JP2000067164A - Method and device for pattern recognition and record medium where template generating program is recorded - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide the pattern recognizing method which is robust to deterioration and deformation by making use of the presence of restrictions on the contiguity relation between characters, etc., when a recognition object is limited. SOLUTION: This device has a learning image storage means 1 which stores an image for learning, a learning pattern extracting means 2 which extracts a learning pattern included in the image for learning including circumferential image information, a template generating means 4 which generates a template by using the learning pattern, a template storage means 5 which stores the template, a pattern storage means 6 which stores an input image pattern, and a recognizing means 7 which recognizes the input image pattern by using the template.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a technology for recognizing image patterns centered on characters and figures. In addition, the present invention relates to a recognition method and an apparatus suitable for a case where the degree of deterioration of these patterns greatly changes, and a recording medium on which a template creation program is recorded.

[0002]

2. Description of the Related Art Here, as an example of pattern recognition,
A description will be given for character recognition. Conventionally, Kanji OCR (Op
character recognition devices such as tical
There is known a method of performing recognition on an input document image pattern by performing pattern matching between the input pattern and a template. Such techniques are commonly used in design-processed characters with unusual textures, such as those found in advertisements and newspaper headlines, as well as blurring, crushing,
It has been reported that the image is relatively robust to an image in which the character background noise is severe.

However, when the character shape fluctuates due to image observation conditions such as lighting conditions, or when there is other information such as other characters in the surroundings, the similarity between the input pattern and the template decreases, and the recognition pattern decreases. There was a problem that the rate decreased.

In order to solve such a problem, conventionally, (1) other information in an input pattern is regarded as noise and removed, or a variation in character shape is corrected, and then recognition is performed using an individual character template. , And (2) a method of absorbing the influence of individual character deformation by using a word template.

[0005] However, in the case of the method using the individual character template, it is difficult to remove only other information in the input pattern. This is because it often happens that character strokes and other information are very similar.
Therefore, other information often remains or information of interest is lost. As a result, the similarity at the time of matching with the template decreases, and the recognition rate decreases. In particular,
Small patterns such as "l" and "r" are often affected by other information around and difficult to recognize.

On the other hand, in the method using a word template, a plurality of characters are collated with each other. Therefore, even when a character such as a small character has a fluctuating character shape and is apt to be erroneously recognized, the entire word can be recognized. Can be expected. However, it is necessary to create a template according to the number of words, and as the number of words increases, the number of combinations increases. In addition, the matching time increases accordingly. Further, if an unknown word is included in the input pattern, it cannot be recognized because there is no corresponding template.

[0007]

As described above, in the conventional pattern recognition technology such as OCR, a technique for correctly recognizing a pattern having large deterioration or deformation has not been sufficiently established.

An object of the present invention is to solve the above-mentioned problem of pattern recognition based on template matching by using the fact that there are restrictions on the adjacency of characters and the like when recognition targets are limited. It is an object of the present invention to provide a pattern recognition method and apparatus which is robust against deterioration and deformation by creating a template in consideration of the above image information. Here, examples of the surrounding image information include character patterns before and after the character pattern of interest, and decorations provided around the underline and the like.

Further, a recording medium storing a computer-readable program for creating a dictionary capable of obtaining a high recognition rate without manual work by automatically extracting a learning pattern from a learning image and learning a template. The purpose is to provide.

[0010]

In order to achieve the above object, the present invention is directed to an image processing apparatus comprising the steps of: providing a similarity degree between an input image pattern and a template of an image category which has been created in advance using a learning image; In a pattern recognition method for calculating a distance value and recognizing a category of the input image pattern, a learning pattern included in a learning image is extracted including surrounding image information, and a template is created using the extracted learning pattern. And recognizing an input image pattern using the template.

According to a second aspect of the present invention, in the pattern recognition method according to the first aspect, a learning pattern is extracted from a learning image by using position information manually designated.

According to a third aspect of the present invention, in the pattern recognition method according to the first aspect, a learning image is recognized by an initial dictionary prepared in advance, and a learning pattern is extracted from the learning image using the result. It is characterized by.

The invention according to claim 4 is the invention according to claims 1 to 3.
The described pattern recognition method is characterized in that the number of templates is reduced by clustering a plurality of created templates.

The invention according to claim 5 is the invention according to claims 1 to 4.
The described pattern recognition method is characterized in that a learning pattern that does not include surrounding image information is extracted from a learning image, and a template is also created from the learning pattern.

According to a sixth aspect of the present invention, a similarity or a distance value between an input image pattern and an image category template created in advance using a learning image is calculated, and the category of the input image pattern is recognized. A pattern storage device for storing an input image pattern, a learning image storage device for storing a learning image, and learning for extracting a learning pattern included in the learning image including surrounding image information. Pattern extraction means, template creation means for creating a template using the learning pattern, template storage means for storing the template, and recognition means for recognizing the input image pattern using the template. Features.

According to a seventh aspect of the present invention, there is provided the pattern recognition apparatus according to the sixth aspect, further comprising a single learning pattern extracting means for extracting a learning pattern not including surrounding image information from the learning image. .

The invention according to claim 8 calculates a similarity or a distance value between an input image pattern and a template of an image category created using a learning image in advance, and recognizes the category of the input image pattern. There is a computer-readable recording medium recording a template creation program for performing a processing process of extracting a learning pattern included in a learning image including surrounding image information, and a template using the extracted learning pattern. And a creating process.

[0018]

Embodiments of the present invention will be described below in detail with reference to the drawings. In the following, a character pattern composed of binary values of white or black will be described as a typical example of an image pattern, but other image patterns can be similarly realized.

FIG. 1 is a block diagram showing a configuration example of a pattern recognition apparatus according to an embodiment of the present invention. 1 is a learning image storage circuit, 2 is a learning pattern extraction circuit, 3 is a single learning pattern extraction circuit, Is a template creation circuit, 5 is a template storage circuit, 6 is a pattern storage circuit, and 7 is a recognition circuit. The pattern recognition apparatus is actually realized using so-called computer hardware and software resources.

The learning image storage circuit 1 stores learning images used for creating a template. As the learning image, an image is used in which a learning pattern can be obtained for all the sets of categories around the input image pattern.
The learning image is composed of n pixels. For example, pixel values are represented by “0” for white pixels and “1” for black pixels.

The learning pattern extraction circuit 2 inputs a learning image from the learning image storage circuit 1 and extracts a learning pattern included in the learning image including surrounding image information.
The method of extracting the learning pattern will be described later. The extracted learning pattern is passed to the template creation circuit 4.

The single learning pattern extraction circuit 3 receives a learning image from the learning image storage circuit 1 and extracts a learning pattern consisting of only the pattern of interest. The extracted learning pattern is passed to the template creation circuit 4. The single learning pattern extraction circuit 3 can be omitted when no single learning pattern is used.

The template creation circuit 4 receives a learning pattern from the learning pattern extraction circuit 2 or the single learning pattern extraction circuit 3 and creates a template. The template created by the template creation circuit 4 is newly registered in the template storage circuit 5, or added / updated.

The template storage circuit 5 stores templates of each category used for recognition determination of an input image pattern.
One or more (generally a plurality) are stored for each category.

The pattern storage circuit 6 stores one or more recognized image patterns as input patterns. These image patterns are composed of M pixels. For example, in the case of a binary image pattern, each pixel is indicated by “0” or “1”.

The recognition circuit 7 receives a template from the template storage circuit 5, inputs an input pattern to be recognized from the pattern storage circuit 5, and performs pattern recognition by comparing the template with the input pattern (shifted matching). And outputs the recognition result.

A feature of the present invention resides in that a learning pattern included in a learning image is extracted including surrounding information, and a template for recognition is created using the learning pattern. In the following, a detailed description will be given of the extraction of a learning pattern and the creation of a template.

First, the process of automatically creating a template from a learning image by the learning pattern extraction circuit 2 and the template creation circuit 4 will be described with reference to the flowchart of FIG. This processing procedure can be provided as a template creation program recorded on a computer-readable recording medium such as a floppy disk, a memory card, or a compact disk (CD-ROM).

A learning image is read from the learning image storage circuit 1 and input to the learning pattern extraction circuit 2 (step 201). It is assumed that the learning image includes one or a plurality of learning patterns. Fig. 3 shows an example of a learning image
The image 301 is shown in FIG.

The learning pattern extraction circuit 2 first reads an initial dictionary from the template storage circuit 5 (step 2).
02). It is assumed that the initial dictionary is created by, for example, manually extracting one template for each category from the learning image. FIG. 4 shows an example of a template in the initial dictionary of the category “e”.

Next, the learning pattern extraction circuit 2 recognizes the learning image using, for example, an initial dictionary (step 2).
03). As a recognition method, for example, a pre-processed image is created by performing pre-processing such as tilt correction on an input image, and a method based on shifted matching using complementary similarity is used for the pre-processed image. Yes (for example, Sawaki et al., “Character Recognition in Bookshelf Images by Automatic Dictionary Selection”, IEICE General Conference 1998 (1998
D-13-33). Specifically, the observation window for recognition is scanned in the vertical direction on the input image after the inclination correction, and each observation window for recognition is compared with the reference pattern in the initial dictionary. Scanning is performed not only in the vertical direction but also in the horizontal direction in consideration of misalignment by using the complementary similarity in this comparison. When the maximum similarity at each collated position exceeds the similarity threshold of each character category that has been learned in advance, the category is used as a recognition result by the initial dictionary. With this recognition method, the image position of the portion showing the high similarity and the category of the template can be obtained. In the image 302 of FIG.
A preprocessed image obtained by performing a tilt correction on the learning image 301 is shown. Further, at 303 in FIG. 3, a category having a high similarity when the line above 302 in FIG. 3 is shifted and compared by the initial dictionary is shown at a position having a high similarity.

Next, the learning pattern extraction circuit 2 extracts a learning pattern by using the position and the category showing the high similarity (snap 204). As an example of a method of extracting a learning pattern, a rectangular area is extracted around an input position. An example of a preprocessed image is shown in an image 501 of FIG. An example of a learning pattern of the category “e” extracted from the image 501 is shown in an image 502 of FIG.

The extracted learning pattern is input to the template creating circuit 8. The template creation circuit 4 creates a template using the input learning pattern (step 205). As a template creation method, for example, when a learning pattern is extracted from a character string, learning patterns having the same set of three categories consisting of the category of the learning pattern and the categories of characters before and after the learning pattern are collected, and averaged at each pixel position. A method of taking a value, a method by clustering, or the like can be used. FIG. 6 shows an example of the created template.

The created template is input to the template storage circuit 5 (step 206). In the template storage circuit 4, the input template is newly stored / added / updated.

Next, as another embodiment, an embodiment in which a learning pattern is manually extracted from a learning image will be described.

In this embodiment, instead of steps 202 and 203 in FIG. 2, the read learning image is displayed on the display, and the position and category of the learning pattern in the learning image are manually input. After that, it is the same as the above embodiment. FIG. 6 also shows an example of a template set created by the method according to this embodiment.

Table 1 shows the results when 3468 characters in the bookshelf image were recognized using the template of FIG. In Table 1, for the sake of comparison, the recognition rate when recognition is performed using a template in which surrounding information has been deleted from the template in FIG. 6 is also shown.

[Table 1]

Next, an example will be described in which a learning pattern that does not include surrounding image information is extracted from a learning image and used as a template. That is, when learning patterns cannot be obtained in advance for all of the set of categories around the input image pattern, it is necessary to prepare a template including a learning pattern that does not include surrounding image information.

In FIG. 1, a learning image is read from the learning image storage circuit 1 and is input to a single learning pattern extraction circuit 3. In the independent learning pattern extraction circuit 3, for example, a learning pattern that does not include surrounding information is manually cut out from the learning image. In this case, a noise component may be removed by filtering or the like, and a learning pattern not including surrounding information may be extracted. The learning pattern is sent to the template creation circuit 4. The template creation circuit 4 creates a template from the input learning pattern and sends it to the template storage circuit 54. In the template storage circuit 5,
Adds / updates the template composed of the input individual patterns to a template including surrounding information.

In any of the embodiments, after the template is created, a process for reducing the number of templates can be performed by using a conventionally known clustering method, and the newly created template can be stored.

In the above description, characters are mainly shown, but the same processing can be applied to image patterns other than characters.

[0042]

As described above, according to the present invention, the following effects can be obtained. (1) By creating a template including surrounding information and performing recognition using this template, even when individual characters are greatly deformed, the advantage that the information such as surrounding characters can be used for correct recognition can be obtained. is there. further,
Since there is no need to create word templates as in the conventional method, the number of templates can be reduced,
Also, it is possible to flexibly cope with a case where an unknown word is input.

(2) By manually inputting the position of the learning pattern in the learning image, it is possible to extract the learning pattern while checking the quality of the deterioration of the learning pattern visually, and there is an advantage that there is no extraction error.

(3) By recognizing a learning image with an initial dictionary prepared in advance, automatically extracting a learning pattern from the learning image using the result, and creating, adding, and updating a template. Thus, there is an advantage that a dictionary suitable for the learning image is automatically formed, and the recognition rate is improved.

(4) By applying clustering to the created dictionary, redundant templates can be reduced.

(5) By providing the independent learning pattern extracting means, the input pattern can be correctly recognized even if the user does not have a template having the same combination as the combination of the character category of the input pattern and the surrounding categories.

[Brief description of the drawings]

FIG. 1 is a block diagram illustrating a configuration example of a pattern recognition device according to the present invention.

FIG. 2 is a flowchart illustrating a template creation method according to an embodiment of the present invention.

FIG. 3 is a diagram illustrating an example of a learning image, a preprocessed image, and a recognition result.

FIG. 4 is a diagram showing an example of a template in an initial dictionary.

FIG. 5 is a diagram for explaining a method of extracting a learning pattern from a preprocessed image.

FIG. 6 is a diagram illustrating an example of a created template.

[Explanation of symbols]

 REFERENCE SIGNS LIST 1 image storage circuit for learning 2 learning pattern extraction circuit 3 independent learning pattern extraction circuit 4 template creation circuit 5 template storage circuit 6 pattern storage circuit 7 recognition circuit

Continued on the front page (72) Inventor Norihiro Hagita 3-19-2 Nishi-Shinjuku, Shinjuku-ku, Tokyo F-term (reference) in Nippon Telegraph and Telephone Corporation 5B064 AB02 AB03 AB13 AB16 AB17 DA27 DB06 5L096 FA19 HA08 JA03 JA09 JA22 KA04 LA17 MA07

Claims (8)

[Claims] 1. A pattern recognition method for calculating a similarity or a distance value between an input image pattern and a template of an image category created using a learning image in advance, and recognizing the category of the input image pattern. Extracting a learning pattern included in a learning image including surrounding image information, creating a template using the extracted learning pattern, and recognizing an input image pattern using the template. Recognition method. 2. The pattern recognition method according to claim 1, wherein a learning pattern is extracted from a learning image using manually designated position information. 3. A pattern recognition method according to claim 1, wherein the learning image is recognized by an initial dictionary prepared in advance, and a learning pattern is extracted from the learning image by using the recognition result. Recognition method. 4. The pattern recognition method according to claim 1, wherein the number of templates is reduced by clustering a plurality of created templates. 5. The pattern recognition method according to claim 1, wherein a learning pattern that does not include surrounding image information is extracted from the learning image, and a template is created from the learning pattern. Pattern recognition method. 6. A pattern recognition apparatus that calculates a similarity or a distance value between an input image pattern and a template of an image category created using a learning image in advance, and recognizes the category of the input image pattern. A pattern storage unit that stores an input image pattern, a learning image storage unit that stores a learning image, and a learning pattern extraction unit that extracts a learning pattern included in the learning image including surrounding image information, Pattern recognition, comprising: template creation means for creating a template using the learning pattern; template storage means for storing the template; and recognition means for recognizing the input image pattern using the template. apparatus. 7. The pattern recognition apparatus according to claim 6, further comprising a single learning pattern extraction unit for extracting a learning pattern not including surrounding image information from the learning image. 8. A template for calculating a similarity or a distance value between an input image pattern and a template of an image category created using a learning image in advance, and recognizing the category of the input image pattern. A computer-readable recording medium storing a program, a processing process of extracting a learning pattern included in a learning image including surrounding image information, and a processing process of creating a template using the extracted learning pattern. A recording medium on which a template creation program is recorded.