JPH09319825A - Character recognition device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a character recognition device which can set the order of candidates to category candidates with a logical discrimination technique as the base. SOLUTION: A phase structure characteristic generation part 11 generates phase structure features of a known pattern, based on a phase structuring method. A candidate-specifying part 12 refers to a recognition dictionary part 13 to specify category candidates. A distance calculation part 14 uses a phase structure feature table 15, where phase structure features of a standard pattern are described in advance with primitives, to calculate the city block distances between phase structure features of the known pattern and those of the category candidates. A order of candidate determination part 16 sets the rankings of category candidates in the order of small city block distances.

Description

Detailed Description of the Invention

[0001]

TECHNICAL FIELD The present invention relates to, for example, OCR (Op (Op
The present invention relates to a character recognition device equipped with a means for identifying an unknown character pattern captured by tical character recognition, etc., and in particular, category identification corresponding to a character pattern by combining the advantages of a logical identification method and an analytical identification method. Related to the technology of doing. In addition, when a character is used in this specification,
In addition to characters in a narrow sense, it also means figures, symbols, numbers, etc.

[0002]

2. Description of the Related Art There are two basic methods for identifying a category corresponding to an unknown character pattern (hereinafter referred to as an unknown pattern) in a character recognition device. One is a logical identification method, and the other is an analytical identification method.

The logical identification method is a category identification method based on primitive features. That is, when an unknown pattern is composed of a simple part shape, the shape as this constituent element is called a primitive, and represents the feature that is the minimum unit for describing the unknown pattern. In the logical identification method, an unknown pattern is described by this primitive, and this is compared and collated with the description of the standard pattern (standard category) which is described in the primitive and stored in the dictionary etc. The category is identified by a scale of true or false. Therefore, the character recognition result by the logical identification method is only the correct category in the dictionary that matches the description of the unknown pattern.

On the other hand, the analytical discriminating method is based on detecting the degree of similarity of the features of the unknown pattern using an analytical function. For example, each pattern feature is given by a set of parameters called a feature vector, and each component of the feature vector corresponds to an element of the pattern feature. Therefore, whether the patterns are similar to each other depends on how similar the feature vectors are. Distance is mainly used for the degree of similarity of feature vectors. The distance is given by the absolute value of the difference between any two feature vectors, for example.

When a category corresponding to an unknown pattern is identified by this analytical identification method, first, a feature vector space that is convenient for identification processing is obtained from the unknown pattern. Then, the category is identified using an analytic function that can introduce the concept of distance as the identification function. Therefore, according to the analytical identification method, the candidate ranks are set in the ascending order of the distance from the feature vector of the unknown pattern, that is, in the ascending order such as the first candidate, the second candidate, the third candidate ,. The character recognition result can be obtained. By looking at this candidate rank, it becomes possible to recognize the category closest to the unknown pattern. For a description of the above character recognition method, see "Introduction to Character Recognition", edited by Shinichiro Hashimoto (Ohmsha, 1982,
The description such as P29-45) is helpful.

[0006]

As described above, the character recognition device using the logical identification method cannot obtain the character recognition result in which the candidate ranks are set. Therefore, when multiple candidate categories appear as a result of character recognition, it is necessary to match the unknown pattern with all the category candidates to identify the best matching category, which complicates the process. was there.

In the analytical identification method, when a plurality of category candidates appear, it is convenient that the optimum category candidate can be specified immediately by looking at the candidate order set for each category candidate. If the recognizer adopts the logical identification method, the identification method must be changed to the analytical identification method. However, since the change of the identification method involves the change of the feature extraction system and the recognition control system in the character recognition, it is technically extremely difficult and disadvantageous in cost.

In order to solve such a problem, conventionally, after a character recognition result in which a candidate rank is not set, which is obtained by a logical identification method, is obtained, the correct answer candidate (= first A technique for squeezing candidates) is known. This technique is called "post-processing". However, since this post-processing uses a scale that lacks objectivity, such as "validity," there is a problem in that it is not possible to accurately know how similar the unknown pattern was to the candidate category. It was

[0009] Therefore, an object of the present invention is to provide a character recognition device which can easily and appropriately set a candidate order for each category candidate when a plurality of category candidates appear when category identification is performed by a logical identification method. To provide.

[0010]

Although not possible with the analytical identification method, the logical identification method can identify a category based on a primitive feature that matches human intuition. In the present invention, for a similar shape category that is difficult to distinguish even by human eyes, the difference portion of the similar shape described by the primitive, for example, the difference in the upper left inclination degree such as “ku” and “wa” A degree of similarity between a part and a standard pattern of the part described in advance, for example, a city block distance between feature vectors is obtained, and a candidate rank is set in the character recognition result according to the degree of similarity (distance value). .

Specifically, the similarity between the first means for specifying at least one category candidate similar to the characteristics of the unknown pattern by the logical identification method and the standard pattern described in advance for the specified category candidate by a primitive. The character recognition device is configured to include a second means for determining the degree of the above and a third means for setting the candidate rank for the category candidate according to the degree of similarity.

The phase structuring method is an example of the logical identification method. This topological structuring method is an extension of Mr. Glucksman's character recognition method, which focuses on the topological structure of an unknown pattern. A device that realizes the phase structured method has an advantage that the device configuration is simple and easy. Moreover, since the global structure of the character is extracted, it is relatively stable against noise and character deformation due to handwriting. The phase structuring method is
From the information obtained by the primary feature extraction that extracts the features of each grid area of the checkerboard-like area in which the unknown pattern is drawn, the integration that accumulates all the local information obtained by the primary feature extraction, and the information obtained by the integration It consists of higher-order generation of more accurate information that reflects global characteristics by adding other local information to each local information. The information that reflects the global structure of unknown patterns obtained through higher order is called topological structure feature. The information obtained by the primary feature extraction in the topological structuring method is a primitive description of the pattern.

The first means in the character recognition apparatus of the present invention includes, for example, a recognition dictionary unit in which the category groups of which the topological structure features are known are accumulated, and a topological structure feature generation for generating topological structure features of unknown unknown patterns. Section, and means for specifying from the recognition dictionary section a category in which the degree of coincidence of the topological structure features of the unknown pattern generated by the topological structure feature generation portion exceeds the specified value, and the identified category is defined as a category candidate. To do.

Further, the second means is, for example, a topological structure feature table in which topological structure features of a standard pattern are described by primitives, a standard pattern in the topological structure feature table for the identified category candidate, and the unknown. And a means for determining the degree of similarity of the phase structure with the pattern.

The topological structure feature table adjusts the candidate ranks by including standardized standard patterns and target-specific standard patterns in which the topological structure features of the standard patterns are modified according to the purpose. be able to.

[0016]

BEST MODE FOR CARRYING OUT THE INVENTION Embodiments of the present invention will be described in detail below with reference to the drawings. FIG. 1 is a configuration diagram of an identification system of the character recognition device according to the present embodiment. In this embodiment, category identification is performed based on the topological structuring method which is one of the logical identification methods, and the degree of similarity of topological structure features is represented by city block distance.

The character identification method in the character recognition device 1 is based on the conventional logical identification method. In order to set an appropriate candidate rank for a category candidate, the phase structure feature generation section 11 and the candidate identification section 12 are used. , The recognition dictionary unit 13, the distance calculation unit 14, and the distance calculation unit 1 which are referred to by the candidate specifying unit 12.
4 includes a topological structure feature table 15 and a candidate rank determining unit 16. Each of the units 11 to 16 may exist as an independent physical unit, or may exist as a functional group logically configured in the information processing device. Further, it may be physically embodied on a storage medium so that the information processing apparatus can read it, and may exist as an instruction set that gives the information processing apparatus the above function when the information is read.

The phase structure feature table 15 is, for example, "O".
CR-HN (JIS X 9006) "," OCR-H
K (JIS X 9005) ”, a table in which the standardized topological structure features of standard patterns are described in primitives, and a table in which the topological structural features of the standard patterns by purpose determined according to the purpose are described in primitives. , Or a combination of both can be used.

The phase structure feature generator 11 generates the phase structure feature of the input unknown pattern, and the candidate specifying unit 12 calculates the phase structure feature of the unknown pattern generated by the phase structure feature generator 11. It compares at least one category candidate by comparing the phase structure features of a plurality of standard patterns stored in the recognition dictionary unit 13.

The distance calculation section 14 calculates the degree of similarity between the phase structure features of the standard pattern and the unknown pattern in the phase structure feature table 15 for the identified category candidate, that is, the city block distance. The candidate rank determination unit 16 analyzes the calculation result of the distance calculation unit 14, and ranks the candidate ranks 1, 2, ...
Is to be set.

Next, the identification method in the character recognition device 1 of the present embodiment will be described with a specific example. First, as the topological structure feature table 15, the standardized “OCR-HN
(JIS X 9006) "and" OCR-HK (JI
S X 9005) ”will be described.

Now, assume that the unknown pattern of the shape feature shown in the upper part of FIG. 2 is input to the phase structure feature generator 11. The phase structure feature generation unit 11 generates a phase structure feature of this unknown pattern based on the phase structuring method, and uses this as a candidate identification unit 1
2 and the distance calculation unit 14. The candidate identifying unit 12 refers to the primitive description of the standard category accumulated in the recognition dictionary unit 13, and matches with a predetermined criterion or more for the primitive description included in the topological structure feature of the unknown pattern. Identify standard categories.

The process of identifying the standard category will be described in more detail. In the recognition dictionary unit 13, as shown in the figure, standard categories “6”, “7”, “8”, ..., “Ki”, “Ku”,
Primitive descriptions of “ke”, ..., “B”, “wa”, “wo”, ... Are accumulated. Of these standard categories, those with a shape relatively similar to the unknown pattern are
"7", "ku", and "wa". Candidate identification unit 12
Compares and collates such a primitive description in each pattern with the primitive description of the unknown pattern to identify three category candidates “7”, “KU”, and “WA”.
Then, these are sent to the distance calculation unit 14.

The distance calculation unit 14 stores the city block distance between the topological structure feature of the unknown pattern and the topological structure feature of the category candidates “7”, “ku” and “wa” in the topological structure feature table 15. Operate by referring to the primitive description. The calculation result in this case is as shown in the lower part of FIG. In the illustrated example, the city block distances between the topological structure features of the unknown pattern and the topological structure features of the category candidates “7”, “ku”, and “wa” are “50” and “10”, respectively.
It shows that it is "0" and "25". Since the unit of distance is not essential, it is omitted here.

As shown in FIG. 4, the candidate rank determining unit 16 sets candidate ranks “1”, “2”, and “3” in order from the category candidate having the smallest city block distance. In the illustrated example, the category candidate “wa” having the city block distance of “25” has the first rank, the category candidate “7” having the city block distance of “50” has the second rank, and the city block The category candidate “ku” whose distance was “100” is ranked third.

As described above, when a plurality of category candidates appear by the logical identification method, the candidate ranks are set for the respective category candidates, so that the operator's work for confirming or correcting the character recognition result can be performed. The burden is reduced.

Next, the above-mentioned "OCR-HN" and "OCR-HN"
A case will be described in which, instead of the phase structure feature table 15 based on “-HK”, a phase structure feature table in which the phase structure features of the purpose-specific standard patterns are accumulated. here,
Include a category corresponding to "characters you want to read" in the standard pattern for each purpose. That is, for example, to prevent misreading,
As shown in FIG. 5 (a), the numeral "7" and the alphabetic character "Z"
The topological features of a character such as may be used after being slightly modified. Such characters are called "characters to be read". In this embodiment, the topological structure feature of the "character to be read" is described in advance by a primitive and stored as one of the purpose-specific standard patterns.

The category candidates when an unknown pattern as shown in FIG. 5A is input are "7", as in the above case.
“Ku” and “wa”, of which the category candidate corresponding to “characters to be read” is “7”. Distance calculator 14
As shown in FIG. 5B, refers to the phase structure feature table 25 that accumulates the phase structure features of the target-specific standard patterns, and refers to the city structure of the topological structure features of the unknown pattern and the category structure. Calculate block distance. An example of the calculation result in this case is shown in the lower part of FIG.
In the illustrated example, each category candidate is “7”, “ku”, “wa”.
City block distances to "10" and "1" respectively
It becomes 50 "and" 110 ".

As a result of the candidate rank determining unit 16 setting candidate ranks in order from the smallest city block distance, as shown in FIG. 6, the category candidate "7" corresponding to "characters to be read" is ranked first. The first rank and the other category candidates “wa” and “ku” are the second and third ranks, respectively.

As described above, by preparing in advance the topological structure feature of the "character to be read", the rank of the corresponding category candidate is likely to be increased, and the accuracy of the character recognition result is increased. It should be noted that this "character to be read" is merely an example of the purpose-specific standard pattern, and the type of language, character type (font), etc. can be used as the purpose-specific standard pattern.

[0031]

As is apparent from the above description, according to the present invention, when a plurality of category candidates are obtained as a result of category identification by the logical identification method, an appropriate candidate for each category candidate is obtained. Since the order is set, there is an effect that the work load for checking and correcting the character recognition result is reduced. In addition, setting of this candidate order is advantageous in terms of cost because it does not require modification of the existing feature extraction system or recognition control system. Further, by using the phase structure feature of the target-specific standard pattern when setting the candidate rank, it is possible to more efficiently specify the category candidate that approximates the unknown pattern.

[Brief description of drawings]

FIG. 1 is a configuration diagram of an identification system of a character recognition device according to an embodiment of the present invention.

FIG. 2 is an explanatory diagram of a procedure for identifying a category candidate by the topological structure method.

FIG. 3 is an explanatory diagram of a distance calculation procedure in a distance calculation unit.

FIG. 4 is an explanatory diagram showing an example of determining a candidate ranking.

FIG. 5 (a) is a diagram showing an unknown pattern to be input,
FIG. 7B is an explanatory diagram of a distance calculation procedure in the distance calculation unit when this pattern is input.

FIG. 6 is an explanatory diagram showing an example of determining a candidate rank when the pattern of FIG. 5A is input.

[Explanation of symbols]

 1 Character Recognition Device 11 Topological Structure Feature Generation Unit 12 Candidate Specification Unit 13 Recognition Dictionary Unit 14 Distance Calculation Unit 15 Phase Structure Feature Table 16 Candidate Rank Determination Unit

Claims (5)

[Claims] 1. A first means for identifying at least one category candidate similar to a feature of an unknown character pattern by a logical identification method, and a similarity between a standard pattern described in a primitive for the identified category candidate in advance. A character recognition device comprising: a second means for obtaining a degree; and a third means for setting a candidate rank according to the degree of similarity to each category candidate. 2. The first means includes: a recognition dictionary unit in which category groups having known topological structure features are accumulated in advance; a topological structure feature generating unit that generates topological structure features of unknown character patterns; Means for specifying a category in the recognition dictionary unit in which the degree of matching of the phase structure features of the character pattern generated by the generation unit exceeds a specified value, and the specified category is set as the category candidate. The character recognition device according to claim 1. 3. The second means comprises: a topological structure feature table in which topological structure features of the standard pattern are described in primitives; standard patterns and character patterns in the topological structure feature table for the identified category candidates. 2. The character recognition device according to claim 1, further comprising: a means for determining a degree of similarity of the topological structure feature with. 4. The character recognition device according to claim 1, 2 or 3, wherein the measure of the degree of similarity is a city block distance of a feature vector. 5. The character recognition apparatus according to claim 3, wherein the topological structure feature table includes a per-purpose standard pattern in which a part of the topological structure feature of the standard pattern is modified according to the purpose. .