JP2002318998A - Pattern input method, program, and storage medium with the program stored therein - Google Patents

Abstract

PROBLEM TO BE SOLVED: To solve the problem that a pattern not defined as dictionary pattern cannot be retrieved when pattern-recognizing a pattern in an input image and converting it to a character code. SOLUTION: This method comprises a processing 20 for cutting an undefined pattern having no character code given thereto in the input image 100 as the dictionary pattern 300; a processing 400 for measuring the characteristic quantity of the dictionary pattern and input image; a processing 500 for retrieving and extracting a pattern 600 similar to the dictionary pattern on the basis of the characteristic quantity; and a processing 700 for giving the same character code to all the extracted similar patterns. These pieces of processing are repeated until the non-defied pattern is eliminated in the image. This method also comprises cutting the patterns appearing in the input image in the order for appearance frequency when cutting the undefined pattern.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a pattern input method for converting various patterns (characters, figures, etc.) existing in an input image into character codes and inputting them.

[0002]

2. Description of the Related Art Conventionally, methods for converting various patterns existing in an input image into character codes and inputting them are divided into an input method by hand and an input method based on pattern recognition using a computer.

In the former method, a person sees and recognizes all patterns in an image, and sequentially inputs a character code given to each pattern by an input device such as a keyboard.

[0004] The latter method is a method generally called pattern recognition (Yoshiaki Shirai, "Pattern Understanding", Ohmsha, 1987). In this method, a pattern that will occur in an image is created in advance as a dictionary pattern, and various patterns existing in the image are classified based on features such as density information and outline information of the created dictionary pattern. How to

[0005]

The former method described in the above "prior art" requires enormous costs in both cost and time. As a method for solving this problem, the latter method using a computer has been studied, but this method has the following problems.

In the latter method, since a search is performed based on a dictionary pattern created in advance, a pattern that is not expected to appear cannot be recognized. Therefore, dictionary patterns for all patterns that will appear must be created first. However, if the variation of the appearance pattern is unknown, it is impossible to create all dictionary patterns in advance.

When the number of pattern variations is very large, it takes a lot of cost to create a dictionary pattern.

As described above, when the variation of the appearance pattern is unknown or very large, it cannot be dealt with by the framework of the conventional method in which a dictionary pattern is created in advance.

An object of the present invention is to provide a pattern input method, a program thereof, and a recording medium on which the program is recorded, which enables efficient pattern input even when the appearance pattern variation is unknown or very large. is there.

[0010]

According to the present invention, in order to solve the above-mentioned problems, an entire image is searched by using an undefined pattern cut out from an image as a dictionary pattern, and a similar pattern is extracted. By repeating the extraction of the definition pattern and the similar pattern search, the undefined pattern is reduced in stages, and the extraction of the maximum number of pattern variations is performed, enabling input of all the patterns. The invention is characterized by a method, a program, and a recording medium on which the program is recorded.

(1) In a method of converting various patterns existing in an input image into a character code and inputting the same, a first step of cutting out an undefined pattern without a character code in the input image as a dictionary pattern A second step of measuring a dictionary pattern obtained by the cutout and a feature amount of the input image; and searching the input image for a pattern similar to the dictionary pattern based on the feature amount obtained by the feature amount measurement. And a fourth step of assigning the same character code to all of the similar patterns extracted in the above step, and the first step until the undefined pattern in the image disappears. A pattern input method characterized by repeating the steps up to a fourth step.

(2) In the pattern input method according to the above (1), in the step of cutting out the undefined pattern from the input image, the pattern is cut out in order from the pattern which appears frequently in the input image. The pattern input method to be performed.

(3) In the pattern input method according to the above (1) or (2), in the step of assigning a character code to the similar pattern extracted by the search, it is determined whether or not the dictionary pattern and the similar pattern are the same. A pattern input method characterized in that a determination is obtained as an operation input.

(4) In the pattern input method according to any one of the above (1) to (3), when the input image is a black-and-white binary image, the contour feature of the pattern is used as a feature amount for search. A pattern input method characterized by using:

(5) In the pattern input method according to any one of the above (1) to (4), in the process of extracting a dictionary pattern and a similar pattern from an image, it is determined whether or not the pattern is similar. A pattern input method characterized in that a judgment scale is a distance value between feature values.

(6) In the pattern input method according to any one of the above (1) to (5), when a weighted average histogram, which is one of the contour features, is measured from both the dictionary pattern and the input image. A small area centering on a black pixel of interest, obtaining a contour direction histogram in the small area, and performing blurring by a two-dimensional Gaussian filter.

(7) In the pattern input method described in any one of (1) to (6) above, when assigning a character code to a similar pattern extracted by a search, the identity is manually confirmed. A dictionary input pattern and a plurality of similar patterns are displayed in a list on the same screen.

(8) In the pattern input method according to any one of the above (1) to (7), when the identity is manually confirmed in the process of assigning a character code to a similar pattern, the same A pattern input method characterized in that when a dictionary pattern and a plurality of similar patterns are displayed in a list on a screen, the plurality of similar patterns are arranged in the order of high or low similarity with the dictionary pattern.

(9) In the pattern input method according to any one of the above (1) to (8), when the identity is manually checked in the process of assigning a character code to a similar pattern, the same A pattern input method characterized in that when a dictionary pattern and a plurality of similar patterns are displayed in a list on a screen, the similar patterns are cut out and displayed in an area large enough to include a pattern present in the vicinity thereof.

(10) A program for causing a computer to execute the pattern input method according to any one of (1) to (9).

(11) A recording medium which records a program for causing a computer to execute the pattern input method according to any one of the above (1) to (9).

[0022]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS As embodiments of the present invention, the following two embodiments will be described in detail with reference to the drawings. The first embodiment is a method in which an undefined pattern is cut out from a binary input image of black and white in the order of appearance, and a similar pattern is searched and extracted using the contour feature of the pattern as a feature amount. The second embodiment is a method of extracting undefined patterns in descending order of appearance frequency.

(Embodiment 1) FIG. 1 shows a processing procedure of Embodiment 1. 100 is a binary image represented by binary values of white and black input by the image input device, 200 is a process of cutting out an undefined pattern as a dictionary pattern, 300
Is a cut-out dictionary pattern, 400 is a feature amount measurement process of the dictionary pattern and the input image, 500 is a process of searching and extracting a similar pattern from the input image, 600 is an extracted similar pattern, and 700 is an extracted similar pattern. Is a process for determining whether or not all patterns have been input.

The input image 100 is an image input as a white and black binary image by an image input device such as a scanner, a facsimile, a camera, or the like. An example of an input image is shown at 110 in FIG.

The dictionary pattern cutout process 200 is a process for cutting out the first pattern that has not been converted into a character code from the input image to create a dictionary pattern image. The clipping method is a method in which an input image developed on a computer memory is cut out by operating a device such as a mouse using a device such as a mouse in the same manner as general image processing software, and a method in which only a dictionary pattern portion is input again from an image input device. be able to.

The dictionary pattern 300 is a binary image extracted from the input image by the dictionary pattern extraction processing 200. An example of the created dictionary pattern is shown in FIG.
It is shown in FIG.

The feature value measuring process 400 is a process for measuring a feature value used for matching between images from each of the dictionary pattern and the input image. There are various features such as density values and contour features. Here, it is one of the contour features.・ Recognition of Hiragana ”, IEICE (D), J70
-D, No. 7, pp. 1390-1397, 1987). However, here, the image is not divided into small regions in advance, but a small region centering on the black pixel of interest is created, and the contour direction histogram is obtained and the two-dimensional Gaussian filter is used in the small region. It shall be.

For each of the dictionary pattern and the input image, a weighted average histogram feature obtained by quantizing the outline direction in four directions, for example, in increments of π / 4, is obtained. A four-dimensional vector representing direction information is provided.

The similar pattern search / extraction process 500
This is a process of searching for a portion where there is little difference between the two while moving the dictionary pattern over the input image, and extracting the portion as a similar pattern. Various measures such as a Euclidean distance, a Mahalanobis distance, and a cross-correlation coefficient between the feature amounts of the dictionary pattern and the input image are used as the scale indicating the difference between the dictionary pattern and the input image. Here, a method using the Euclidean distance will be briefly described.

The components of the four-dimensional vector given to each pixel of the dictionary pattern having a size of M × N pixels are made continuous,
Create a 4 × M × N-dimensional vector. Similarly, for an input image overlapping with a dictionary pattern, 4 × M
Create a × N-dimensional vector. Assuming that the movement amount of the dictionary pattern at this time is (X, Y), the Euclidean distance R (X, Y) between the feature amounts is

[0031]

(Equation 1)

## EQU2 ## Here, T _i is a component of the dictionary pattern, and I _i is a component of the input image. The smaller the value of the Euclidean distance R, the smaller the difference between the images. If the processing time when calculating the Euclidean distance R becomes a problem, increase the dictionary pattern movement amount per time,
Alternatively, the speed can be increased by a method such as reducing the resolution of an image.

The result of searching the input image 110 using the dictionary pattern 310 and calculating the Euclidean distance between the feature values at each point is shown as 510 in FIG. The higher the density, the smaller the distance value. By extracting an image of a dictionary pattern size from the input image based on the coordinates of a point having a high density at 510, a pattern similar to the dictionary pattern can be extracted.

The similar pattern 600 is searched and extracted 5
This is a similar pattern extracted by “00”. 6 in FIG.
10 shows a similar pattern extracted from the input image 110 by the dictionary pattern 310.

The character code assigning process 700 is a process of assigning the same character code to the similar pattern 600. At this time, it is assumed that all the extracted similar patterns are the same as the dictionary pattern, and a method of automatically assigning a character code, and a method of manually confirming the identity between the similar pattern and the dictionary pattern while manually confirming the identity. Can be provided. The confirmation by hand can be performed at high speed by performing a simple operation such as clicking a mouse button.

The pattern input determination processing 800 is a processing for determining whether or not input processing has been performed on all patterns in an image. If the undefined pattern still exists in the input image, the process returns to the dictionary pattern cutout processing 200 for manually cutting out the undefined pattern again,
Repeat the whole process. Whether an undefined pattern exists in an input image can be determined by a method automatically performed by a computer or a method performed manually.

The automatic method is as follows, for example. Since a pattern is a set of black pixels, a portion where black pixels are densely packed to a certain extent within a certain range is considered to be a part of a certain pattern. Therefore, it is possible to determine whether or not the input of all the patterns is completed by extracting a portion in which a black pixel having a certain value or more exists within a certain range and confirming whether or not a part of the pattern has already been input. If there is a black pixel dense portion that is not a part of the already input pattern, it means that an undefined pattern still exists in the input image.

In the manual method, a previously input pattern is colored and displayed, and the entire input image is listed to easily find an uncolored undefined pattern.

(Embodiment 2) As Embodiment 2 of the present invention, a method of extracting an undefined pattern in order from a pattern having a high frequency of appearance in an image will be described.

The appearance frequency of the pattern appearing in the input image is biased. Therefore, by performing pattern cutting in the order of appearance frequency, the speed at which undefined patterns are reduced is higher than in the first embodiment. Therefore, an undefined pattern to be extracted next can be efficiently found. In the case where the labor cost is limited, a method of cutting out a pattern having a high appearance frequency is the most efficient.

(Embodiment 3) As Embodiment 3 of the present invention, a method for easily determining whether or not a dictionary pattern and a similar pattern are the same in assigning a character code to a similar pattern extracted by a search. Will be described.

Reference numeral 710 in FIG. 6 shows an example of a window display for manually confirming whether or not the dictionary pattern and the similar pattern are the same. A dictionary pattern and a plurality of similar patterns are displayed in a list on the same screen. Here, a total of 1 similar patterns are arranged in a line by 5 patterns.
This is the case where 0 pattern is displayed.

Reference numeral 720 denotes a dictionary pattern cut out from the input image, which is the same as 310 in FIG.

As described above, the list display of the dictionary patterns and the similar patterns enables not only the comparison between the dictionary patterns and the respective similar patterns but also the comparison between the similar patterns.
It is more efficient to determine whether a dictionary pattern and a similar pattern are the same.

Reference numeral 730 denotes a plurality of extracted similar patterns. Here, the similar patterns are arranged in the order of higher similarity with the dictionary pattern. By arranging the patterns in the order of similarity in this manner, the tendency of correctly extracted patterns and erroneously extracted patterns being individually arranged increases. In addition, by extracting and displaying not only the similar pattern itself but also an area including a pattern existing in a neighboring area (upper and lower in this example), the pattern existing in the neighboring area can be used as a reference when confirming identity. it can.

Reference numeral 740 denotes a command button for performing confirmation processing (cancel processing) of a plurality of similar patterns by one input operation. Since similar patterns are arranged in descending order of similarity, correctly extracted patterns have a strong tendency to be solidified, and it is efficient if there is a command button for which a certain number of similar patterns are collectively processed. In this example, the five patterns arranged in the upper row are all correct patterns, so that batch processing is possible. Efficiency is improved five times as compared with a case of operating once for each of the five patterns. be able to.

When only a part of the similar patterns to be collectively processed is correctly extracted, the area within the frame in which the correctly extracted pattern is displayed is individually clicked with a mouse button. Only the correct pattern can be selected by the operation described above.

It is to be noted that a part or all of the processing functions of the pattern input method shown in FIG. 1 or the like may be implemented as a program and implemented using a computer, or the processing procedure shown in FIG. 1 or the like may be implemented as a program. Computer. Further, a computer-readable recording medium readable by a computer, the program for realizing the processing function of each unit, or the program for causing the computer to execute the processing procedure,
For example, FD (floppy (registered trademark) disk), M
O, ROM, a memory card, a CD, a DVD, a removable disk, etc., and can be stored and provided, and can be distributed via a communication network such as the Internet. It is possible.

[0049]

As described above, according to the present invention, the extraction of the undefined pattern and the extraction of the similar pattern are repeated, and the undefined pattern is reduced step by step.
It enables efficient pattern input with minimum necessary human intervention in cases where the appearance pattern variation is unknown or very large. In addition, taking into account the fact that there is a bias in the frequency of appearance of the pattern, by performing the cutout in the order of appearance frequency, more efficient pattern input becomes possible.

[Brief description of the drawings]

FIG. 1 is a processing procedure diagram showing an embodiment of the present invention.

FIG. 2 is an example of a pattern input image.

FIG. 3 is an example of a dictionary pattern.

FIG. 4 is an image example in which a distance value between a dictionary pattern and an input image is obtained.

FIG. 5 is an example of a similar pattern in an image.

FIG. 6 is an example of a confirmation screen for a similar pattern.

[Explanation of symbols]

 100: binary image 200: dictionary pattern cutout processing 300: dictionary pattern 400: feature quantity measurement processing 500: similar pattern search / extraction processing 600: similar pattern 700: character code assignment processing 800: determination processing

 ──────────────────────────────────────────────────続 き Continuing on the front page (72) Inventor Akio Shio 2-3-1 Otemachi, Chiyoda-ku, Tokyo Within Nippon Telegraph and Telephone Corporation (72) Inventor Sakuichi Otsuka 2--3, Otemachi, Chiyoda-ku, Tokyo No. 1 Nippon Telegraph and Telephone Corporation F term (reference) 5B064 AA01 AB02 AB03 DA03 DA34 DC00 EA08

Claims (11)

[Claims] 1. A method of converting various patterns present in an input image into character codes and inputting the converted characters, wherein a first step of cutting out an undefined pattern without a character code in the input image as a dictionary pattern A second step of measuring a dictionary pattern obtained by the extraction and a feature amount of the input image; and searching the input image for a pattern similar to the dictionary pattern based on the feature amount obtained by the feature amount measurement. A third step of assigning the same character code to all of the similar patterns extracted in the above step, and a fourth step of assigning the same character code to the similar pattern extracted in the above step. A pattern input method characterized by repeating the steps up to step 4. 2. The pattern input method according to claim 1, wherein in the step of cutting out an undefined pattern from the input image, the pattern is cut out in order from a pattern having a high frequency of appearance in the input image. Method. 3. The pattern input method according to claim 1, wherein in the step of assigning a character code to the similar pattern extracted by the search, a determination is made as to whether or not the dictionary pattern and the similar pattern are the same. A pattern input method characterized by being obtained as an input. 4. The pattern input method according to claim 1, wherein, when the input image is a black and white binary image, the outline feature of the pattern is used as a feature amount used for the search. Characteristic pattern input method. 5. The pattern input method according to claim 1, wherein in the step of extracting a dictionary pattern and a similar pattern from an image, a determination criterion for determining whether or not the pattern is similar is a feature quantity. A pattern input method characterized by using a distance value between the patterns. 6. The pattern input method according to claim 1, wherein a weighted average histogram, which is one of the contour features, is measured from both the dictionary pattern and the input image. A pattern input method comprising: creating a small area centered on a black pixel; acquiring a contour direction histogram and performing blurring by a two-dimensional Gaussian filter in the small area. 7. The pattern input method according to claim 1, wherein when assigning a character code to the similar pattern extracted by the search, the step of manually confirming the identity includes: A pattern input method characterized by displaying a list of dictionary patterns and a plurality of similar patterns on the same screen. 8. The pattern input method according to claim 1, wherein in the process of assigning a character code to a similar pattern, when the identity is manually checked, the dictionary pattern is displayed on the same screen. And displaying a plurality of similar patterns in a list, wherein the plurality of similar patterns are arranged in the order of high or low similarity with the dictionary pattern. 9. The pattern input method according to any one of claims 1 to 8, wherein in the case of confirming the identity by hand in the process of assigning a character code to a similar pattern, the dictionary pattern is displayed on the same screen. And displaying a plurality of similar patterns in a list, wherein the similar patterns are cut out and displayed in an area large enough to include a pattern present in the vicinity thereof. 10. A program for causing a computer to execute the pattern input method according to any one of claims 1 to 9. 11. A recording medium on which a program for causing a computer to execute the pattern input method according to claim 1 is recorded.