JPH07200738A - Character recognition method - Google Patents

Abstract

PURPOSE:To improve the performance of recognition on the conditions of extension possibility for the character type of an object. CONSTITUTION:Strict conditions are decided for each model. At the same time, dedicated identification processing is decided to the set of models to be matched as needed. In the case of character recognition due to a sub phase structure analyzing method, the strict conditions for each model are checked at the time of detailed classification concerning candidates due to structure matching (rough classification) or when the dedicated identification processing is decided to the set of models, that identification processing is performed.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to character recognition technology. More specifically, the present invention relates to matching processing in character recognition, and more particularly to a technique for performing detailed classification using a large classification result as a starting point.

Although the present invention can be applied to a drawing recognition apparatus and a document processing apparatus in general including a character recognition apparatus, the principle of the present invention can be applied to general pattern recognition including voice recognition and image recognition. .

[0003]

2. Description of the Related Art FIG. 1 shows an outline of a conventional identification method in character recognition. In the following description, it is assumed that the scalar amount that evaluates the entire pattern of one character is the distance,
Even the degree of similarity does not essentially change. The only difference is that the closest one is selected for the distance and the highest one is selected for the similarity.

In the most basic form, as shown in FIG. 1A, a candidate class matched by structure matching is subjected to distance calculation with a model, and the closest one is used as an identification result. Is the method. Here, there is no classification such as large classification and detailed classification, and there is only one identification process.

In this method, in an experiment conducted on all numerical data (156337 characters) possessed by the applicant, the correct reading rate is 91.8%, the rejection rate is 3.6%, and the misreading rate is 4.6%.
The result was obtained. It should be noted that "rejection" here means that it did not structurally match any model.

On the other hand, the present applicant previously filed Japanese Patent Application No. 1-
In the quasi-topological structure analysis method proposed in Japanese Patent Application No. 245507 or Japanese Patent Application No. 2-339696, some special processing is performed according to the candidates obtained by the structure matching, as shown in FIG. I was doing. For example, if it is a number, a special process is performed if it matches a certain model in '7', some are changed to match '1' and some are rejected. . Moreover, in order to reduce misreading, a threshold that determines the recess according to the distance was introduced, and it was a constant value regardless of the candidate model.

The advantage of the constant value is that the overall recognition performance can be easily controlled. However, as the recognition performance is improved in the development stage, more detailed control is required. Is a fact that cannot be ignored.
According to this method, the correct reading rate was 93.7%, the rejection rate was 5.5%, and the misreading rate was 0.8%.

[0008]

The detailed classification method of the conventional quasi-phase structure analysis method will be further examined. In this method, as shown in FIG. 2, distance calculation is performed for structurally matched classes, and then those that do not satisfy the constraints for each class due to special processing are excluded from candidates (example in FIG. 2). Candidate class 5 of 2
b) or another candidate is generated (in the example 1 of FIG.
It can be said that 1'is also added to the candidates), and the remaining one leaves one optimal candidate according to the previous distance (one of the candidate classes 5a and 9b in Example 2 of FIG. 2). Here, when no candidate is obtained as in Example 3, no consideration is given.

Here, the purpose of increasing correct reading and the purpose of reducing erroneous reading have not been clearly separated. If the target is only a specific character type, such a form is sufficient, but it is considered to be a problem when the expansion of the target character type is considered. Because reducing misreading is mostly a very narrow question of whether or not it should really match the model listed as a candidate, while increasing correct reading is This is because not only the model but also the entire model set depending on the character type to be read.

The problem to be solved by the present invention is mainly concerned with the improvement of the recognition performance among the three basic problems in the character recognition, namely the improvement of the recognition performance, the improvement of the recognition speed, and the reduction of the storage capacity. is there. The object of the present invention is to achieve this improvement in recognition performance under the condition that the target character type can be expanded.
Further, the present invention includes a device for improving the recognition speed.

[0011]

[Means for Solving the Problems] What kind of viewpoint is needed when the recognition performance is improved in consideration of the expansion of the target character type? With the dual purpose of increasing correct reading and reducing false positives,
In addition, knowing the relationship between each purpose and model set, you can always use it with a certain target character type (example: number), you can always use it with another target character type (example: English character), both It is considered to be "to increase the independence of the function according to the purpose" such as what can be used when mixed. Furthermore, ideally, it is ideal if it can meet individual requests such as "I want to read this character and only this character."

First, terms necessary for describing the present invention will be described. In the present invention, four viewpoints are introduced in order to classify / understand the result of large classification. These are introduced from the viewpoint of importance in detailed classification.

The viewpoints are the following four.

(1) Pattern (2) Class (model) (3) Class set (model set) (4) Category Of course, from another viewpoint, if it is a quasi-topological structure analysis method, for example, it is a primitive, It is also possible to consider primitive sequences or singular points. It is possible to consider the viewpoint of a combination of a specific primitive sequence and a singular point even in individual models. Here, we have introduced the above four viewpoints as the most basic ones.

The individual viewpoints will be described below.

[0016] A "pattern" is embodied through a certain form. Specifically, here
It is a character representation by an image array.

A "model" is an item in a dictionary, and is considered as a representative that summarizes the tendency of the appearance of characters.

On the other hand, a "class" is considered as a set of patterns that match the model. It should be noted that the model here is represented by several primitive sequences and some singular points, and there are several selection possibilities for each. From the point of view of large classification, the pattern is considered to belong to the class if at least the model is one of the candidates.

The "model set" is literally a combination of models, and means a combination of models that can be matched at the same time by a major classification called structural matching.

A set of patterns corresponding to a model set is called a "class set". This is a very important point of view when considering detailed identification of similar characters. In some cases, there is only one candidate at the same time at the time of major classification, and this is also considered as one model set.

The "category" is the character itself, or the character itself with the transformation removed, and is the most abstract. In general, categories often do not correspond to one group in the pattern space. Therefore, it is usual to prepare multiple models and cover the categories with their classes, which is also an example here. FIG. 3 illustrates the concept of categories and classes.

For example, the numeral "4" may have an open top or a closed top. Although these are one category of "4", they are prepared separately as models. It is in good condition. In FIG.
The classes 4a, 4b, and 4c mean the classes corresponding to one model representing the character category "4". Similarly, the classes 9a and 9b mean the classes corresponding to one model representing the category "9".

The object of the present invention is to realize the recognition performance while considering the expandability of the target character type by embodying the policy of "increasing the independence of the function according to the purpose" as described above. Is to raise. This is embodied according to the invention as shown in FIG.

According to the present invention, first, a constraint condition is set for each of all classes (models). This constraint condition is a necessary condition as an allowable range for the class. That is, if the constraint condition (check standard) is violated, it is clearly a condition that the condition is outside the allowable range. This constraint condition can be set for each class (model), and therefore the independence of each class is extremely high. This is indicated by a special arrow symbol in FIG. 4, and one must be prepared for all classes.

According to the present invention, the candidates are narrowed down by checking the constraint conditions of the corresponding model in the detailed classification of the candidate class obtained by the structure matching (major classification).

According to the present invention, in the detailed classification, with respect to a class set (model set), an identification process for an individual class set is introduced "if necessary". The identification process for a class set is "For this class and this class set,
It is a separate process of “identifying in this way”. By "if necessary" is meant that this is not the case if the original separation is so good that it can be easily identified by the size of the distance. The identification processing for such class sets may be created for each class set, and is highly independent. With this class group identification processing, it is possible to narrow down candidate classes for structural matching (major classification).

[0027]

As described above, by separating the two concepts of the constraint condition for each class and the checking thereof and the class group identification processing into the detailed classification, it becomes easy to deal with the expansion of the target character type. Also, in terms of improving recognition performance,
For those that are very difficult to separate, it is extremely convenient because it is possible to individually create and replace an identification process that introduces new features. It is also possible to use local image features down to the level of the image, rather than at the level of skeletal or contour representation. This enables more accurate identification and improves the recognition performance.

[0028]

[Embodiment] Next, each of the constraint conditions for each class and the class group identification processing will be described more specifically.

First, the constraint conditions for each class (model) will be described. The constraint conditions for each class are considered here in two stages. The first row is the condition related to the entire character of the pattern, and the second row is the local condition.

The "condition relating to the whole" in the first step is a condition relating to the distance between the pattern and the model. This was introduced as a threshold.

Conventionally, a fixed threshold value which does not depend on the model (class) is adopted, which is very convenient for controlling the capacity of the entire system. But,
It is rather an obstacle to improving the recognition performance. Because not all classes are evenly distributed in the pattern space, it is quite possible that one class is widespread and another is only allowed to exist in a very narrow range. Is. Also, it is not easy to select a certain threshold
It is also confirmed in the actual analysis results. Therefore, in order to improve the recognition performance, it is necessary to prepare a threshold related to distance for each class. This is adopted as the first step in order to suppress an increase in processing amount,
Patterns that are sufficiently far from the model can be removed without looking at local features, and patterns that are sufficiently close to the model are candidates as candidates without looking at local features. It is based on the idea that it will be acceptable. On the other hand, the second-stage "local thing"
This is the condition for individual primitive sequences and singularities. The characteristic parameters of the primitive series include the coordinates of both end points, the center of gravity of the circumscribing rectangle, the width, and the height.
Further, there is a coordinate as a characteristic parameter of the singular point. In addition to these feature parameters, we introduced a parameter that represents the degree of unevenness of the primitive sequence, and set a limit range for each scalar quantity. It is also possible to set a limited range using these combinations as vector quantities.

Although the above is an example of implementation by the quasi-phase structure analysis method, it is possible to similarly set a limit range in the feature vector in the ordinary pattern matching method. Further, if necessary, it is possible to re-extract the feature at the image level and set a limited range also for this feature. By doing so, the recognition performance can be improved.

Next, the class (model) set identification processing will be described. By introducing a constraint for each class, candidates that are not actually needed are canceled out of matching, so the number of class sets is reduced.

According to the experiment by the inventor, in handwritten numeral recognition (the number of models is 68) by the quasi-topological structure analysis method, there are 106 kinds of class sets (the number of models is 0: 1) and the number of models at the time when the major classification is completed. 1:68 types, number of models 2: 3
(1 type, 3 models, 6 types, 4 or more models: none)
However, by introducing the constraint conditions for each class described above, there are 82 types of class sets (0: 1 types of models, 1:68 types of models, 2:13 types of models, 3 types of models). Above: None).

Of these, 13 types in which the number of classes is 2 or more require class identification processing. Of these, 5 types can be correctly identified depending on the size of the distance and are special. Only eight types required identification processing.

For each of these eight types, the identification conditions were individually determined based on the distribution of the distances for each model. Basically, by changing the weight to the distance for each model, most correct identification was possible.

According to the experiment, the correct reading rate was 94.9%, the rejection rate was 4.7%, and the misreading rate was 0.4%. It is considered that the recognition performance will be further improved by introducing local features and new features at the image level.

FIG. 5 shows an example of a device configuration for carrying out the present invention described above. In this example, the present invention is applied to character recognition by the quasi-phase structure analysis method.

In FIG. 5, an image such as a handwritten character is read by the image input unit 100 and input as binary image data. The feature extraction unit 102 extracts the skeleton line of the binary image, analyzes the curved line structure, and extracts the global structure of the character shape by the quasi-topological feature and the singular point.
This global structure expresses each stroke component that constitutes a character skeleton line by a chain (primitive series) of monotonous curve constituent elements (primitives), and the combination of the series elements and the connection relation between the series elements, and Characteristic features of a character are described by binary relations between series elements, such as connection relationships of series elements at singular points. For details of such a global structure, for example, Japanese Patent Application No. 1-245507 and Japanese Patent Application No. 2-245507 previously filed by the present applicant.
No. 339696, so further explanation is omitted.

The feature extraction unit 102 further calculates the parameter vector of the local feature of the character shape based on the global structure expression. The parameters to be obtained include, for example, the coordinates of both end points of the primitive series, the center of gravity and the width and height of the circumscribing rectangle, the degree of unevenness, and the coordinates of the singular point. Regarding this, for example, Japanese Patent Application No. 2-
Since I am familiar with No. 339696, I will omit further explanation.

Reference numeral 110 is a dictionary section for character recognition,
As in the conventional case, the global structure description of the character shape and the local parameter vector which are obtained in advance are registered therein. Further, according to the present invention, the dictionary unit 110 also performs a constraint condition determined for each model (class) regarding each character of the recognition target character type as described above, and a process of identifying a necessary model (class) set. It is registered.

In the structure matching unit (major classification unit) 104, the global structure of the character shape input from the feature extraction unit 102 and the global structure registered in the dictionary unit 110 are matched and the distances are close ( Or high similarity)
Find a candidate class.

Reference numeral 106 denotes a detailed classification section (constraint condition check / model group recognition processing section) introduced by the present invention. Here, the candidates obtained by the structure matching (major classification) are registered in the dictionary section 110. The constraint condition is checked for each corresponding model (class), and if necessary, identification processing of the corresponding model (class) set registered in the dictionary unit 110 is performed to narrow down the candidates. (Detailed classification).

The local feature matching unit 108 has a dictionary unit 110 for the candidates narrowed down by the detailed classification unit 108.
The final recognition result is obtained by matching with the character shape parameter vector registered in. It is also possible to omit the local matching unit 108 by setting the result of the detailed classification as the final recognition result.

With respect to the detailed classification section (constraint condition check / model set identification processing section) 106 and the restraint conditions, various aspects as described below can be adopted.

(1) Only the constraint conditions determined for each model are checked. As one of the constraint conditions, one scalar quantity such as the distance or the similarity of the feature (global feature) for the large classification is used.

(2) Only the constraint conditions determined for each model are checked. As one of the constraint conditions, a limit range is set for each feature element (scalar amount) or combination (vector amount) of features for large classification.

(3) Only the constraint conditions determined for each model are checked. As one of the constraint conditions, not only features for large classification but also predetermined special features (scalar amount or vector amount) are used.

(4) Only the constraint conditions determined for each model are checked. Of the constraint conditions for each model, one scalar quantity such as the distance or similarity of features for large classification is used as the first condition, and if this is within a predetermined range, the model is examined without examining other features. Cancel the alignment with.

(5) Only the constraint conditions determined for each model are checked. Of the constraint conditions for each model, one scalar quantity such as the distance or similarity of features for large classification is set as the first condition, and if this is within the predetermined range, other features will not be examined. Allows matching with the model.

(6) The constraint condition determined for each model is checked and the model (class) set is identified. When the identification process is not defined for the model set, the identification is performed by one scalar amount such as distance or similarity.

(7) The constraint condition determined for each model is not checked, but only the model group identification process is performed.
When the identification process of the corresponding model set is not defined, identification is performed by one scalar amount such as distance or similarity.

[0053]

As can be understood from the above description, according to the present invention, the following effects can be obtained.

According to the first to sixth aspects of the present invention, the recognition performance can be improved by having the constraint condition for each model.

Since the invention of claim 2 uses one scalar quantity such as general distance and similarity as one of the constraint conditions, it can be used for most character recognition techniques.

According to the third aspect of the present invention, as one of the constraint conditions, a limit range is set for each element (scalar amount or combination thereof (vector amount)) of a feature for large classification.
This is an amount for evaluating the entire pattern of one character to which a scalar amount such as distance or similarity in the invention of claim 2 is given, whereas it limits a local feature, and is therefore more detailed. Since the control becomes possible, the recognition performance can be improved.

According to the fourth aspect of the invention, not only the features for the large classification but also the specific special features (scalar amount or vector amount) that are set in advance are used as the constraint conditions. It is equivalent to introducing features according to the model other than the features specified for the model,
Therefore, flexible processing is possible, and recognition performance can be improved.

According to the fifth aspect of the invention, the scalar quantity for evaluating the entire pattern of a given character such as the distance or the similarity is first checked in the constraint condition for each model, and if it is not within a certain range, the other is obtained. It is possible to cancel the matching without evaluating the individual characteristics of each, and by doing so, the processing time can be shortened.

Contrary to the invention of claim 5, in contrast to the invention of claim 5, the amount of scalar for evaluating the entire pattern of one character given the distance, the degree of similarity and the like is first checked in the constraint condition for each model. However, if this is within a certain range, it is possible to allow matching without evaluating other individual characteristics, and by doing so, the processing time can be shortened.

According to the invention of claim 7, it is possible to set a dedicated identification process for each set of models that can be matched at the same time. When the dedicated identification process is not set,
By adopting a configuration in which identification is performed using a single scalar quantity such as distance or similarity, it is possible to improve the identification performance of a set of models that are difficult to identify, and avoid the introduction of unnecessary identification processing. Can be suppressed.

According to the eighth aspect of the present invention, it is possible to reduce the number of types of model sets by defining a dedicated identification process for each set of models that can be matched at the same time and providing a constraint condition for each model. Therefore, it is possible to suppress an increase in the storage capacity of the dictionary.

[Brief description of drawings]

FIG. 1A is an explanatory diagram of a most basic character recognition method. (B) It is explanatory drawing of the conventional quasi-phase structure analysis method.

FIG. 2 is an explanatory diagram of an identification method in a conventional quasi-phase structure analysis method.

FIG. 3 is an explanatory diagram of a relationship between a pattern space and classes / categories.

FIG. 4 is an explanatory diagram of a detailed classification method according to the present invention.

FIG. 5 is a block diagram showing an example of a device configuration for implementing the present invention.

[Explanation of symbols]

100 image input unit 102 feature extraction unit 104 structure matching unit (major classification unit) 106 detailed classification unit (constraint condition check / model group identification processing unit) 108 local feature matching unit 110 dictionary unit

Claims (8)

[Claims] 1. A character recognition method in which one or more models are defined for each character of a character type to be read, and a character is recognized by matching with these models. A character recognition method characterized in that a constraint condition is checked in a detailed classification of a candidate having a condition and obtained by a large classification. 2. The character recognition method according to claim 1, wherein one of the constraint conditions is one scalar quantity such as a distance or similarity of features for large classification. 3. The character recognition method according to claim 1, wherein, as one of the constraint conditions, a limit range is set for each scalar amount which is an element of a feature for large classification or for each vector amount which is a combination thereof. 4. The character according to claim 1, wherein, as one of the constraint conditions, not only a feature for a large classification, but also a scalar amount or a vector amount, which is a predetermined special feature, is used. Recognition method. 5. Among the constraint conditions for each model, one scalar quantity such as the distance or similarity of features for large classification is set as the first condition, and if this is within a predetermined range, another The character recognition method according to claim 1, wherein the matching with the model is canceled without checking the characteristics. 6. Among the constraint conditions for each model, one scalar quantity such as the distance or similarity of features for large classification is set as a first condition, and if this is within a predetermined range, another The character recognition method according to claim 1, wherein matching with the model is allowed without checking the characteristics. 7. One or more models are defined for each character of the character type to be read, and in a character recognition method for recognizing a character by matching with these models, the character classification methods are classified into two groups. Dedicated identification processing is established as needed for each set of models that can be matched, and in the detailed classification of candidates obtained by large classification, if dedicated identification processing is defined, it is identified by that identification processing The character recognition method is characterized in that when the identification processing is not specified, the identification is performed by one scalar amount such as distance or similarity. 8. A character recognition method in which one or more models are defined for each character of a character type to be read, and a character recognition method for recognizing a character by matching these models restricts each model. In addition to having conditions, each set of models that can be matched at the same time has a dedicated identification process as needed, and in the detailed classification of the candidates obtained by the large classification, the constraint conditions of the corresponding model are checked and When the identification processing is defined, the identification processing is performed, and when the dedicated identification processing is not provided, the identification is performed by one scalar amount such as distance or similarity. Character recognition method described.