JPS5852267B2 - Character analysis method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a character analysis method, and more particularly to a character analysis method for recognizing characters (symbols) by analyzing character image data obtained from a character reading device.

Character images such as alphabets, kana characters, numbers, symbols (including punctuation marks, etc.) are recognized based on information effective for identifying the character image, that is, the characteristics of the character image.

Such features include size, loops, end points, branch points, recesses, protrusions, etc., and their number and positional relationships.

For example, the number 6 shown in Figure 1 has features such as one end point at the top, one loop at the bottom, one 3-branch point at the left of the center, and a recess opening to the right in the upper half. be identified.

Each character image has an extremely large number of features, and the more features selected, the more accurately the character image can be recognized.However, in order to efficiently analyze a character image, it is necessary to reduce the number of features as much as possible. It is required to select less.

Therefore, in order to satisfy both of these requirements, character images have been analyzed based on the characteristics of, for example, about 160 species of deer.

The conventional character analysis method is explained using Figure 3. Figure 3 shows the extraction of all 160 types of features necessary for analyzing each character image, the analysis procedure for character analysis, and the analysis results, that is, the answers. This figure shows a storage device called a dictionary that stores .

On the top row of the horizontal axis, 160 types of features to be checked are arranged in the order of analysis from the left side, and from the next row, the presence or absence of 160 types of features for each character image is shown, that is, "l" if the feature is present, and "l" if it is absent. "O" is stored if the feature does not need to be checked, and "x" is stored if the feature does not need to be checked.

The right end stores the answer to the analysis result, that is, the name of a character image that has features that completely match the features stored in the horizontal axis column.

The number of character images that correspond to this answer is kept within three at most.

Now, when a character image is read by a character reading device, for example, a scanner such as OCR, 1
It is converted into a character pattern that indicates the presence or absence of 60 types of characteristics using binary codes.

This character pattern is compared with the characteristics of each character stored in the dictionary, and when all the characteristics match, the rightmost answer is displayed as output.

For example, the generated character pattern is "0111000...
. . 1", the analysis result indicates that the character image is the character "ho".

Such a conventional character analysis method is advantageous in that it is easy to create a dictionary that stores analysis procedures and analysis results.

In other words, there are about 120 types of character images such as alphabets, kana characters, numbers, and memory, and one character image can be written in many ways depending on the typeface and individual differences, but in general
It is considered statistically sufficient if we consider a set of 10,000 characters.

Therefore, even if 160 types of characteristics are stored for all of these character images, a storage capacity of 160 million bits, that is, 20 million bytes (=20 megabytes) is sufficient.

Then, statistics are collected from the frequency of occurrence and the answer for each character image, characteristics for classification are detected, and character analysis can be performed by storing the characteristics.

However, this conventional character analysis method has the following drawbacks.

One of the problems is that different answers, that is, character images, produce the same characteristics as character patterns, so that the answer is that they are the same character image.

For example, this is likely to occur when the font of a character image can be written.

This drawback is caused by the insufficient number of features, but if the number of features is increased to improve this, the storage capacity of the dictionary will first increase and the dictionary structure will become larger.

As the number of features to be checked increases, the time required for character analysis also increases.

Furthermore, new drawbacks arise, such as the need for a lot of time and effort to develop new features.

Another disadvantage is that character parsing takes time.

In other words, in the dictionary, the features that do not need to be checked are set as "x", the ones that are not checked as [j or OJ], and mask pitches are set for each feature, and only the features that need to be checked are checked and the characters that are analyzed as features in the dictionary are set. When the characteristics of the patterns match, the analysis is terminated and the answer is obtained.

For this reason, analysis takes time because all unnecessary features are extracted from among the features.

Furthermore, if the number of features is increased to prevent errors from occurring, the analysis time will increase accordingly.

SUMMARY OF THE INVENTION It is therefore an object of the present invention to eliminate the above-mentioned drawbacks of the conventional methods, to facilitate the increase of features and to reduce the occurrence of errors. This provides a character analysis method that requires only a very small increase in time.

To this end, the character analysis method of the present invention converts character image data into contour line analysis information, stores all positions of contour line segments in plane coordinates, and creates a feature dictionary of character sets to be classified according to their contents. In a character analysis method in which a character image is analyzed using a feature dictionary, the feature dictionary divides a set of character images to be analyzed by the features that can be divided with maximum efficiency, and further divides each divided character image set into two parts. The character image is divided according to the features that can be divided with maximum efficiency, and this process is repeated multiple times to create multiple stages. When analyzing a character image, the character image is analyzed for each divided feature, and based on the analysis results, the next stage is divided. The feature is that character analysis is performed by determining the characteristics.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Before describing the present invention in detail, an outline thereof will be explained below.

Character images such as alphabets, kana characters, numbers, symbols, etc. can be divided into groups according to their characteristics.

For example, there is a clear difference in size between a character group consisting of alphabets, kana characters, and numbers and a symbol group (for example, punctuation marks).

Therefore, you can determine the size from the character pattern to be analyzed, and if you know that it belongs to a character group, you only need to check the characteristics that belong to the character group.
Checking for features belonging to symbol groups becomes unnecessary.

Character groups are also classified into two groups depending on the presence or absence of a predetermined feature, such as the presence or absence of a loop, and if it is found to have the features of one group, it is further classified based on another feature.

That is, as shown in Fig. 4, character images are subdivided in a tree format according to the presence or absence of a predetermined feature, and the character pattern to be analyzed is sequentially classified, starting from the presence or absence of feature (1) in the top row, to the presence or absence of the next feature. By checking the following, you can complete the character pattern analysis and obtain the answer in much less time than traditional methods.

This invention was made based on this idea.

The tree shown in FIG. 4 is created as follows.

First of all, character images are characterized by being able to divide all character images (approximately 120 types) into two with maximum efficiency, i.e., character images can be divided into two so that items belonging to one group do not belong to the other group. A database is created by simulating all of
).

Next, for each of the divided groups, the characteristic that can divide the group into two with maximum efficiency was similarly found by computer statistical processing, and as shown in Figure 4, the characteristic (2), the % characteristic (3 ).

Repeat the same process to complete the entire tree.

Each school stealing the completed tree will fall under one of the 160 characteristics used in the conventional method dictionary.

For example, feature (1) is the size of the character image, feature (2)
) includes the presence or absence of loops, and feature (3) includes the number of loops.

Other major features include end points, bifurcations, unevenness, and their number and position.

In actual analysis, the position of these features in the tree will be statistically determined based on the distribution and frequency of use of character images used by each user.

Next, details of the present invention will be explained based on FIG. 5, which is an embodiment.

In the figure, 1 is a scanner, 2 is a processor, 3 is a video memory, 4 is a contour data memory, 5 is a dictionary memory, 6 is a feature extraction language analysis microprogram memory (hereinafter referred to as microprogram memory), and 7 is a photoelectric This is the conversion section.

The scanner 1 scans and reads the input data input from the photoelectric converter 7, and converts the binary value, for example, the character part into "
A character pattern such as "l" and the blank part is "0" is generated and stored in the video memory 3 in scanning order.

The processor 2 performs various controls based on the instructions of the control program stored in the microprogram memory 6, and sequentially reads out character patterns stored in the video memory 3 and prints them on each character image frame. The contour data is calculated and stored in the contour data memory 4.

Contour data is coordinate data that represents the outline of a character image.
In the number 6 shown in Fig. 1, point P 1 indicates its outline.
t P 1 + P2 , P2' y P 3 y
q3 y q3 zP!・・・・・・The coordinate data of Plo and Pll, and the number l shown in FIG. 2, are the points r, tr rr r2 z r2'+ r
3 y r3 "song" The coordinate data of Plo represents the outline data of the character.

Since the character image changes from "0" to rlJ or from "1" to "O" at its outline, the coordinates of the outline can be found from this.

Then, scanning is first performed in the X-axis direction and then in the Y-axis direction to obtain coordinate data of the outline of the character image that intersects each scanning line, and this is stored in the outline data memory 4 as outline data.

By analyzing this contour data, the features of the character image can be extracted.

For example, the size can be determined from contour coordinate data in the X-axis direction.

In the case of a loop, as shown in Figure 1, the difference in contour data in the X-axis direction ("t'-"t)y (P2'-P2)...
The value of is large, and the rate of change of the difference is also large.

On the other hand, in the case of end points, as shown in Figure 2, the difference in contour coordinate data in the X-axis direction is % (rl''rl) y
This can be determined from the fact that the value of ("-r2)"" is small and the rate of change of the difference is small.

The dictionary memory 5 stores the features to be analyzed, that is, the features stored in the conventional dictionary shown in FIG. 3 (which may be the same or different) and their execution instructions. It is.

This dictionary memory is created by statistically processing the data based on the data to be classified as described above so as to create an optimal dictionary with the characteristics of maximum efficiency.

The microprogram memory 6 is a control memory in which a microprogram for performing an operation for determining the presence or absence of each feature from contour data is stored.

This microprogram memory 6 is activated when the outline data of a character image is written into the outline data memory 4, and first, the first feature to be analyzed by the dictionary memory 5 is determined by the start address, that is, the feature shown in FIG. 1) and its execution instruction data are taken into the processor 2.

The processor 2 determines what feature should be extracted from the character image from the feature (1) data, retrieves the outline data of the character image from the outline data memory 4, and stores it in the microprogram memory 6. The above characteristics (1)
) is used to analyze the contour data and determine the presence or absence of feature (1).

Then, based on the determination result, it is determined which feature should be determined next, the address of the feature is specified, and data for executing the feature is fetched from the dictionary memory 5 into the processor 2.

For example, in the case of FIG. 4, when feature (1) is extracted, the presence or absence of feature (2) is determined.

The operation regarding feature (2) is the same as that for feature (1).

This operation is then repeated to proceed with character analysis, and the analysis is terminated when the final answer is obtained.

Therefore, in the present invention, the number of times of feature discrimination to be performed before an answer is obtained is far less than in the conventional method, so that character analysis can be completed in much less time than in the conventional method.

Furthermore, even if a new feature to be checked arises, it is only necessary to add that feature to the dictionary memory 5 and add a program to analyze the feature to the microprogram memory 6, so the capacity of the dictionary memory 5 can be increased. Less is needed.

Also, in character analysis, only the group that needs to check this new feature checks it, and other groups do not need to check this new feature, so the analysis time is not affected.

As explained above, according to the present invention, the analysis is performed using the minimum required analysis procedure, so the number of steps required is far less than that of the conventional method.

Moreover, it is extremely easy to add new features to be analyzed to compensate for the lack of features, and based on this, there is only a slight increase in the dictionary memory capacity, and the analysis time is It has no effect on anything.

[Brief explanation of drawings]

Figures 1 and 2 are explanatory diagrams of features and contour coordinates of character images, Figure 3 is an explanatory diagram of a conventional dictionary, Figure 4 is a tree diagram showing the analysis procedure of the present invention, and Figure 5 is an explanatory diagram of the character image. FIG. 1 is a configuration diagram of an embodiment of the invention. In the figure, 1 is a scanner, 2 is a processor, 3 is a video memory, 4 is a contour data memory, 5 is a dictionary memory, 6 is a microprogram memory, and 7 is a photoelectric conversion unit.

Claims (1)

[Claims] 1. Convert character image data to contour line analysis information, store all positions of contour line segments in plane coordinates, create a feature dictionary of character sets to be classified according to their contents, and classify character images using the feature dictionary. In the character analysis method to be analyzed, the feature dictionary is divided by a feature that can divide a character image set to be analyzed with maximum efficiency, and is further divided by a feature that can divide each divided character image set with maximum efficiency. It is configured in multiple stages by repeating multiple times, and when analyzing a character image, the character image is analyzed for each divided feature, and based on the analysis results, the next stage's features are determined and character analysis is executed. A character analysis method characterized by the following.