KR100573392B1 - Method and System for digitalizing a large volume of documents based on character recognition with adaptive training module to real data - Google Patents

Abstract

The present invention relates to an efficient configuration of character recognition in a document digitization process, and more particularly, to automatic generation of a representative pattern model mounted on a character recognition engine to enable adaptive learning of actual data in a large and diverse document digitization process.

The present invention as described above comprises the steps of extracting the frequency of appearance of the character pattern included in the digitization target document from the document data containing the structure and text information of the document image; Dividing the individual image for each character pattern using the document structure and the segmentation information; Extracting statistical features of each character pattern image using the individual image segmentation information; And generating a representative model for each character pattern using the statistical feature and comparing the input character pattern with an input character pattern. Therefore, according to the present invention, the character recognition engine of the document digitization system may be equipped with a new representative pattern model for actual data to maximize its performance.

Document Digitization, Character Recognition, Representative Patterns, Feature Extraction, Adaptive Learning

Description

Method and System for digitalizing a large volume of documents based on character recognition with adaptive training module to real data}             

1 is a view illustrating a classification according to a character recognition object,

2 is a configuration diagram of a document digitization system using character recognition;

3 is a view showing a character recognition process,

4 is a view showing the configuration of the adaptive learning module of the character recognition engine according to an embodiment of the present invention;

5 is a diagram showing a distribution example of the frequency of appearance of a character pattern appearing in a document;

6 is a view showing character pattern information of a document image structure and segmentation information DB;

7A is a view showing a nonlinear form normalization method of a text area;

FIG. 7B is a diagram illustrating a direction feature of contour pixels extracted from a normalized image; FIG.

8A illustrates an exemplary model representation of a character pattern (printed character) represented by outline pixels;

8B is an example of representative model reproduction (written characters) of a character pattern represented by outline pixels;

9A is a diagram showing an example of an interface (step 1) of an adaptive learning module;

9b is a diagram showing an example of an interface (step 2) of an adaptive learning module;

9C is a diagram showing an example of an interface (step 3) of an adaptive learning module;

9D is a diagram showing an example of an interface (step 4) of an adaptive learning module;

10 is a diagram showing a hierarchical input / output relationship of four steps of an adaptive learning process;

11A is a view showing an example of a letter pattern appearance frequency survey generated in step 1;

11B is a view showing 100 'ga' patterns (step 2) divided from a document image;

12 is a view illustrating a change in the recognition performance and the number of character models of the character recognition engine.

The present invention relates to a character recognition technology that has been utilized in response to the automation requirements for input in the process of digitizing a large document, and more particularly, to an efficient document digitization method and system using an adaptive learning module. Character recognition is largely divided into on-line and off-line character recognition, and the digitalization process of the document corresponds to the latter. The classification according to the character recognition object is well illustrated in FIG. 1.

In general, handwritten characters include more complex variations than printed characters, and in order to recognize handwritten characters, it is most important how to absorb complex variations. In the case of constrained cursive writing, cross-shaped guide lines may be shaped to minimize deformation. In the case of printed matters, it is common to use mono-font OCR rather than omni-font OCR engine to minimize typographical variations. This means that the deviation of the performance of the recognizer according to the actual data is large.

On the other hand, the character recognition consists of a process of extracting features from input data and classifying them into one of a plurality of predetermined models. Models that are built in advance during the training process are the same as the criteria for mapping them into a concept through recognition, which limits the recognition data or affects the performance of character recognition.

There are qualitative and quantitative aspects of computer difficulty in recognizing characters. The excellent character recognition function of human beings is based on intuition and experience, and thus quality difficulties occur because the general method of objectifying, quantifying, and formulating the recognition process on the computer is not completed. For example, in the case of printed characters, difficulties arise due to inconsistencies in various font shapes, sizes, slants, and the like.

The quantitative aspect of character recognition occurs when it is impossible to implement practically in engineering large memory capacity or recognition time, and is particularly severe when characterizing a large amount of information content. For example, as the recognition object is changed from numbers (10 types) to English letters (52 types), Korean characters (11,172 types), and furthermore to Chinese characters (about 50,000 types), this increases the number of characters. This is because, as the type increases, the shape of the character to be stored increases, leading to an increase in the size of the recognition device for comparing and determining various possibilities.

The qualitative and quantitative difficulty of character recognition as described above is a factor that makes it difficult to implement general-purpose character recognition having a certain performance on all data. Therefore, in order to solve the problem of efficiently digitizing a large amount of documents, it is inevitable to mount an additional module that can be applied to the existing document digitization system in a form in which the difficulty of character recognition is minimized.

Accordingly, an object of the present invention to solve the problems of the prior art is to provide a method and system for digitizing a document that is most suitable for real data and maintains optimum performance without considering the qualitative and quantitative aspects of real data.

Another object of the present invention is to provide a method and system for adaptively driving a character recognition engine so that the digitization system of a document works best with actual document data.

It is still another object of the present invention to provide a method and system for adaptively configuring a digitization system of a document so that an operator (operator) can generate a real pattern model using the digitized document data at any time. have.

The present invention for achieving the above object of the present invention is to mount the offline adaptive learning module to the character recognition engine constituting the digitization system of the document. The offline adaptive learning module proposed in the present invention creates a new model for the patterns by examining the distribution of patterns appearing in the actual document data. Therefore, the character recognition engine does not have to consider the qualitative and quantitative aspects of the actual data because the adaptive learning module can generate the actual pattern model by using the digitized document at any time by the operator. It is possible to maintain proper and consistent optimum performance.

The present invention for adaptively driving a digitization system of a document utilizes the structure, split information, and text information of a document generated in the process of digitizing a document. In detail, the operation of the adaptive learning module may include extracting a frequency of occurrence of a character pattern included in a digitization target document from document data including a structure of a document image and text information; Dividing the individual image for each character pattern using the document structure and the segmentation information; Extracting statistical features of each character pattern image using the individual image segmentation information; And generating a representative model for each character pattern using the statistical feature and comparing the input character pattern with an input character pattern.

DETAILED DESCRIPTION A detailed description of preferred embodiments of the present invention will now be described with reference to the accompanying drawings. It should be noted that reference numerals and like elements among the drawings are denoted by the same reference numerals and symbols as much as possible even though they are shown in different drawings. In the following description of the present invention, if it is determined that a detailed description of a related known function or configuration may unnecessarily obscure the subject matter of the present invention, the detailed description thereof will be omitted. In the following description, the adaptive learning or learning process refers to a representative model of a character pattern by using an adaptive learning module according to an embodiment of the present invention by using the document information and the segmentation information DB of the document image generated during the digitization of the document. It means the process of generating.

Document digitization system to which the present invention is applied is as shown in FIG. Referring to FIG. 2, the digitization of the document goes through an automation step and a verification and correction step by a verification interface to extract text information from an input document image. Here, the automation step is composed of text area extraction and character recognition process by structural analysis and segmentation of document image. In FIG. 2, the technique for clustering and character verification by character recognition is described in Korean Patent Application No. 2002-0067286 (high document input and correction system through Chinese character image clustering) and Korean Patent Application 2003-0088230 filed by the present inventor. (Digital verification system and method for character clustering technology based documents).

The character recognition process of the large-capacity document digitization system is shown in FIG. 3. The recognition result is generated through preprocessing, feature extraction, classification, and postprocessing on the input pattern. In the classification process, the most likely text code is determined by comparing the input pattern with a previously learned character model. The adaptive learning module of the present invention automatically generates a character model suitable for the input pattern. The information used in the learning process is the fragment information DB and text DB of the document image generated during the digitization process.

Principle of the Invention

4 is a view showing the configuration of the adaptive learning module of the character recognition engine according to an embodiment of the present invention. Referring to FIG. 4, the document structure / division information DB 110 stores the character structure / division information data generated by the document digitization performing unit 100 during the digitization process of the document. In addition, the document text information DB 130 stores the document text information in which the automation step and the verification and correction step by the verification interface are completed. The character recognition module 120 determines the most likely text code in the classification process through the character pattern representative model that is most suitable for the actual document data provided from the adaptive learning module unit 200, and uses it for document digitization. .

The operator in charge of the digitization of a large amount of documents from time to time, if the performance of the character recognition module 120 is not optimal, using the adaptive learning module 200 to maintain the constant of the character recognition to suit the actual document data There is a number. The adaptive learning module 200 is often driven as in the present invention has two advantages that can solve the quantitative difficulty of character recognition. First, since the quality difficulty arises from the generality problem that the character recognition engine does not absorb the different elements between the data at the learning stage and the actual data well, the character recognition engine learns by using the actual data. This means that the shape, size, tilt, and quality of the image are easily absorbed. In addition, quantitative difficulty is particularly severe when there is a large amount of information to be recognized. Learning only the patterns that appear in the actual document data is a possibility of character shapes that need to be remembered by the character recognition engine and various possibilities of comparison and determination. This aspect can be solved by reducing the number.

The adaptive learning module unit 200 is a character pattern distribution research unit 210 for examining the appearance pattern distribution of each of the various character patterns included in the digitization target document from data stored in the document text information DB 130, the document structure / Character pattern divider 230 for dividing the character pattern information from the data stored in the partition information DB 110, Statistical feature information extraction unit 250 for extracting statistical feature information from each of the divided character pattern information, The statistical feature It comprises a representative model generation unit for generating a representative model of the character pattern using the average or the weighted average of the feature values extracted by the information extraction unit 250.

Hereinafter, the adaptive learning process through the adaptive training module unit 200 shown in FIG. 4 will be described in detail.

A. Examining the Distribution of Character Patterns

The operator drives the adaptive learning module unit 200 so that the character recognition module 120 exhibits optimal performance. When the adaptive learning module unit 200 is driven, the character pattern distribution research unit 210 examines the appearance pattern distribution of each of the various character patterns included in the digitization target document from data stored in the document text information DB 130. The document contains various character patterns and the character patterns have their respective frequency of appearance, so that some character patterns are frequently used, while others are rarely used. Therefore, if the character recognition engine does not have or do not have a comparative model for a pattern that is rarely used, it does not significantly affect the overall performance. Inflicted. This means that the character recognition engine should be sensitive to the frequency of occurrence of the character pattern during the digitization of large documents.

The information on the appearance pattern distribution of each of the character patterns irradiated by the character pattern distribution research unit 210 is stored in the pattern appearance frequency survey 220.

B. Segmentation of appearance character pattern areas

The division of the appearance character pattern region means that the character image for each pattern is extracted in the following manner.

When the examination of the actual appearance pattern distribution is completed from the document area, the character pattern division unit 230 displays a character image of each pattern from data stored in the document structure / division information DB 110 and the pattern appearance frequency survey 220. Extract. In this process, as well as the distribution information of the emerged character pattern from the document, the character pattern to form the representative model is determined, and the character image for each pattern is extracted by using the structure information and the segmentation information DB of the document image. The extracted character image information for each pattern is stored in the character pattern image DB 240.

6 shows an example of character pattern information to be divided by the structure of the document image and the segmentation information DB. Here, examples of the character pattern information may include page numbers (Page Idx), paragraph numbers (Para and Idx) to which the character belongs, lines other than the coordinates of the character area and code values of the character pattern in the document. (Line Idx), character order (Char Idx), and so on.

C. Statistical Characteristic Information Extraction of Character Patterns

When the division of the actual appearance pattern is completed from the document image, the statistical feature information dividing unit 250 extracts statistical feature information from each of the divided character pattern images using data stored in the character pattern image DB 240. . This step is performed in the same manner as the preprocessing and the feature extraction of the character recognition process. First, in the preprocessing step, the position, size, and shape information of the image are normalized, and in the feature extraction step, statistical information extracted from the normalized image is digitized. Accordingly, the representative model generator 270 may be implemented as a means for normalizing the position, size, and shape information of the image, and a means for digitizing statistical information extracted from the normalized image.

A method of obtaining a normalized image uses a nonlinear shape normalization technique such as dot density, crossing count, line interval, and the like. After normalizing to 64 * 64 size, 64 (8 * 8) block is formed using 8 * 8 pixel area as one block. Statistical features for absorbing variations in various character patterns include stroke count, contour-pixel count, contour direction, backgound area, and projection. 7B illustrates an example of extracting a contour directional feature (CDF). If the direction of the contour pixel of the character pattern is one of four directions (horizontal, vertical, diagonal, inverse diagonal) as shown in FIG. 8B, the contour direction feature that can be extracted from any character pattern is 256 (8 * 8). 4) It becomes a vector of dimensions.

Statistical feature information extracted from the segmented character pattern image is stored in the character pattern feature information storage unit 260.

D. Create Representative Model of Character Pattern

When statistical feature information extraction is completed from the divided character pattern, the representative model generator 270 generates a representative model of the character pattern from the feature information of the pattern stored in the character pattern feature information storage unit 260. The representative model of the character pattern can be easily obtained by taking the average or the weighted average of feature values extracted from the learning character images. If the contour direction feature is used as the statistical feature of the character pattern, a representative model of the character pattern can be reproduced as a gray-scale image of 64 * 64 size as shown in FIGS. 8A and 8B. This is composed of 256 gray images of the distribution of contour pixels extracted from 100 character images for each pattern. The white pixels have no contour pixels of the pattern located at that point, and the black pixels have the contour pixels of the pattern. It always exists at that location. 8A shows a representative model of a printed character pattern having a relatively small shape variation, the representative model also shows a standardized result and the presence of outlines is very clearly shown. On the other hand, since FIG. 8B is a text image extracted from a manuscript with a very free handwritten style, the representative model does not have a distinct shape, and the outline shape is faint.

<Examples>

9A to 9D illustrate an interface configuration of an adaptive learning module according to an embodiment of the present invention. 9A to 9D illustrate a process of generating a representative model of a character pattern suitable for actual document data.

"AWD_Inspection" of FIG. 9A is a step of examining the distribution of appearance patterns of a document, and stores a directory in which the text information of the document is located, that is, a region specifying the document text information DB 130 and a distribution of the character pattern of the document. An area for specifying a file, that is, the pattern appearance frequency survey 220, is provided.

"Random Selection" in FIG. 9B is a step of dividing an area of a character pattern from a document image. The "Random Selection" of FIG. An area for designating a structure / division information DB 110, an area for designating an image storage directory for the character pattern to be divided, which is the character pattern image DB 240, and an area for displaying the maximum number of character pattern images to be divided are provided. .

"FeatureEx" of FIG. 9C is a step of extracting a feature of a character pattern. The character pattern feature information storage unit 260 is an area for specifying a statistical feature storage directory for a character pattern and an area for specifying a divided image storage directory. It provides a range of choices for both English and statistical feature extraction methods to choose from and nonlinear shape normalization techniques.

In FIG. 9D, "representative" is a step of generating a representative model of a character pattern, and an area for designating a statistical feature storage directory of the character pattern and a representative model storage directory of each character pattern which is the character pattern representative model DB 280. A designated area is provided.

The hierarchical input / output relationship of the above four steps of generating a representative model of the actual character pattern from the digitized document information is shown in FIG. 10. FIG. 11A shows an example of “CharFrequency.txt” generated from about 740,000 characters of document in the “AWD_Inspection” step of FIG. 9A, and FIG. 11B is randomly divided in the “Random Selection” step of FIG. 9B. 100 "ga" pattern text images are shown.

Meanwhile, in the detailed description of the present invention, specific embodiments have been described, but various modifications may be made without departing from the scope of the present invention. Therefore, the scope of the present invention should not be limited to the described embodiments, but should be determined not only by the scope of the following claims, but also by the equivalents of the claims.

As described above, the present invention maintains optimal performance by minimizing quantitative and qualitative problems by providing an adaptive learning module that can be directly adapted to actual data in a character recognition engine in an environment for efficient digitalization of large documents. There is an advantage.

In order to show the effect of the present invention through specific data, the recognition performance and the number of models of the character recognition engine equipped with the character recognition engine and the periodic adaptive learning module learned by general training data were measured. The experimental data for quantitative evaluation is a 7 million-character booklet published in 1974, and the quality of document images is very low because it is printed by lead typeface. The experimental data is a document image scanned at 400dpi of about 700,000 characters from the above booklet. Before the character recognition, noise and skew were removed by using basic image preprocessing techniques.                     

Fig. 12 shows the experimental results on the recognition performance and the number of character models. The character recognition engine (a) includes 2,350 class character models that can represent 99.999% of Korean documents. 100 character images were generated. The character recognition engine (b) is an adaptive learning step for about 5% (about 35,000 characters) of the experimental data, and (c), (d), and (e) are 20%, 40%, and 50, respectively. This is the result of using% in the adaptive learning stage.

The character recognition engine (a) includes a character model for recognizing almost all data. However, the character recognition engine (a) has a low recognition rate due to severe image quality and shape variation of the actual document data. On the other hand, from the character recognition engines (b) to (e), since the adaptive learning was performed using the actual data, the recognition performance and the number of the character models were maintained at the optimal state. In particular, the engine (b) includes only 894 character models because the training was performed on only about 5% of actual data, but the recognition rate is much higher than that of the engine (a). can do.

In the environment of digitizing large documents through the use of the present invention, even a small character recognition engine combined with an adaptive learning step utilizing a small number of actual data may be more efficient to maintain constant performance than a very large character recognition engine. Expect to be.

Claims (6)

In the digitizing method of the document comprising the step of comparing the input character pattern with a previously stored character model, Extracting a frequency of appearance of a character pattern included in a digitization target document from document data containing document structure and text information; Dividing the individual image for each character pattern using the document structure and the segmentation information; Extracting statistical features of each character pattern image using the individual image segmentation information; Providing a representative model for each character pattern from the digitization target document from time to time by using the statistical feature and comparing the input character pattern with the input character pattern; Character recognition-based digitization method of a large document comprising a. The method of claim 1, wherein the information for dividing the individual image for each character pattern comprises: A character including at least one of a coordinate value of a character area of the digitization target document and a code value of a character pattern, a page number of the digitization target document, a paragraph number to which the character belongs, a line to which the character belongs, and a sequence number of the character Recognition-based digitization of large documents. The method of claim 1, wherein extracting the statistical feature comprises: Normalizing the position, size, and shape information of the image; Digitizing statistical information extracted from the normalized information Character recognition based on the digitization method of a large document comprising a. The method of claim 1 or 3, wherein the statistical characteristic information, A method of digitizing large text documents based on text recognition, comprising at least one of stroke count, outline pixel, outline direction, background area, and projection information. The representative model of any one of claims 1 to 3, wherein Character recognition-based large document digitization method comprising at least one or more of the average or weighted average of the extracted feature value. In the digitization system of a document comprising a character recognition module for comparing the input character pattern with a previously stored character model, A character pattern distribution research unit for examining the frequency of occurrence of the character pattern included in the digitized document from data stored in the document text information DB; A character pattern divider for dividing an individual image for each character pattern using document structure and segmentation information; A statistical feature information extraction unit for extracting statistical features of each character pattern image using the individual image segmentation information; Representative model generation unit for generating a representative model for each character pattern from time to time using the statistical feature to compare with the input character pattern Digitization system of a document comprising a.

Images