JPH06251158A - Pattern recognition dictionary generating method - Google Patents

Abstract

PURPOSE:To efficiently generate a pattern recognition dictionary whose performance is satisfactory by using a shielding means for shielding a specific part of the pattern recognition dictionary from other category in its periphery. CONSTITUTION:This method is constituted of a phase 1 for selecting a learning category and a phase 2 for sequential learning. The phase 1 consists of an initial dictionary generating part 01, a pattern recognizing part 02, a learning category set selecting part 03 and a shielding category set selecting part 04. The phase 2 consists of a learning pattern selecting part 05, a pattern recognizing part 06, a dictionary correction deciding part 07, a dictionary correction executing part 08, and an end deciding part 09. Also, in the recognition dictionary, only a necessary part of learning becomes an object of learning, and a necessary part of learning is shielded by a category having possibility of erroneous recognition, or a necessary part of learning is shielded by an artificial pattern, therefore, even if sequential learning is executed in a part of the dictionary, it is suppressed to the minimum that adverse influence is exerted on the whole dictionary.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for creating a dictionary used in a pattern recognition device such as an OCR (optical character reading device), and more particularly, to efficiently create a recognition dictionary with high recognition accuracy. Regarding how to create.

[0002]

2. Description of the Related Art Conventionally, when recognizing a pattern of characters or voices, a reference pattern which is a standard pattern in each category and this reference pattern according to the distribution shape of the category are used. A weight vector for correcting the positional relationship with the recognition target pattern is prepared in advance as a recognition dictionary, and the similarity between the pattern to be recognized and the reference pattern of each category is evaluated by some evaluation function (for example, , Euclidean square distance), and the category with the highest similarity is output as the recognition result. This reference pattern
Usually, it is given as a pattern average value to be learned, and the weight vector is given as an inverse number of standard deviation. Further, as a sequential learning method, a method of deciding the reference pattern and weight by correcting an error at the time of learning, such as a potential function method and LVQ (learning vector quantization), has been proposed. The literature on this type of conventional method is "Introduction to Character Recognition" by Shinichiro Hashimoto and Ohmsha (198
2), (This section describes commonly used character recognition methods, standard pattern creation methods using arithmetic mean, etc.), “Self-Organization and associative Memory
′ 3rd.edition 』T. Kohohonen, Springer-Verlag (198
9), (This describes the methodology LVQ (learning vector quantization) that moves standard patterns sequentially) and "Optimization of discriminating hyperplanes by dynamic model" Takeshi Sakano, Electronic Information Communication Academic Technical Report, PRU-91-149 (1992)
(This describes a method of sequentially moving and optimizing the weights for correcting the shape of the distribution).

[0003]

As described above, conventionally, when a pattern recognition dictionary having a reference pattern for each category and a weight for correcting the similarity measure according to the distribution of the category is created, The data constituting the dictionary is calculated and added by the sequential learning optimization method using many learning patterns. However, when the number of categories increases, it is difficult to use techniques such as the amount of calculation and the possibility of convergence for performing recognition experiments for all categories.
Even if categories are added, it is necessary to perform optimization on all categories, which is difficult in operation. Further, the conventional pattern recognition method has been used in a fairly wide range because the algorithm is simple and the effect can be expected in many cases. However,
In sequential function dictionary creation methods such as the potential function method and LVQ (learning vector quantization), it is necessary to perform learning for all categories used for recognition.
As described above, there are difficulties in the technique such as the amount of calculation and the possibility of convergence. That is, when the category used for recognition is, for example, a character, and the number thereof is large as in the JIS first level 3201 category, there are aspects that are not practical in terms of both the convergence possibility and the processing amount. Therefore, when the learning is sequentially performed using only a part of the categories, there is a problem that the performance of the entire dictionary is deteriorated because erroneous recognition increases in the categories that are not learned. An object of the present invention is to solve the conventional problems as described above and to provide a pattern recognition dictionary creating method capable of efficiently creating a high-performance pattern recognition dictionary by shielding the influence on the surroundings. To provide.

[0004]

In order to achieve the above object, the method for creating a pattern recognition dictionary of the present invention is used for (a) pattern recognition of characters, voices, etc. It consists of a standard reference pattern and a set of weights,
For each learning pattern, the arithmetic mean and standard deviation of the features are calculated for each category and each feature axis, and the arithmetic mean is referred to.
, An initial recognition dictionary with the reciprocal of the standard deviation as a weight is created in advance, and the reference pattern or the weight is changed so as to minimize misrecognition while repeatedly giving the learning pattern. In the method of creating a pattern recognition dictionary completed by doing so, the dictionary is partially improved in performance by using a shielding means that shields a specific part of the pattern recognition dictionary from other categories around it. It is characterized by that. Further, (b) as a shielding means, a pattern of a category having a high possibility of misreading a specific set of categories is used. Furthermore, (C)
As a shielding means, it is also characterized in that a pseudo pattern is generated in the periphery of a set of specific categories so as to shield it.

[0005]

In the present invention, first, for all categories used for recognition, reference patterns, which are the arithmetic mean of several patterns, and weights, which are the reciprocal of the standard deviation, are created for each category, and these are calculated. By registering as an initial dictionary and then recognizing the pattern used for learning by the initial dictionary, a category set that requires further sequential learning (called a learning category set) is selected, and then the selected category set is selected. Focusing on the recognition result of the category set, other than that category set, the category that is the target of misrecognition or the one whose evaluation function value used for recognition is close is selected as the category that shields the learning category set from other categories. Finally, iterative learning is performed using the selected learning category set and occlusion category set. Instead of the above process, paying attention to the recognition result of the selected category set, the pattern farthest from the center of the learning category set among the learning category sets in terms of the evaluation function used for recognition. The pseudo pattern that measures the distance R of the
It is also possible to use a process in which a large amount of the generated pattern is generated and a shield is formed by using the generated pattern (the pseudo pattern is called a shield pattern). As described above, in the present invention, the performance of only a part of the recognition dictionary is improved by using the method of creating a pattern recognition dictionary of the sequentially changing type which reduces erroneous recognition. That is, by shielding the periphery of a minority category to be optimized by a pseudo pattern, the minority category can be optimized without affecting the entire system. As a result, the optimizing operation can be performed only in the part having a particularly low recognition rate in the recognition dictionary, the newly added category, and the periphery thereof, and an efficient recognition dictionary can be created.

[0006]

Embodiments of the present invention will now be described in detail with reference to the drawings. FIG. 1 is a pattern diagram showing a first embodiment of the present invention.
FIG. 3 is a functional block diagram of a recognition dictionary creating system. Figure 1
As shown in FIG. 5, the dictionary creating method of the present invention comprises Phase 1 of learning category selection and Phase 2 of sequential learning. First, Phase 1 (learning category selection unit) includes an initial dictionary creation unit 01, a pattern recognition unit 02, a learning category set selection unit 03, and a masked category set selection unit 04. Select target category set,
And each process of selection of an occlusion category set is performed. Fee
Step 2 (sequential learning section) comprises a learning pattern selection section 05, a pattern recognition section 06, a dictionary correction determination section 07, a dictionary correction execution section 08, and an end determination section 09, and is selected by Phase 1. Recognition is performed for each pattern that constitutes the learning category set and the masked category set, and a dictionary is created by an arbitrary sequential learning algorithm so that erroneous recognition is reduced in both category sets. The processing operation of each block will be described below.

First, the initial dictionary creating section 01 of Phase 1 finds a reference pattern as an arithmetic mean for each category,
Weights are obtained as the reciprocal of the standard deviation, and these are used as the initial dictionary. Next, the pattern recognition unit 02 uses the initial recognition dictionary created by the initial dictionary creation unit 01 to recognize the pattern used for learning. Next, the learning category set selection unit 03 selects a category having a particularly low recognition rate from the learning patterns recognized by the pattern recognition unit 02. Then, a learning category set is constructed by using the category in which the category is often misrecognized. Next, the occlusion category set selection unit 04 selects each pattern forming the learning category set.
The recognition result is checked for all the registered information, and all the categories that are erroneously recognized are registered as a masked category set. In Phase 2, a dictionary is created by an arbitrary sequential learning algorithm so that the misrecognition rate is reduced. At that time, the selected algorithm targets only the categories belonging to the learning category set for learning and masks them. Do not change the categories belonging to the category set.

The learning pattern selecting section 05 of Phase 2 selects only one pattern for learning the recognition dictionary, and then the pattern recognizing section 06 selects the selected pattern. To recognize. Next, the dictionary correction determination unit 07 determines whether to update the dictionary according to the conditions of the sequential update algorithm adopted based on the recognition result of the selected pattern. If the result of determination is that the conditions are satisfied, the dictionary correction execution unit 08 updates the dictionary. However,
If the selected pattern is a pattern belonging to the masking category set registered by the masking category set selecting unit 04, the learning pattern selecting unit 05 does not update the dictionary.
Return to and select a new learning pattern. Further, the dictionary correction execution unit 08 updates the dictionary according to the dictionary change rule of the adopted sequential update algorithm. Next, the end determination unit 09 determines the end of learning according to the end determination condition of the adopted sequential update algorithm. If it is determined that the learning has not ended, the process returns to the learning pattern selection unit 05 to select a new learning pattern and continue the learning.

Next, a specific processing operation of the present invention will be described. Here, T. An example using LVQ (Learning Vector Quantization) proposed by Kohonen will be described. However, the present invention is not limited to this, and other sequential learning algorithms are used for the purpose of simplifying the description. Of course, it is okay. The LVQ method is a method used to create a multi-template dictionary in pattern recognition, and the learning pattern is recognized after the initial dictionary is created by the clustering method, but it cannot be recognized correctly. If the learning pattern exists near the discrimination boundary between the reference pattern on the correct reading side and the reference pattern on the misreading side, the most distance is found among the reference patterns of the category that should be recognized correctly. The position of the recognition dictionary is optimized by repeating the process of bringing the reference pattern which is close to the reference pattern closer to the reference pattern and moving the reference pattern of the incorrectly recognized category away from the reference pattern. A specific example of the process of moving the reference pattern will be described by the following equations (1) and (2). In addition, this formula (1) (2)
Is a one-dimensional movement formula of the feature vector expressing the pattern, it is necessary to execute this movement formula for each feature axis. [Correct reading side] m (t) = m (t-1) + α (t) (x−m (t-1)) ... (1) [Error reading side] m (t) = m (t-1) -α (t) (x-m (t-1)) (2) where m (t) is the reference pattern at time t, m (t) t-
1) is a reference pattern at time t-1, x is a learning pattern α (t) is a decreasing function (0 <α <1).

Under the above conditions, the operation flow of the present invention will be described. FIG. 2 is an operation flowchart of the pattern recognition dictionary creating system shown in FIG. First, in Phase 1, for all categories used for recognition in the initial dictionary creation unit 01, the reference pattern, which is the arithmetic mean of several patterns for each category, and the reciprocal of the standard deviation. A recognition dictionary M0 having one weight vector is created (step 101). Next, the pattern recognition unit 02 recognizes all learning patterns using this dictionary M0 (step 102). Next, the learning category set selection unit 0
In step 3, the recognition result is analyzed, and a category having a poor recognition rate, in particular, a category misreading each other is selected as a learning category set (step 103). Next, the occlusion category set selection unit 04 analyzes the recognition results for all the patterns belonging to the learning category, and selects, as the occlusion category set, all misreading categories that do not belong to the learning category (step 104). . The dictionaries of the selected learning category set and occlusion category set are registered as the initial recognition dictionary M (step 105).

Next, in Phase 2, the learning pattern
The pattern selection unit 05 selects only one learning pattern (step 106), and the pattern recognition unit 06 recognizes this learning pattern using the recognition dictionary M (step 107). Next, the dictionary correction determining unit 07 determines whether the recognition result is a misread or a correct reading (step 108). If it is a misread, the learning pattern is learned for each of the reference pattern of the correct category whose distance from the learning pattern is the closest and each reference pattern which is closer than the distance value. It is determined whether it belongs to the category set or the occlusion category set (step 109). If the reference pattern belongs to the learning category set, the reference pattern is the target of movement. That is, as a result of the determination as to whether or not to update the dictionary, if the condition is satisfied, the dictionary correction execution unit 08 performs the pattern.
(Step 110). Further, the end determination unit 09 determines whether or not the end condition is satisfied (step 111). If the condition is satisfied, the learning ends, the recognition dictionary at that time is registered, and the process ends.

FIG. 3 is a functional block diagram of a pattern recognition dictionary creating system showing a second embodiment of the present invention, and FIG. 4 is a processing flowchart by the system of FIG. The functional block of FIG. 3 is different from the functional block configuration of FIG. 1 in that a masking pattern generation unit 10 is provided in place of the masking category set selection unit 04, and other configurations are the same. Exactly the same. The occluding category set selecting unit 04 in FIG. 1 occludes a portion that needs learning in a category that may be erroneously recognized.
In the shielding pattern generation unit 10 of (1), the pseudo pattern blocks the portion that needs learning. In any case, by shielding, only a part of the recognition dictionary can be improved in performance. In FIG. 4, first, the initial dictionary creating unit 01 of the phase 1 refers to the reference pattern which is an arithmetic mean of several patterns and the standard deviation for all categories used for recognition. A recognition dictionary M0 having one weight vector which is the reciprocal of 1 is created (step 201). Next, the pattern recognition unit 02 recognizes all learning patterns using this dictionary (step 202). Next, in the learning category set selection unit 03,
The recognition result is analyzed to select a category having a poor recognition rate, particularly a category misreading each other as a learning category set (step 203). Next, the occlusion category set selection unit 04 analyzes the recognition results for all the patterns belonging to the learning category, and forms occlusion patterns at the same distance as the pattern farthest from the center of the learning category set. (Step 204). The dictionary of the selected learning category set is registered as the initial recognition dictionary M (step 2
05). These two processes (steps 204 and 205) are different from the flow of FIG.

Next, in Phase 2, a learning pattern
The learning pattern selection unit 05 selects one learning pattern (step 206), and the pattern recognition unit 06 recognizes this learning pattern using the recognition dictionary M (step 20).
7). Next, the dictionary correction determination unit 07 determines whether the recognition result is erroneous reading or correct reading (step 208), and if it is correct reading, the process proceeds to the end determination unit 09. If it is a misread, the learning pattern is set as the learning category set for each of the reference pattern of the correct category whose distance to the learning pattern is the closest and all the reference patterns which are closer than the distance value. It is determined whether or not it belongs to the shield pattern (step 209). If it belongs to the shield pattern, the learning pattern
Then, the process returns to the selection unit 05 and the learning pattern is selected again. If the reference pattern belongs to the learning category set, the reference pattern is the target of movement. That is, the dictionary correction execution unit 08 executes the movement of the reference pattern (step 210), and then the end determination unit 09.
At step 211, it is determined whether the learning pattern has been erroneously read or the set number of learnings has been reached (step 211). If the conditions are satisfied, the learning is terminated, and the recognition dictionary at that time is registered and processing is performed. finish.

[0014]

As described above, according to the present invention,
Only the parts that need to be learned in the recognition dictionary are targeted for learning, and the parts that need to be learned are blocked by categories that have the possibility of misrecognition, or the parts that need to be learned are blocked by pseudo patterns. Therefore, even if the sequential learning is performed in a part of the dictionary, it is possible to minimize the adverse effect on the entire dictionary. As a result, when designing a pattern recognition dictionary, adding categories, etc., a recognition dictionary having a high recognition rate can be created with a relatively small amount of processing and a simple algorithm.

[Brief description of drawings]

FIG. 1 is a functional block diagram of a pattern recognition dictionary creating system showing a first embodiment of the present invention.

2 is an operation flowchart of the method for creating a pattern recognition dictionary in FIG.

FIG. 3 is a functional block diagram of a pattern recognition dictionary creating system showing a second embodiment of the present invention.

4 is an operation flowchart of the method for creating a pattern recognition dictionary in FIG.

[Explanation of symbols]

 01 initial dictionary creation unit 02 pattern recognition unit 03 learning category set selection unit 04 masked category set selection unit 05 learning pattern selection unit 06 pattern recognition unit 07 dictionary correction determination unit 08 dictionary correction execution unit 09 end determination unit 10 Shielding pattern generator

Claims (3)

[Claims] 1. A standard reference pattern for each category, which is used for pattern recognition of characters, voices, etc., and a set of weights, and a plurality of learning patterns for each category. , The arithmetic mean and standard deviation of the features are calculated for each characteristic axis, the arithmetic mean is used as a reference pattern, and an initial recognition dictionary in which the reciprocal of the standard deviation is used as a weight is prepared in advance, and the learning pattern is prepared. In the method for creating a pattern recognition dictionary, which is completed by changing reference patterns or weights so as to minimize misrecognition while repeatedly giving
A pattern recognition dictionary creating method characterized in that the performance of the dictionary is partially improved by using a shielding means that shields a specific part of the computer recognition dictionary from other categories around it. 2. The pattern recognition dictionary creating method according to claim 1, wherein a pattern of a category having a high possibility of misreading a set of specific categories is used as the shielding means. A method for creating a pattern recognition dictionary. 3. The pattern recognition dictionary creating method according to claim 1, wherein the shielding means shields a set of specific categories by generating a pseudo pattern around the set. A method for creating a pattern recognition dictionary.