JPH01259475A - Character recognizing device - Google Patents

Abstract

PURPOSE:To improve a recognizing rate without lowering whole processing speed by detecting the character code of a highest appearing rate as the first candidate character of a final result and further executing other selection processing when the appearing rate of the first candidate character is equal. CONSTITUTION:A candidate character selecting means 9 selects the character code, whose appearing rate is the highest out of the character codes to be selected as the respective first candidate characters of plural identifying means 7a-7c, as the first candidate character of the final result. Then, when the appearing rates of the first candidate characters are equal, the other selection processing is further executed. Thus, to the characters of different character shapes or a deformed character, the improvement of the recognizing rate can be achieved without lowering the whole processing speed. Then, by providing the plural identifying means, a character recognizing device can be obtained to allow the deformed characters as much as possible to correspond to the characters as much as possible.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to a character recognition device, and particularly to a character recognition device that reads character image data from an optical reading means and converts it into a character code.

[Prior Art] Conventionally, this type of device creates at least one character dictionary for each character to be recognized, performs identification processing on a given input character, and performs identification while referring to the character dictionary. The system is configured to perform processing and output the final candidate characters. In this case, many rely on statistical approaches, so even if the characters are extremely different, e.g.
° may be mistakenly recognized as the character ``snow''. Therefore, by providing post-processing that further evaluates candidate characters at the word level, the recognition rate is improved and extreme errors are eliminated. However, in this case, a word dictionary is required in addition to a character dictionary.

[Problems to be Solved by the Invention] However, in the above conventional example, if the restrictions on character shapes are loosened even slightly in the process of identifying input characters with a character dictionary, the recognition rate of the first candidate character drops dramatically. In addition, in post-processing at the word level, there are drawbacks such as the word dictionary becoming enormous and consecutive characters being erroneously changed to misread characters.

Therefore, (1) Set limits when writing by hand. For print, limit fonts.

(2) Limit the number of characters.

(3) Post-processing is limited depending on the text (for example, a dictionary corresponding to the address is provided for an address).

The current situation is that we are dealing with this.

The present invention eliminates the drawbacks of the conventional example and provides a character recognition device that can improve the recognition rate for characters with different shapes and modified characters without reducing the overall processing speed.

Specifically, by providing a plurality of identification processes, a character recognition device is provided that can handle as many modified characters as possible for as many single characters as possible.

[Means for solving the problem] In order to solve this problem, the character recognition device of the present invention has the following features:
A character recognition device that recognizes a character pattern by comparing it with a pre-stored pattern includes a plurality of identification means each performing different identification processing on the character pattern, and a first identification means of each of the plurality of identification means. Among the character codes selected as candidate characters,
The character code having the highest appearance rate is selected as the first candidate character in the final result, and when the appearance rates of the first candidate characters are equal, the candidate character selection means further performs another selection process.

Here, when the appearance rates of the first candidate character are equal, the candidate character selection means selects the final result from the sum of the appearance rates of the second and subsequent candidate characters to the appearance rate of the first candidate character. Select the first candidate character.

Further, when the appearance rate of the first candidate character is equal, the candidate character selection means multiplies the plurality of candidate character strings from the plurality of identification means by a weighting coefficient according to the candidate ranking, and calculates the final number from the sum of each character code. Generate candidate strings.

[Operation] In such a configuration, each of the plurality of identification means executes different identification processing, and the candidate character selection means selects the most appearing character code among the character codes selected as the first candidate character of each of the plurality of identification means. A character code with a high rate is selected as the first candidate character in the final result, and when the appearance rates of the first candidate characters are equal, another selection process is performed.

This candidate character selection means selects the final result of the first candidate character from the sum of the appearance rates of the second and subsequent candidate characters to the appearance rate of the first candidate character when the appearance rates of the first candidate characters are equal. Select candidate characters.

Further, when the appearance rate of the first candidate character is equal, the candidate character selection means multiplies the plurality of candidate character strings from the plurality of identification means by a weighting coefficient according to the candidate ranking, and calculates the final number from the sum of each character code. Generate candidate strings.

In other words, in the present invention, although the candidate character group of recognition results obtained from one identification process and the candidate character group of recognition results obtained from another identification process are almost the same in terms of characters that are easy to identify, For characters with shapes that are easily misrecognized, different output results are calculated through identification processing.
This method takes advantage of the fact that even if a character is erroneously recognized in one identification process, it will hardly be misrecognized if another identification process is performed.

Embodiment FIG. 1 is a block diagram of the character recognition device of this embodiment. The input section 1 in Fig. 1 is an input section of an optical image reading nail device, etc., and the normalization section 2 has a fixed size because the size of the characters cut out one by one varies depending on the shape of the character even if the size of the same character is the same. Normalize to size. A filtering section 3 shapes the deformation of the character shape due to scaling, and a feature extraction section 4 extracts features. This feature extraction requires a very large vector space for character image data, so the information is compressed while keeping the features of the data intact. The quality of this information contraction has a great influence on the subsequent processing results, but will not be explained in detail in this embodiment.

Next, the major classification section 5 performs major classification processing based on the feature extraction results. In this major classification process, after relatively simple arithmetic processing, the sorting section 6 performs sorting to select candidate characters from about 50th to 100th place. Regarding the character code data obtained by sorting, the identification unit 7
a, the identification section 7b, and the identification section 7c perform three different types of identification processing.

In the three types of discrimination processing, the following quadratic discrimination function G(x) □ [II x-, tM II 2 is defined.

Here, X is the feature vector found by feature extraction, μ
2 is the average feature vector of the letter M, 11...11
represents the norm. Further, H and A are constants shown for each character, λ1 is an eigenvalue, and φ is an eigenvector for the eigenvalue. All characters other than X are converted into a dictionary.

Based on the above formula, calculation is performed by setting 5 in the identification section 7a, 7 in the identification section 7b, and 9 in the identification section 7c. And sorting sections 8a and 8b.

Sort by 8c. Sorting section 8a.

Based on the output results of 8b and 8c, a final candidate character selection section 9 performs the final selection of candidate characters, and outputs them from an output section 10.

It should be noted that when using a different identification process according to the present invention, the value of the above equation =5. ni=7 is considered to be an intermediate result of the calculation up to k=9. Therefore, after calculating k = 5, perform the first sorting, and then use the values in [ ] of the formula up to k = 5 to calculate = 7 (6, 7) and perform the second sorting. do. The same applies to k=9 below. Also, k
=5. : The sorting results up to =7 are the same (first
When the position candidate characters are the same), the calculation ends, and k=
It is not a difference in the present invention that the operation of 9 is omitted.

By omitting operations in this manner, the microprocessor described below can be used for other purposes, making it more efficient.

FIG. 2(a) shows an example of the hardware configuration of the character recognition device of this embodiment. The character data input from the input unit 20 is sent to the CPU according to the program stored in the ROM 22.
21 performs calculations and processing while using the RAM 23 as an auxiliary memory. The character dictionary 24 stores in advance character patterns used for major classification and identification. The recognition result is output from the output section 25. The above units are connected by a bus 26 and exchange data via this. However, in the configuration shown in FIG. 2(a), if the process from input to output is performed sequentially, the elapsed time until recognition becomes long.

Therefore, as shown in FIG. 2(b), the preprocessing section 27 operates from the input section 1 to the feature extraction section 4 in FIG. If the components from ~7C to the output section 10 are combined into the identification processing section 29, and each section 27, 28, and 29 is realized by a respective microprocessor, the processing time delay due to the use of three identification sections can be avoided. It can be absorbed. Furthermore, the identification section 78
7C and the sorting units 88 to 8c may be performed in parallel.

What is shown in FIG. 3 is the first to third rankings of the sorting units 8a, 8b, 8c for the input character data example "Ku".
This is an example of output results up to Based on the output results, the candidate character selection section 9 performs final candidate character selection, and outputs the result from the output section 10.

Several examples of selection by the candidate character selection section 9 will be shown below.

<Example 1> Focusing on the first candidate character output from the sorting sections 88 to 8c, "Ku", which is the first candidate character of the sorting sections 88 to 8c, is selected as the final first candidate character.

Example 2 If selection cannot be made due to the high or low appearance rate as in Example 1, the second and third place candidates are accessed in order and the one with the higher appearance rate is selected.

For example, the first and second ranks output from the sorting unit 7b
Exchange positions and explain. The first place is °゛present”, “true”
, the "Emperor" is selected, and each appearance rate is equal. Next, when the second candidate character is accessed, "゛present" and "Ku" survive.Next, the third candidate character is accessed, and "Ku" is selected as the first candidate for the final output. That will happen.

<Example 3> Candidate characters are weighted in order of priority and prioritized in descending order of the total sum of each character.

As in Example 2, a description will be given with the output of the sorting section 8b being exchanged. Here, considering the weighting coefficients of 3 for the first place, 2 for the second place, and 1 for the third place, 3 x "presentation" is obtained from the output of the sorting section 8a.

2 x "Hoshi" 11 x "Kou", 3 x "Shin" from the output of the sorting unit 8b, 2 x "Kou", 1 x "Chang", and 3 x "Kou" from the output of the sorting unit 8c.

2 x “Sei”, IX “Wataru”, resulting in 6 x “Kou”
, 5X "presentation"... is obtained. Therefore, "Kou" will be selected as the first candidate.

The processes in Examples 1 to 3 above may be performed independently, or a plurality of processes may be combined and priorities may be given to each process. Note that the processing is not limited to the above example.

FIG. 4 shows a timing chart of the processing of this embodiment. As shown in the figure, it can be seen that performing the three identification processes has no effect on the total throughput. Here, the detailed classification is a portion indicating sorting by the identification sections 78 to 7C and the respective sorting sections 88 to 8c.

In other words, what is indicated by ■-1 is the identification of the first character by the identification section 7a and the sorting by the sorting section 8a.

[Effects of the Invention] According to the present invention, it is possible to provide a character recognition device that can improve the recognition rate for characters with different shapes and modified characters without reducing the overall processing speed.

Specifically, by providing a plurality of identification processes, it is possible to provide a character recognition device that can handle as many modified characters as possible.

[Brief explanation of the drawing]

Figure 1 is a block configuration diagram of the character recognition device of this embodiment. Figures 2 (a) and (b) are hardware configuration diagrams of the character recognition device of this embodiment. Figure 3 is an example of output from the sorting section. FIG. 4 is a diagram showing a processing timing chart of this embodiment. In the figure, ■... Input section, 2... Normalization section, 3... Filtering section, 4... Feature extraction section, 5... Major classification section, 6... Sorting section, 7a, 7b, 7c...
Identification section, 8a, 8b, 8c... Sorting section, 9.
...Candidate character selection section, 1o...Output section, 20...Input section, 21...CPU, 22...ROM,
23...RAM. 24... Character dictionary, 25... Output section, 26... Bus, 27... Preprocessing section, 28... Main classification processing section, 2
9...Identification processing section.

Claims (3)

[Claims] (1) In a character recognition device that recognizes a character pattern by comparing it with a pre-stored pattern, a plurality of identification means each perform a different identification process on the character pattern, and each of the plurality of identification means Among the character codes selected as the first candidate character, the character code with the highest appearance rate is selected as the first candidate character in the final result, and when the appearance rates of the first candidate characters are equal, the other character codes are selected. 1. A character recognition device comprising candidate character selection means for performing a selection process. (2) When the appearance rates of the first candidate character are equal, the candidate character selection means selects the final result from the sum of the appearance rates of the second and subsequent candidate characters to the appearance rate of the first candidate character. 2. The character recognition device according to claim 1, wherein the first candidate character is selected. (3) When the appearance rate of the first candidate character is equal, the candidate character selection means multiplies the plurality of candidate character strings from the plurality of identification means by a weighting coefficient according to the candidate ranking, and from the sum of each character code. 2. The character recognition device according to claim 1, wherein the character recognition device generates a final candidate character string.