the present embodiment will be described.

  FIG. 3 shows an operation flowchart when registering a document image. The document image registration operation will be described with reference to FIG. 1 (broken lines indicate the operation during registration) and FIG.

  In step S001, the user instructs the server apparatus 110 to register image data and the registered image data 112 to be registered from the client apparatus 100 via the application program 102.

  In step S002, the registered image data 112 is input to the server apparatus 110 along with file information such as a file name and a creation date via the external communication path 104, and an ID number is assigned to the image DB 114 via the external I / F 111. Registered. At the same time, the thumbnail generation processing unit 113 performs a scaling process on the registered image 112 to generate a thumbnail image having a predetermined size or less, and registers the image DB 114 with an ID number. When the registered image data 112 is a plurality of pages of image data, thumbnails are generated for each page.

In step S003, the registered image data 112 is input to a document type identification processing unit 115 described later, and the document type attribute is identified. The identified document type attribute is registered in the image information DB 117 together with the following image information data.
・ File name, creation date ・ Image data ID
・ Thumbnail image data ID
Document Type Attribute Note that the image information DB 117 can easily realize processing such as information registration, management, and search by using a general RDB (relational database). Further, if the image DB 114 and the image information DB 117 satisfy the above-described functions, a hierarchical data structure or the like may be constructed and stored in the same DB using a language such as XML (extensible Markup Language). Alternatively, each different server may be stored as a DB. In the image registration, the image data may be registered in the server device 110 directly from an image input device such as a scanner or a digital camera.

  FIG. 4 illustrates a configuration of the document type identification processing unit 115 according to the first embodiment. A pre-processing unit 301 performs image processing on the input registered image 112 to reduce the processing amount (processing cost) of subsequent processing and improve the accuracy of the local region identification processing unit 303. For example, processing for reducing the processing cost by reducing the number of pixels includes conversion processing of a color image into a gray image, reduction scaling processing, etc., and accuracy improvement processing includes γ correction for correcting the tone curve of the image. Processing includes, for example, color correction processing for converting a device-specific color space into a standard color space, spatial filter processing for correcting the MTF of an image, conversion processing to a predetermined resolution (similar to scaling processing), and the like.

  A frequency conversion processing unit 302 converts a plurality of pixels from real space to frequency space. There are various image frequency conversion processes, such as discrete Fourier transform (DFT), discrete cosine (DCT), and discrete wavelet transform (DWT). It is assumed that conversion (hereinafter referred to as DCT) is performed in units of 64 pixels of 8 pixels vertically and horizontally. DCT is realized by Equation 1.

Ｎ＝８：ブロック内水平（垂直）画素数
ｕ，ｖ＝０，１，〜Ｎ−１：ブロック内ＤＣＴ係数座標
ｘ，ｙ＝０，１，〜Ｎ−１：ブロック内画素座
ｆ（ｘ，ｙ）＝入力画素値
Ｆ（ｕ，ｖ）＝ＤＣＴ係数値

N = 8: number of horizontal (vertical) pixels in the block u, v = 0, 1, to N−1: DCT coefficient coordinates in the block x, y = 0, 1, to N−1: pixel position in the block f (x , Y) = input pixel value F (u, v) = DCT coefficient value

  Reference numeral 303 denotes a local region identification processing unit for identifying whether or not the image is a line drawing based on the DCT coefficient output from the frequency conversion processing unit 302. Reference numeral 304 denotes an identification result constituted by a line drawing identification result (on / off). An image 305 is a feature amount calculation processing unit that calculates a feature amount from the line drawing identification result image 304, and 306 identifies the document type attribute of the registered image 112 based on the feature amount calculated by the feature amount calculation processing unit 305. A type attribute identification processing unit 307 for identifying document type attribute information is identified.

The preprocessing unit 301 performs predetermined preprocessing on the registered image 112. In this embodiment, as an example, (1) when the input image is a color image, a gray image conversion process for converting to a gray image is performed in order to reduce the processing cost. Resolution conversion processing is performed to improve the accuracy.
(1) The gray image conversion process is a process of reducing the image data amount to 1/3 when a color image is input, and the processing cost is reduced. There are various conversion methods. If the input registered image 112 is an R (ed) G (reen) B (lue) image, the image is converted to luminance Y. A conversion formula from RGB to luminance Y is expressed by Formula 2.
Y = 0.299R + 0.587G + 0.114B Formula 2
However, Y: Luminance R: Red pixel value G: Green pixel value B: Blue pixel value In addition, you may use Formula 3 simply.
Y = (R + 2G + B) / 4 Formula 3
(2) The resolution conversion process is performed for the purpose of unifying the frequencies of the frequency conversion coefficients (DCT coefficients) output during the frequency conversion performed by the frequency conversion processing unit 302. As shown in the present embodiment, when frequency conversion is performed with an 8 × 8 pixel area fixed, the output spatial frequency is different when the resolution of the registered image 112 is different. By unifying the image resolution input to the processing unit 302, the processing cost is reduced as a whole. Even if the resolution conversion is not performed, the same effect can be obtained by changing the area (number of pixels) used for frequency conversion in accordance with the resolution of the registered image 112 in the frequency conversion processing unit 302.

  The resolution is preferably determined according to the area of the area used for frequency conversion. In the present embodiment, it is necessary to extract the line drawing separately from the characters in the area area of 8 × 8 pixels. For this reason, if the resolution is too high, there is a high possibility of misidentifying the side or bar of the character as a line drawing. Conversely, if the resolution is too low, for example, a line drawing such as a table portion and the character portion are combined. This increases the possibility of misidentifying line drawings.

  There are various methods for resolution conversion processing, so-called image enlargement / reduction processing, such as nearest neighbor method, linear interpolation method, cubic convolution method, etc. For example, a linear interpolation method that does not perform a thinning process is preferable. In the linear interpolation method, the MTF of the image is lowered. Therefore, MTF correction may be performed by spatial filter processing or the like after the resolution conversion.

  Next, the frequency conversion processing unit 302 performs DCT as shown in Equation 1 using a total of 64 pixels of 8 pixels in the vertical and horizontal directions, and outputs 64 DCT conversion coefficients. The local region identification processing unit 308 identifies whether the 8 × 8 pixel region is a line drawing from the DCT coefficient.

  A local region identification method using DCT coefficients will be described below. FIG. 5 is a diagram in which F (u, v), which is the DCT coefficient of Expression 1, is two-dimensionally arranged. Mainly, the power of the DCT coefficient concentrates when vertical lines, horizontal lines, and diagonal lines are input. Are shown in black. Coordinates (0, 0) represent a direct current (DC) component.

  FIG. 5A shows DCT coefficients that are concentrated when a vertical line image is input, and a vertical line component is used for convenience. 5B corresponds to a horizontal line image corresponding to the horizontal line image, and FIG. 5C corresponds to a diagonal line component corresponding to the diagonal line image. The white portions other than the DC component in each figure are non-vertical lines, non-horizontal lines, and non-diagonal line components, respectively.

  As shown in FIG. 5, the DCT coefficient has a property that a coefficient at which power is concentrated with respect to the input image can be almost predicted, and the non-linear component shows a low value. By using such a property, it becomes possible to distinguish a line drawing from the rest from the DCT coefficient.

  Specifically, as shown in the flowchart of FIG. 13, a total value or maximum value of “absolute value of DCT coefficient” for each line component and non-line component is calculated in advance, and the line component is calculated for each of the vertical, horizontal, and diagonal lines. The non-linear component difference is compared with a predetermined value (steps 21 to 23), and when the difference is equal to or larger than the predetermined value, it is identified as a line drawing.

  As another method, a learning machine such as a support vector machine (hereinafter referred to as SVM) is used to input an absolute value of an alternating current (AC) component other than a DC component as a feature amount, and is calculated from a line drawing in advance. A line drawing and the others are discriminated by using a model in which DCT coefficients are learned. Such a property shows the same property in other frequency conversion methods. For example, the same identification can be performed by using DWT for frequency conversion processing and performing local region identification using a DWT coefficient.

  FIG. 6A shows a registered image 112 of a form, and FIG. 6B shows a line drawing identification result image (local area identification result) 304 for the registered image 112. In FIG. 6B, the line drawing portion is represented as black and the other portion is represented as white binary image. FIG. 6 shows an example in which local area identification is performed on a rectangular area unit obtained by performing DCT of 8 pixels in the vertical and horizontal directions on the input image. Therefore, the image of the identification result is 1/8 in length and width compared to the image input to the frequency conversion processing unit 302.

  As described above, when local region identification using frequency conversion processing is performed using a plurality of pixels in a rectangular region, whether the identification unit is performed in a rectangular unit or a pixel unit differs depending on the accuracy to be obtained. If processing is performed in units of rectangular areas as in this embodiment, the amount of processing can be greatly reduced and processing costs can be reduced, but processing for identifying in units of pixels (that is, by overlapping pixels for DCT). DCT is performed for each pixel of interest, and the identification result is output for each pixel.The line drawing identification result image 304 has the same size as the image input to the frequency conversion processing unit 302), and the accuracy is low.

  Next, the feature amount calculation processing unit 305 calculates a feature amount from the line drawing identification result image 304. The image feature amount includes an image moment, texture, edge amount, and the like. As a calculation method, the image is divided into several regions, and the feature amount is calculated for each region. It is also possible to use image arrangement information as a feature amount.

  In the present embodiment, an example in which a higher-order local autocorrelation feature is used is shown. When higher-order local autocorrelation feature values are used for a binary image, the feature values are calculated for each of the 25 types of 3 × 3 local patterns shown in FIG. That is, a 25-dimensional feature amount is calculated. The calculation of each feature amount is the product of the pixel values corresponding to the local pattern (the product of the pixels corresponding to “1” of the pattern in FIG. 7. In the case of a binary image, the pixel corresponding to “1” of the local pattern This is realized by scanning all pixels and adding them together. However, since the calculated feature amount is affected by the image size, normalization is performed when images of different sizes are input to the registered image 112. In this way, the line drawing identification result identified by the local area identification processing 303 is converted into a 25-dimensional feature amount.

  A 25-dimensional feature amount is input to the type attribute identification processing unit 306, and the type attribute identification processing unit 306 identifies the document type attribute 307 of the registered image 112. The type attribute identification processing unit 306 according to the present exemplary embodiment identifies document types such as “form”, “drawing”, and “others” as an example.

In order for the type attribute identification processing unit 306 to identify the document type attribute, it is only necessary to identify from the input 25-dimensional feature amount by paying attention to the feature amount in “form” and “drawing”. FIG. 8 is a graph showing the feature amount when each document image of a form, a drawing, and a paper is input. Each graph in FIG. 8 shows a characteristic tendency of each document type. That is,
-The form has many feature quantities in the dimensions of No. 3 and No. 6 and has few feature quantities in the dimensions from No. 10 to No. 25.
-Drawings generally have more feature values (higher numerical values) than feature values of forms.
・ The paper has less overall feature values (lower numerical values) than the feature values of forms and drawings.

  Therefore, by using these features, “form” and “drawing” can be identified. The identification is performed by comparing feature value values or by threshold processing for determining whether or not a predetermined threshold value is exceeded.

  FIG. 9 is a flowchart of the type attribute identification processing unit 306. In step S011, the total value of the feature amounts No2 to No25 is calculated, and it is compared whether or not it is a predetermined threshold value or less. If the value is equal to or less than the predetermined value, the “other” attribute is selected.

  In step S012, if the total value of the feature amounts No2 to No25 is larger than the predetermined value, the ratio of the feature amounts No3 + No6 and No4 + No5 is compared to be less than or equal to the predetermined value. If it is larger than the predetermined value, the “form” attribute is selected.

  As another method, a document type may be identified by using a learning machine such as SVM, inputting a 25-dimensional feature value, and using a model learned in advance.

  Through the above process, the document type identification process of the first embodiment is completed, and the document type attribute of the registered image 112 is identified.

  FIG. 10 is an operation flowchart when a document image is searched using the document classification process based on the document type identification process.

  In step S <b> 101, the user uses the application program 102 in the client device 100 to instruct the document classification to the server device 110. The instruction means at this time displays a thumbnail list display screen as shown in FIG. 11 on the display device 101 of the client apparatus 100, for example.

  In FIG. 11, 401 is a classification radio button for instructing a display screen by document classification, 402 is a thumbnail radio button for instructing thumbnail display, 403 is a frame for displaying an image, and 404 is a thumbnail of the image. A plurality of image thumbnails 404 are displayed on the frame 403, and thumbnails used in a general image DB are displayed in a list. Normally, a large number of images are registered in the image DB 114, but for image thumbnails that cannot be displayed at once, the upper and lower sliders are provided on the frame to scroll the images, or a page feed function is provided to display a group of displayed images. Browse by changing.

  When the user clicks the classification radio button 401 using a pointing device such as a mouse of the input device 103, the screen control data 120 that is a document classification instruction is transferred to the server side via the external communication path 104.

  In step S102, when the server apparatus 110 receives the screen control data 120, which is a document classification instruction, the display screen control processing unit 118 counts the number of document images for each document type identification data (hereinafter, classification category) in the screen information DB 117. To determine the layout of the display screen and the document image data to be displayed.

  Next, document image data to be displayed or a thumbnail of the document image data is input from the image DB 114 to generate a classification result display screen 119, which is transmitted from the external I / F 111 to the client apparatus 100 via the external communication path 104. .

  FIG. 12 shows an example of a display screen for the classification result. Characters such as “form” and “drawing” in FIG. 12 represent categories. Reference numeral 411 represents a classification category and shows an example of classification into three categories. In addition, the size of the ellipse 411 schematically represents the number of documents included in each category (the number of documents may be directly expressed as a number), and the thumbnails in the categories are document images included in each category. Is due to. The image thumbnail displayed here displays all images when the number of images registered in the image DB 114 is small, and displays some representative images in each category otherwise. By reducing the number of display images in this way, the display time on the client device 100, the transfer time via the external transfer path, and the processing time on the server device 110 are both shortened. If a sufficient processing speed is obtained, all the images may be superimposed or displayed with a slider.

  There are various methods for creating the display screen and the communication method between the server and client as described above. As a commonly used method, by using the World Wide Web-based technology using the server device 110 as a Web server. It becomes feasible. The display screen 119 is described in HTML (Hyper Text Markup Language), and the application 102 may use a general Web browser.

  In step S <b> 103, the client apparatus 100 displays a display screen 119 on the display device 101.

  In step S <b> 104, the user selects a category close to the document image being searched using the input device 103, and transmits the selected category data to the server apparatus 110. For example, it is assumed here that the “form” category of FIG. 12 is selected. As a selection method, a category is selected by clicking in the category with a pointing device such as a mouse. It is also possible to select / deselect for each category, and to be able to select a plurality of categories.

  In step S105, the server apparatus 110 receives an instruction for selecting a category in the display screen control processing unit 118, and creates a display screen for a thumbnail list in the selected category (form category) as shown in FIG. To the client device 100.

  In step S <b> 106, the client apparatus 100 displays the display screen 120 illustrated in FIG. 11 on the display device 101.

  Thus, by identifying the document type and classifying and displaying the document type, the user can narrow down the search target images while confirming the features such as the overview of the image even in the image DB having a large number of document image registrations.

  As described above, according to the present embodiment, in the image processing apparatus and method for identifying the document type of a bitmap image, whether or not the image is a line drawing for each local area is identified, and the feature amount is calculated from the identification result of the local area Since the document type is identified and the document is classified based on the calculated feature amount, the document type can be accurately identified even in a document image having a complicated layout shape. In addition, since it is not necessary to scan the image a plurality of times and it can be realized by a single scan, the processing cost can be reduced, and the processing time can be easily predicted and easy-to-use processing can be realized.

Example 2:
In the second embodiment, the local area identification processing unit 303 identifies a line drawing, identifies a character image, and improves the accuracy of the document type attribute. The system configuration of the second embodiment is the same as that of the first embodiment.

  The DCT coefficient of the character image includes a plurality of line components as compared with that of the line drawing. FIG. 14 is a flowchart of the local area identification processing unit 303 according to the second embodiment. In the first embodiment, when the difference between the line component and the non-linear component is equal to or larger than the predetermined value, the line drawing is identified. However, in the second embodiment, the maximum value of the AC component is calculated in advance, and the maximum value is less than the predetermined value. If there is (yes in step S031), “other” is set, the maximum value is not less than or equal to the predetermined value (no in step S031), and if the difference between the line component and the non-linear component is greater than or equal to the predetermined value (in step S032). yes), “line drawing”, and when the difference between all line components and non-line components is equal to or smaller than a predetermined value (no in steps S033 and S034), “line”. Note that a learning machine such as SVM may be used as in the first embodiment.

  In the local area identification processing unit 303 according to the second embodiment, “line drawing” and “character” are identified, so that there are two types of identification result images: a line drawing identification result image and a character identification result image. FIG. 6C shows a character identification result image in the second embodiment. The line drawing identification result image is the same as that in the first embodiment.

  In the second embodiment, the feature amount calculation processing unit 305 calculates a feature amount for each identification result image 304 as in the case of the first embodiment. For example, when calculating higher-order local autocorrelation feature quantities, 25-dimensional feature quantities are calculated from the line drawing identification result image, 25-dimensional feature quantities are calculated from the character identification result image, and these 50-dimensional feature quantities are calculated. Is input to the type attribute identification processing unit 306.

  In the second embodiment, as an example, an example in which four types of attributes of “form”, “drawing”, “book”, and “other” are identified in the input registration image 112 is shown. FIG. 15 illustrates a configuration of the type attribute identification processing unit 306 according to the second embodiment. In the second embodiment, an example is shown in which the type attribute identification processing is identified by a classifier using a learning machine such as SVM.

  In FIG. 15, reference numeral 501 denotes a 50-dimensional feature quantity output from the feature quantity calculation processing unit 305, 502 denotes a form identifier for identifying whether or not the document attribute type is a form based on the inputted feature quantity 501, and 503 denotes A form model that has been learned and created based on the feature amount based on the learning data of the form image in advance, 504 is the identification result of whether or not it is a form, 505 is whether or not the document attribute type is a drawing based on the input feature quantity 501 A drawing classifier 506 for learning based on a feature amount based on learning data of a drawing image and created in advance, 507 for identifying whether or not it is a drawing, and 508 based on the inputted feature amount 501. A book discriminator for identifying whether or not the document attribute type is a book, 509 is a book model that has been previously learned and created based on a feature amount based on learning data of a book image, and 510 is a book or not. It is the identification result.

  The SVM is a discriminator that automatically creates a discrimination plane by spatially projecting multi-dimensional feature values and discriminates using a discrimination function called a kernel. In order to actually identify, it is necessary to perform learning using learning data in advance and to model learning results. The learning result modeling result is stored in a file called a model. In learning, a set of “feature amount” and “correct answer of identification result” actually used for identification is prepared and learned. Once learning is performed, high-speed identification is possible only by using a model file. In this embodiment, the learning results of forms, drawings, and books are stored in the respective model files 503, 506, and 509.

  Hereinafter, the operation of the type attribute identification processing unit 306 according to the second embodiment will be described. The 50-dimensional feature quantity output from the feature quantity calculation processing unit 305 is input to the form discriminator 502, the drawing discriminator 505, and the book discriminator 508, respectively, and identified using the models 503, 506, and 509, Each identification result is output. If all the identification results are negative (off), the “other” attribute is selected. As the document type attribute, one attribute is usually given to one registered image. However, for example, even when the user identifies and classifies an image, it may be confused about what attribute should be given. If the user's subjectivity is divided and an image is searched using the classification display, the user's subjectivity is different from the identification result, and the search result desired by the user cannot be obtained.

  Therefore, in the case of an image in which the user's subjectivity is divided in advance, it is desirable to provide a plurality of attributes so that the search target document is hit regardless of which category the user selects.

  In order to realize this, it is necessary to prepare and learn learning data that should overlap type attributes during learning. FIG. 16 is a diagram illustrating learning of duplicate data. A rectangle 601 in FIG. 16 represents an image, and 602 and 603 represent categories having the same attribute. That is, the model of FIG. 16 represents the positions that each image can take on the map when the image is mapped in two dimensions using the feature amount. The black image 604 in FIG. 16 (hereinafter referred to as a duplicate image) is included in both categories 602 and 603 of the form and the drawing. When learning these, the form is learned as a correct answer when learning the form model 503, and the drawing is learned as a correct answer when learning the drawing model 506. In this way, by learning the form data redundantly, an image that is unclear in both the form and the drawing is identified in both the form and the drawing.

  Normally, a document image such as a book is often input in a plurality of pages. However, as in the present embodiment, processing is performed in units of pages, identified in units of pages, and finally has many type attributes identified. A thing is used as a representative document type attribute of a multi-page document image. In addition, the attributes of each page and a plurality of pages may be different from each other to generate a “page unit display screen” and to generate a “document unit display screen”.

  As described above, according to the present embodiment, the line drawing / character / others are identified for each local region, and the feature quantity dimension is increased, so that the identification accuracy is improved. Also, when identifying type attributes, in addition to improving the accuracy of identification by allowing a plurality of classifiers to be used in parallel and allowing the identification results to be duplicated, search for an image in which the subjectivity of the user is divided. Even in this case, the search can be performed by the classification display using the document type identification.

Example 3:
The third embodiment is an embodiment in which the local region identification processing unit 303 identifies line drawings and character images and also identifies photographic images to improve the accuracy of document type attributes. The system configuration of the third embodiment is the same as that of the first embodiment.

  The DCT coefficient of a photographic image basically has a lower AC component power than that of a line drawing or character image, and the DCT coefficient power is distributed over a relatively wide range. FIG. 17 is a flowchart of the local region identification processing unit according to the third embodiment. In the third embodiment, the maximum value of the AC component is compared with the first threshold value. When the maximum value of the AC component is equal to or less than the first threshold value (yes in step S041), “other” is identified. When the maximum value of the AC component is not less than or equal to the first threshold value (no in step S041) and the maximum value of the AC component is less than or equal to the second threshold value (yes in step S042), it is identified as “photograph”. .

  In the local area identification processing unit 303 according to the third embodiment, in order to identify “line drawing”, “character”, and “photograph”, there are three types of identification result images: a line drawing identification result image, a character identification result image, and a photo identification result image. Thus, for each identification result image 304, the feature amount calculation processing unit 305 calculates a feature amount as in the second embodiment. For example, taking the case of calculating a higher-order local autocorrelation feature amount as an example, a 75-dimensional feature amount is calculated and input to the type attribute identification processing unit 306. The type attribute identification processing unit 306 identifies a plurality of document types using a plurality of classifiers in the same manner as in the second embodiment using 75-dimensional feature values.

  In the third embodiment, since the photographic image is identified by the local area identification processing unit 303, it is possible to identify a document type including a character and a photograph such as a catalog or a flyer, an image of only a photograph, and the like.

  As described above, according to the present embodiment, line drawings / characters / photos / others are identified for each local region, and the feature dimension is increased. It is possible to identify the category of the document type including

The system configuration | structure of Example 1 of this invention is shown. The structure of a server apparatus / client apparatus is shown. The operation | movement flowchart at the time of document image registration is shown. The structure of a document type identification process part is shown. The distribution of DCT coefficients is shown. The result of local region identification output is shown. A pattern for higher order local autocorrelation is shown. Indicates the feature amount of each image of the form, drawing, and paper. 6 is a flowchart of a type attribute identification processing unit according to the first embodiment. 6 is an operation flowchart when retrieving a document image according to the first exemplary embodiment. An example of a thumbnail list display screen is shown. An example of classification display is shown. 3 is a flowchart of a local region identification processing unit according to the first embodiment. 6 is a flowchart of a local region identification processing unit according to the second embodiment. The structure of the type attribute identification process part of Example 2 is shown. It is a figure explaining learning of duplication data. 10 is a flowchart of a local area identification processing unit according to a third embodiment.

Explanation of symbols

DESCRIPTION OF SYMBOLS 100 Client apparatus 101 Display device 102 Application program 103 Input device 104 External communication path 110 Server apparatus 111 External interface 112 Registered image data 113 Thumbnail generation process part 114 Image DB
115 Document Type Identification Processing Unit 117 Image Information DB
118 Display Screen Control Processing Unit 119 Display Screen Data 120 Screen Control Data

Claims (14)

  An image processing apparatus for identifying a document type of an image, wherein local line identifying means for identifying whether or not the image is a line drawing area with reference to pixels in a predetermined area of the image, and the line drawing identified by the local area identifying means An image processing apparatus comprising: a feature amount calculating unit that calculates a feature amount from an identification result; and a document type identifying unit that identifies a document type of the image according to the feature amount.   An image processing apparatus for identifying a document type of an image, wherein a local area identifying means for identifying a line drawing area and a character area with reference to pixels of a predetermined area of the image, and the line drawing identified by the local area identifying means An image processing apparatus comprising: a feature amount calculating unit that calculates a feature amount from an identification result and a character identification result; and a document type identifying unit that identifies a document type of the image according to the feature amount.   An image processing apparatus for identifying a document type of an image, wherein a local region identifying unit for identifying a line drawing region, a character region, and a photographic region with reference to pixels in a predetermined region of the image, and the local region identifying unit A feature amount calculating unit that calculates a feature amount from the line drawing identification result, a character identification result, and a photo identification result, and a document type identifying unit that identifies a document type of the image according to the feature amount. An image processing apparatus.   The local area identifying means performs frequency conversion on a plurality of pixels in the predetermined area, outputs a plurality of frequency conversion coefficients, and identifies each area based on the frequency conversion coefficients. The image processing apparatus according to 1, 2, or 3.   4. The image processing apparatus according to claim 1, wherein the document type identification unit includes a plurality of classifiers for each document type, and allows duplication of identification results by the plurality of classifiers.   The document type identification unit performs identification based on learning data learned in advance using a feature amount for each document type, and the learning data includes data included in a plurality of document types. 5. The image processing apparatus according to 5.   An image processing method for identifying a document type of an image, wherein a local region identification step for identifying whether or not the image is a line drawing region with reference to pixels of a predetermined region of the image, and the line drawing identified by the local region identification step An image processing method comprising: a feature amount calculating step for calculating a feature amount from an identification result; and a document type identifying step for identifying a document type of the image according to the feature amount.   An image processing method for identifying a document type of an image, wherein a local region identifying step for identifying a line drawing region and a character region with reference to pixels of a predetermined region of the image, and the line drawing identified by the local region identifying step An image processing method comprising: a feature amount calculation step of calculating a feature amount from an identification result and a character identification result; and a document type identification step of identifying a document type of the image according to the feature amount.   An image processing method for identifying a document type of an image, wherein a local region identifying step for identifying a line drawing region, a character region, and a photographic region with reference to pixels in a predetermined region of the image, and the local region identifying step A feature amount calculating step of calculating a feature amount from the line drawing identification result, the character identification result, and the photo identification result, and a document type identifying step of identifying the document type of the image according to the feature amount. Image processing method.   The local area identifying step performs frequency conversion on a plurality of pixels in the predetermined area, outputs a plurality of frequency conversion coefficients, and identifies each area based on the frequency conversion coefficients. The image processing method according to 7, 8 or 9.   The image processing method according to claim 7, wherein the document type identification step includes a plurality of identification steps for each document type, and allows duplication of identification results by the plurality of identification steps.   The document type identifying step performs identification based on learning data learned in advance using a feature value for each document type, and the learning data includes data included in a plurality of document types. 11. The image processing method according to 11.   A program for causing a computer to implement the image processing method according to any one of claims 7 to 12.   A computer-readable recording medium on which a program for causing a computer to implement the image processing method according to any one of claims 7 to 12 is recorded.

Images