JPH06259565A - System for preparing pattern recognition dictonary - Google Patents

Abstract

PURPOSE:To provide a system for reconstituting a pattern recognition dictionary having a high recognition ratio. CONSTITUTION:An error pattern sorting part 5 sorts a learning pattern read out in error into a pair of a correct side category group and a misread category group. A reference vector preparing part 6 finds out, the mean vector of patterns in each group and newly prepares a reference vecotor for providing a vector element close to a correct category side reference vector with the feature value of the correct category reference vector and providing a vector element close to a misread category side reference vector with hte veature value of the mean vector. The reference vector expresses the specific deformation of each misread learning pattern. Thereby a reference vector adding part 7 adds the reference vector to an existing pattern recognition dictionary.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a dictionary creation system for pattern recognition of character patterns, voice patterns, etc.
In particular, the present invention relates to a pattern recognition dictionary creation system used to reconstruct an existing pattern recognition dictionary (pattern recognition dictionary) to improve accuracy.

[0002]

2. Description of the Related Art When a character, a figure, or a voice is input to a computer as information for various kinds of control or processing, it is necessary to mechanically recognize the character or the like. Therefore, in general, a character or the like is expressed by a vector (hereinafter referred to as “feature vector”) composed of a sequence of a plurality of feature values to be patterned, and pattern recognition based on an input pattern such as the character pattern is performed. A method of discriminating input characters and the like is adopted.

In the above pattern recognition, standard patterns of each category (for example, character type in the case of character pattern recognition) are prepared in the pattern recognition dictionary as reference vectors. Then, the similarity between the input pattern to be recognized and the reference vector of each category in the pattern recognition dictionary is calculated using a predetermined evaluation function,
Based on this calculation result, the category with the highest similarity is output as the recognition result of the input pattern.

Further, since there are various modes of characters, figures, or voices even in the same category, when creating the above reference vector, first, the average of several different patterns is calculated for each category, and this is calculated. Prepare in the pattern recognition dictionary. And by learning by a worker involved in character recognition or so-called expert system,
This pattern recognition dictionary is reconfigured to improve accuracy and the recognition rate of input patterns by the pattern recognition dictionary is increased.
As a technique for reconstructing such a pattern recognition dictionary, conventionally, the following methods have been proposed.

(1) In the first method, an input pattern for learning (hereinafter referred to as "learning pattern") is recognized by an existing pattern recognition dictionary. Then, the pattern recognition dictionary is reconstructed by repeatedly performing the process of optimizing the position of the reference vector so that the misreading (misrecognition) of the input pattern is minimized. As a method for optimizing the reference vector, for example, a method adopting a simulated annealing method applied in LSI layout design or the like has been proposed (Matsunaga, Abe, Kida: “The simulated annealing method is used. Optimized character recognition dictionary ", IEEJ, PRU90-39 (1990)).

(2) The second method is the same as the above method in that the learning pattern is recognized by the existing pattern recognition dictionary, but the input pattern that cannot be correctly recognized is used as the reference vector of the corresponding category. The pattern recognition dictionary is reconstructed by newly adding it to the recognition dictionary.

(3) The third method is similar to the above second method, but one reference vector is added to one misread input pattern in order to reduce the size of the pattern recognition dictionary as much as possible. Instead of providing it in the recognition dictionary, the average of multiple misread input patterns is added to the pattern recognition dictionary as a reference vector of the corresponding category (Ide, Wakahara: "A Study of Subcategory Generation Method for Character Recognition", 1990). IEICE Autumn National Conference Proceedings, D-348, p. 6-350 (199)
0)).

[0008]

However, the above-mentioned prior art has the following problems to be solved.

First, the first method is excellent in that the pattern recognition dictionary is optimized so that the recognition accuracy becomes high in consideration of the difference in pattern distribution between categories. However, when the pattern distributions are complicated and the pattern distributions are intermingled between the categories, or when the pattern distributions are not continuous but separated (dispersed), one reference vector cannot handle it. Therefore, in the case of such a pattern, it is difficult to improve the accuracy of the pattern recognition dictionary, and conversely the recognition accuracy is deteriorated.

Further, in the second method, since a plurality of reference vectors can be provided for each category, the problem of the first method can be solved. However, in this method, since the created reference vector corresponds to the input pattern itself at the time of learning, the versatility for pattern recognition of the category is poor. Further, since the learning pattern is used as the reference vector, if the learning pattern is similar to other categories, the input pattern may be misread as a pattern of the other category.

Further, in the third method, by using the average of a plurality of patterns as the reference vector, it is possible to ensure versatility in pattern recognition. However, this method has a drawback that it is difficult to secure versatility in the reference vector when the number of input patterns that are misread during learning is small, and the effect is not obtained unless there are many input patterns that are misread.

The present invention has been made in view of the above problems, and an object thereof is to make it possible to create a highly versatile reference vector even when the number of input patterns that are misread during learning is small, and to improve the recognition rate. It is to provide a pattern recognition dictionary creating system that can efficiently create a high pattern recognition dictionary.

[0013]

The pattern recognition dictionary creating system of the present invention for achieving the above object is a system for reconstructing an existing pattern recognition dictionary composed of a set of reference vectors to be collated at the time of recognition of an input pattern. A dictionary setting means for setting the pattern recognition dictionary as an initial dictionary, a feature extracting means for extracting a feature from the input pattern, and a pattern recognizing means for recognizing a category of a feature vector of the input pattern using the initial dictionary. And a pattern extracting means for extracting misreads from the input patterns recognized by the pattern recognition means, and the extracted misreading input patterns are classified and grouped for each combination of the correct answer side category and the misreading destination category. Reference vector creation that creates new reference vectors based on different values for each group and each classification means Pattern recognition with a high generalization ability by creating a reference vector expressing a modification peculiar to a misreading pattern, and having a means and a reference vector adding means for adding the newly created reference vector to the pattern recognition dictionary. It was configured to get a dictionary.

In the above structure, the reference vector creating means obtains an average vector consisting of an average value sequence of each feature value of the input pattern groups included in the same group, and the average vector is the correct category reference vector and the misread category reference. It is determined which of the vectors is closer to each feature value, and the reference value is created by adopting the average value for the feature values close to the misread category and the feature values of the correct category reference vector for the other values. It comprises vector decision means.

[0015]

In the pattern recognition dictionary creating system of the present invention having the above-described structure, first, misread input patterns are grouped by the combination of the correct answer side category and the misread destination category, so that those having similar misreading tendencies are grouped. Collect. Then, in the average vector obtained from the input patterns included in the same group, the feature value close to the reference vector of the correct category is the value of the correct category reference vector,
On the other hand, for the feature values close to the reference vector of the misreading destination category, by adopting the average value of the average vector, respectively, a highly versatile reference vector is created in which only the common deformation of the misreading input patterns is expressed differently. To be done. Since this newly created reference vector is effective even if the number of misread input patterns is small, adding this reference vector to the pattern recognition dictionary improves the recognition rate in the pattern recognition dictionary.

[0016]

Embodiments of the present invention will be described below with reference to the drawings.

FIG. 1 is a block diagram of a pattern recognition dictionary creating system according to an embodiment of the present invention. In the figure, solid arrows indicate the flow of processing in this system, and broken arrows indicate the flow of data.

Referring to FIG. 1, the pattern recognition dictionary creating system according to the present embodiment has an initial dictionary setting unit 1, a feature extraction unit 2, a pattern recognition unit 3, an error pattern extraction unit 4, an error pattern classification unit 5, and a reference vector. It includes a creation unit 6, a reference vector addition unit 7, and a recognition dictionary storage unit 8.

The initial dictionary setting unit 1 sets an existing pattern recognition dictionary as an initial dictionary (dictionary setting means). In this pattern recognition dictionary, standard patterns for each category are prepared as reference vectors. The feature extraction unit 2 performs feature extraction of the input learning pattern (feature extraction means). In the extraction process, for example, when the pattern recognition target is a character, a character or a character string to be learned is binarized and input by a reading unit such as a scanner or a character slicing device, and these input characters are input. Is represented by a feature vector composed of an array of feature values, which is a plurality of vector elements, and is used as a learning pattern.

In the pattern recognition unit 3, the initial dictionary setting unit 1
The learning pattern is recognized using the initial dictionary set to (pattern recognition means). This recognition processing is performed by comparing the reference vector extracted from the initial dictionary and the feature vector of the learning pattern with the array of feature values forming these vectors to obtain the one with a high degree of similarity. Recognizes a particular category of learning patterns. The error pattern extraction unit 4 extracts a learning pattern that was misread by the pattern recognition unit 3 without being correctly recognized (pattern extraction means). Such an extraction process
Similar to the conventional method, it is performed by an operator who is involved in character recognition, for example.

The error pattern classifying unit 5 classifies the extracted misreading learning patterns into groups by classifying them into combinations of correct categories, which are correct categories, and misreading destination categories, which are misreading categories. (Classification means). The reference vector creation unit 6 creates a reference vector for each of these classified groups (reference vector creation means). This reference vector will be described later. This newly created reference vector is used by the reference vector adding unit 7
From the recognition dictionary storage unit 8 and is added to the initial dictionary in the recognition dictionary storage unit 8 (reference vector adding means).

Next, the processing procedure of the reconstruction work of the pattern recognition dictionary according to this embodiment will be concretely described. FIG. 2 is a flowchart showing the above processing procedure, and S represents each step. Referring to FIG. 2, first, the initial dictionary setting unit 1 stores the initial dictionary in the recognition dictionary storage unit 8 (S101).
Next, the feature extraction unit 2 extracts the features of the learning pattern (S102), and the pattern recognition unit 3 recognizes the learning pattern using the initial dictionary (S103). The error pattern extracting unit 4 extracts misread patterns from the recognized learning patterns (S104), and then the error pattern classifying unit 5 extracts the misread learning patterns from the correct answer side category and the misread destination category. Are grouped into a plurality of groups Ge (e = 1, 2, ... E) (S105). By this processing, for example, G1 is "A"
Group of learning patterns misread by "King" among the learning patterns of G, and G2 is "bunch" among the learning patterns of "East"
It becomes a group of learning patterns misread by.

The above-mentioned classified groups are average vectors m for each group Ge in the reference vector creating section 6.
_Ge (i) (i = 1, ..., N) is created (S10).
7). Next, in the reference vector creating unit 6, the created average vector is either the reference vector of the correct answer side category (for example, “A” in the case of group G1) or the reference vector of the misreading destination category (the same, “King”). Is determined for each vector element, that is, for each feature value (S10).
8).

As a result of this determination, for the feature value close to the misread category, the average vector m _Ge (i) created above is used.
The average value of (i = 1, ..., N) is (S109), and for the feature values close to the correct category other than that, the feature value of the reference vector mc (i) (i = 1, ..., N) of the correct category. (S110) are adopted respectively. By performing this process for all feature values (S111), the group Ge
New reference vector m ′ _Ge (i) (i = 1, ..., N)
To create. The new reference vector m'created in this way
_Ge (i) (i = 1, ..., N) is added to the reference vector adding unit 7
Is added to the initial dictionary (S112) and stored in the recognition storage unit 8. This process is applied to all groups G1, G2, ...
GE is performed (S106).

In this way, new reference vectors expressing transformations peculiar to the respective learning patterns are sequentially added to the pattern recognition dictionary, and when pattern distributions are intermixed between categories or when they are not continuous but separated. In this case, the deterioration of the recognition accuracy is suppressed. In addition, generalization ability is improved, and it is not necessary to draw in patterns of other categories.
Moreover, since a reference vector having high versatility can be created even if the number of input patterns misread during learning is small, it is possible to efficiently create a pattern recognition dictionary having a high recognition rate. The present invention is not limited to this embodiment, and the configuration of FIG. 1 or the processing procedure of FIG. 2 can be changed without departing from the scope of the invention.

[0026]

As described in detail above, in the pattern recognition dictionary creation system of the present invention, the misreading learning patterns are grouped by the correct answer side and the misreading destination pair, and
For the pattern groups included in each group, an average vector composed of a column of average values of feature values is obtained, and it is determined for each feature value whether the average vector is closer to the correct category reference vector or the misread category reference vector. , The average value for the vector element close to the misreading destination, and the feature value of the correct category reference vector for the other cases to form a new reference vector, and add this new reference vector to the existing pattern recognition dictionary. I decided to do so,
It becomes possible to form a reference vector that represents a transformation peculiar to each error learning pattern, and a recognition dictionary with high generalization ability can be obtained.

Since a reference vector having high versatility can be created even if the number of input patterns misread during learning is small, it is possible to efficiently create a pattern recognition dictionary with a high recognition rate and improve the accuracy of the existing pattern recognition dictionary. The pattern recognition dictionary for can be efficiently reconstructed.

[Brief description of drawings]

FIG. 1 is a configuration diagram of main parts of a pattern recognition dictionary creating system according to an embodiment of the present invention.

FIG. 2 is a flowchart showing a processing procedure for creating a pattern recognition dictionary in this embodiment.

[Explanation of symbols]

 1 initial dictionary setting unit (dictionary setting unit) 2 feature extraction unit (feature extraction unit) 3 pattern recognition unit (pattern recognition unit) 4 error pattern extraction unit (pattern extraction unit) 5 error pattern classification unit (classification unit) 6 reference vector Creation unit (reference vector creation unit) 7 Reference vector addition unit (reference vector addition unit) 8 Recognition dictionary storage unit

Claims (2)

[Claims] 1. A system for reconstructing an existing pattern recognition dictionary composed of a set of reference vectors to be collated at the time of recognizing an input pattern, the dictionary setting means for setting the pattern recognition dictionary as an initial dictionary, and the input. Feature extraction means for performing feature extraction from a pattern, pattern recognition means for recognizing the category of the feature vector of the input pattern using the initial dictionary, and extraction of misreads from the input patterns recognized by the pattern recognition means Pattern extracting means, classifying means for classifying the extracted misreading input patterns for each combination of the correct answer side category and the misreading destination category, and creating a new reference vector based on a different value for each group Reference vector creating means and the newly created reference vector are added to the pattern recognition dictionary. A pattern recognition dictionary creating system comprising: a reference vector adding means. 2. The reference vector creating means obtains an average vector consisting of an average value sequence of respective feature values of input pattern groups included in the same group, and the average vector is the correct category reference vector and the misread category reference vector. Which one is closer to each of the characteristic values is determined, and a reference vector is created by adopting the average value for the characteristic value close to the misreading destination category and the characteristic value of the correct category reference vector for other than that. The pattern recognition dictionary creating system according to claim 1, further comprising a vector determining means.