JPH06139410A - Template matching type character recognition system using neural network - Google Patents

Abstract

PURPOSE:To ensure a high recognition ratio by means of a feature value conversion neuro even though the learning of this conversion neuro is not complete by learning at a high speed the conversion neuro that converts a feature value vector into a standard one. CONSTITUTION:In a character recognition mode, a feature value vector 1 extracted from a read character pattern is converted into a standard feature value vector 3 where the feature of each category is reflected by a feature value conversion neuro 2. The template matching is carried out between the vector 3 and a template 4 and the characters are recognized. In a learning mode, a template 4 is produced from a feature vector 1' and the neuro 2 is learnt by using the data that converts the template 4 into the same one as a teacher signal. If the correct recognizing result cannot be obtained despite the learning of the neuro 2, the neuro 2 is learnt again by means of the relevant feature value vector.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a character recognition method applying template matching and a neural network, and in particular, the present invention can train a neural network at high speed, and even if learning is not completed, it is high. The present invention relates to a template matching type character recognition method by a neural network capable of obtaining a recognition rate.

[0002]

2. Description of the Related Art FIG. 7 shows a neural network which is a premise of the present invention.
It is a block diagram of a template matching type character recognition device by a network (hereinafter abbreviated as neuro). In the figure, 101 is a scanner for reading handwritten characters to be recognized, 102 is a normalizing means for normalizing the bit map data read by the scanner, and 103 is a feature of the normalized bit map data. A feature extraction means for extracting a quantity vector, 104 a feature quantity conversion neuro for converting a feature quantity vector, 105 a template storage means for storing a template that is a feature vector for each category, and 106 a feature quantity conversion neuro 104 It is a collation unit that collates the feature amount vector with the template stored in the template storage unit 105 and outputs the recognition result.

FIG. 8 (a) is a diagram showing an example of the configuration of the feature amount conversion neuro 104 in FIG. 7, and FIG. 8 (b) is a diagram showing the configuration of the element neurons of the feature amount conversion neuro 104. In FIG. 8A, 111 is an input layer, 1
Reference numeral 12 is an intermediate layer, 113 is an output layer, and each neuron of the intermediate layer 112 and the output layer 113 has weights w1, w2, respectively for input signals x1, x2, ... Xn, as shown in FIG.
It is composed of an adding unit 115a for multiplying by and adding and a function unit 115b for converting the output y of the adding unit 115a.

Input layer 101 of feature quantity conversion neuro 104
When a feature amount vector as shown in FIG. 8A is input to each neuron of FIG. 8, the feature amount vector is input to the intermediate layer 112 via the input layer 111, and as shown in FIG. In the addition unit 115a, the loads w1, w2, ... Are multiplied and added, and in the function unit 115b, for example, the function z
= 1 / (1 + e- ^y ).

The output of the intermediate layer 112 is given to each neuron of the output layer 113, and the output layer 113, like the intermediate layer 112, applies a weight to the output of the intermediate layer 112, adds the outputs, and then converts them in the function part. , A converted feature vector is obtained as shown in FIG. Next, the character recognition by the template matching type character recognition device shown in FIG. 7 will be described.

In FIG. 7, a scanner 101 reads a handwritten character and outputs bit map data as shown in the figure. The normalizing means 102 normalizes the bit map data read by the scanner 101 to a predetermined size, and supplies the normalized bit map data to the feature amount extracting means 103. FIG. 9 is a diagram showing an example of the feature amount vector, and the feature amount extraction means 103 counts the number of dots in each of the vertical direction and the horizontal direction of the normalized bit map data shown in FIG. further,
The number of bits in the diagonal direction may be counted), the number of dots is divided by the maximum value, and all the obtained numerical values are arranged to form a feature quantity vector.

The feature amount vector extracted by the feature extracting means 103 is given to the feature amount converting neuro 104 and converted. The feature amount conversion neuro 104 is trained to output a standard (average) vector of a category to which a character belongs, when a feature amount vector of a certain character read by the scanner 101 is input in advance. The feature amount vector extracted by 103 is converted into a standard (average) feature amount vector of the category to which the character belongs in the feature amount conversion neuro 104.

Standard (average) of the category to which the character belongs
Feature conversion neuro 10 converted into the feature vector
The output of No. 4 is given to the collating means 106, collated with the template stored in the template storing means 105, and the recognition result is output. FIG. 10 is a diagram showing a block diagram in the case of learning the feature quantity conversion neuro 104 shown in FIG.

In the figure, when learning the feature amount transforming neuro 104 of FIG. 7, first, the random number is input to the input layer 111 (see FIG. 8) and the intermediate layer 1 of the feature amount transforming neuro 104.
It is given to the connection connecting 12 and the connection connecting the intermediate layer 112 and the output layer 113, and initialized. Next, a feature vector 131 that can be used for learning is extracted from various characters belonging to each category, and one or a plurality of templates 132 corresponding to each category are created by using the average value of the feature vector 131. .

Next, the feature vector 1 that can be used for the above learning
31 and the created template 132 are given to the input layer 111 (see FIG. 8) and the output layer 113 (see FIG. 8) of the feature amount transforming neuro 104, respectively, and the feature amount transforming neuro 1 is subjected to back propagation algorithm or the like, for example.
Train 04. As described above, the feature amount conversion neuro 1
By learning 04, the feature transformation neuro 10
When the feature amount vector 4 of a character is input, 4 outputs the standard (average) feature amount vector of the category to which the character belongs.

[0011]

By the way, in the learning method shown in FIG. 10, it is necessary to learn the feature quantity conversion neuro by using all the data of the feature quantity vector that can be used for learning, which is time-consuming. In addition, since learning for all data is performed at the same time, correct learning cannot be performed unless the learning as a whole is completed, resulting in a low recognition rate.

The present invention has been made in order to solve the above-mentioned problems of the prior art. It is possible to quickly learn a feature quantity conversion neuro that transforms a feature quantity vector, and even if learning does not end. Another object of the present invention is to provide a template matching type character recognition method by a neural network that can obtain a high recognition rate by using the feature amount conversion neuro.

[0013]

1 and 2 are block diagrams of the principle of the present invention. In order to solve the above-mentioned problems, the invention of claim 1 of the present invention, as shown in FIG. 1A, uses a feature quantity conversion neuro 2 to transform a feature quantity vector 1 extracted from a read character pattern into a feature of each category. Is converted into the standard feature vector 3 and the feature conversion neuro 2
In the template matching type character recognition method for recognizing a character by performing template matching 5 between the standard feature amount vector 3 converted by the above and one or a plurality of templates 4, as shown in FIG. , A template 4 is created from a feature vector 1 ′ that can be used for learning, and data for converting the created template 4 into the same template is given as a teacher signal to the feature conversion neuro 2 for learning, thereby reading the read character pattern. The feature amount vector 1 extracted from is converted into the standard feature amount vector 3 reflecting the features of each category.

According to a second aspect of the present invention, in the first aspect of the invention, the feature quantity conversion neuron 2 is trained by using the data for converting a given template into the same template as a teacher signal, and then, FIG. As shown in a)
When another feature amount vector 1 ′ is transformed by the feature amount transforming neuro 2 and the transformed feature amount vector 3 is not classified into the correct category by the template matching 5, it corresponds to another feature amount vector. From the templates 4 in the correct category, the template closest to the conversion result of the other feature amount vector by the feature amount conversion neuro is selected to generate re-learning data 6, and the other feature amount vector 1 ′ is used as the feature amount. As shown in FIG. 2B, the feature transformation neuro 2 is relearned by using the input signal of the transformed neuro 2 and the template closest to the transformation result of the other feature amount vector as the output signal. It is a thing.

According to a third aspect of the present invention, in the first aspect of the invention, the feature amount conversion neuron is learned by using the data for converting a given template into the same template as a teacher signal, and then, FIG. ), The feature amount vector 4 of the template is transformed by the learned feature amount transforming neuron to create a template 4 ′,
The converted template 4'is used as a template for performing template matching.

According to a fourth aspect of the present invention, in the second aspect of the present invention, another feature amount vector is used as an input signal of the feature amount conversion neuron, and the template closest to the conversion result of the other feature amount vector is used. 2 is used as an output signal to relearn the feature amount transformation neuro, and then the feature amount vector 4 of the template is transformed by the relearned feature amount transformation neuro 2 as shown in FIG.
And the converted template 4'is used as a template for performing template matching.

[0017]

In the first aspect of the present invention, since only the template 4 is used as the learning data of the feature transformation neuro 2, the feature transformation neuro 2 becomes an identity map with respect to the template 4, but other The data is also slightly pulled in, and an output close to the template can be obtained as the output of the feature amount conversion neuro 2.

Since most of the actual data is close to the corresponding template, when the transformation with the feature transformation neuro learned as described above is performed, pull-in occurs on the correct template side, and the features learned as described above. The recognition rate in the template matching can be increased as compared with the case where the quantity conversion neuron is not passed. Further, since the learning is performed using only the template as the teacher data, it is not necessary to learn the feature quantity conversion neuro using all the data of the feature quantity vector as in the conventional example.

In the invention of claim 2 of the present invention, in the invention of claim 1, when the feature quantity vector 1'is transformed by the feature quantity transforming neuro 2, the transformed feature quantity vector 3 is the template matching 5 If not classified into the correct category by, the template closest to the conversion result of the above-mentioned feature amount vector by the feature amount conversion neuro is selected from the templates 4 of the correct category to generate the re-learning data 6, and the feature amount Since the transformed neuro 2 is re-learned, the feature vector distant from the template can be transformed into the correct feature vector, and which template is used to easily learn the feature transformed neuro is easy. You can decide. Since most of the actual data is close to the corresponding template, the number of feature vectors to be relearned is small and the burden of relearning is small.

In the third and fourth aspects of the present invention, the template feature quantity vector 4 is transformed by the learned or re-learned feature quantity transforming neuron to create the template 4 ', which is transformed. Template 4 '
Since is used as a template for performing template matching, even a feature quantity conversion neuro that has not been learned can be used for character recognition.

[0021]

FIG. 3 is a flow chart showing a first embodiment of the present invention. In this embodiment, as shown in FIG. 1 (b), a feature quantity conversion neuron is learned using only a template as learning data. It shows an example. A first embodiment of the present invention will be described with reference to FIG.

In step S1, the operation of normalizing the character data and converting it into a feature amount vector is repeated for the number of input character data. In step S2, one or more templates are created from the feature amount vector obtained in step S1, and this is repeated for the type of the input character. In step S3, one learning pattern in which the template created in step S2 is used as an input pattern and an output pattern is created, and this operation is repeated for each template.

In step S4, a feature quantity conversion neuro having the same number of input units and output units as the dimension number of the feature vector is constructed, and in step S5,
The feature transformation neuron is made to learn all learning patterns. In the present embodiment, since the feature amount conversion neuron is learned by using only the template as the learning data, as described above, the feature amount vector close to the template may be slightly pulled in and an output close to the template may be obtained. it can. Also, since most of the actual data is close to the corresponding template, most of the data can be converted into the correct template by learning the feature quantity conversion neuron by the method of this embodiment.

In the first embodiment, as described above, the feature quantity vector close to the template can be converted into the correct template, but the feature quantity vector far from the template can be converted into the correct template. I can't. The second embodiment shown below shows a method for re-learning a neural network for a feature quantity vector for which a correct recognition result cannot be obtained by the feature quantity transforming neuron learned by the first embodiment. is there.

FIG. 4 is a flow chart showing a second embodiment of the present invention. In this embodiment, as shown in FIGS. 2 (a) and 2 (b), re-learning data is generated and re-learning data is stored. An example is shown in which a feature amount conversion neuro is learned by using it. A second embodiment of the present invention will be described with reference to FIG. In step S11, the feature quantity conversion neuro is learned by the method of the first embodiment shown in FIG.
Next, in step S12, the feature amount vector is transformed by the learned feature amount transforming neuro.

In step S13, step S12
The feature quantity vector converted in step S15 is compared with the template created by the method shown in the first embodiment, and if it is not recognized correctly, in step S15, the converted feature quantity in the template of the correct category. Choose the one that is closest to the vector. Then, step S16
In, in the input data is the feature vector before conversion,
A learning pattern is created using the template selected in step S15 as output data.

Then, as a result of the comparison in step S13, the feature amount vector that is not correctly recognized is processed in steps S15 and S16 to create a learning pattern. As described above, a learning pattern is created for a feature quantity vector that has not been correctly recognized, and in step S17, a feature quantity conversion neuron is learned using the created learning pattern.

As described above, in this embodiment, the first
By creating a learning pattern for a feature amount vector that could not be correctly recognized by the feature amount transforming neuron learned by the embodiment and re-learning the feature amount transforming neuron, a correct recognition result can be obtained for the feature amount vector. Further, it is possible to easily determine which template is used to train the feature amount transformation neuro.

FIG. 5 is a flow chart showing a third embodiment of the present invention. This embodiment is an embodiment for coping with the case where the learning of the feature amount conversion neuron is not completed in the first embodiment. This is shown and corresponds to FIG. 2C described above. In FIG. 5, in step S21, the feature quantity conversion neuro is learned by the method shown in the first embodiment, and then in step S22, the feature quantity of each template is converted by the feature quantity conversion neuro. Then,
In step S23, the converted feature amount is used as a new template.

Then, the above processing is repeated for the number of templates, the old template is discarded in step S24, and the new template obtained in step S23 is adopted. FIG. 6 is a flow chart showing a fourth embodiment of the present invention. In the present embodiment, when the learning of the template is not yet finished after re-learning the feature transformation neuro in the second embodiment. FIG. 2 shows an embodiment for dealing with the problem, and like FIG.
Corresponds to (c).

In FIG. 6, in step S31,
After re-learning the feature amount conversion neuro by the method shown in the second embodiment, the feature amount of each template is converted by the feature amount conversion neuro in step S32. Then, in step S33, the converted feature amount is used as a new template. Then, the above processing is repeated for the number of templates, the old template is discarded in step S34, and the new template obtained in step S23 is adopted.

In the first and second embodiments, when the learning of the feature amount conversion neuron is not completed, the template itself which has not been learned cannot be converted into the correct template, but the above-mentioned third embodiment is used. , Like the fourth embodiment, the template is created by converting the template with the feature transformation neuro, and template matching is performed using the created template.
Even a feature amount conversion neuro whose learning has not been completed can be used for character recognition.

[0033]

As is apparent from the above description,
In the present invention, since only the template is used as the learning data of the feature amount conversion neuro, it is not necessary to learn the feature amount conversion neuro by using all the data of the feature amount vector, and the feature amount conversion neuron can be learned at high speed. You can Further, when character recognition is performed using the feature amount conversion neuron learned as described above, a feature amount vector that was not correctly recognized and a template closest to the feature amount vector converted by the neural network are set. By using the re-learning data, the re-learning data can be easily created.

Further, by converting the template by the feature quantity conversion neuro and performing template matching using the converted template, a high recognition rate can be obtained even if a neural network in which learning has not been completed is used. it can.

[Brief description of drawings]

FIG. 1 is a principle block diagram of the present invention.

FIG. 2 is a block diagram (continuation) of the principle of the present invention.

FIG. 3 is a flowchart of the first embodiment of the present invention.

FIG. 4 is a flow chart of a second embodiment of the present invention.

FIG. 5 is a flow chart of a third embodiment of the present invention.

FIG. 6 is a flow chart of a fourth embodiment of the present invention.

FIG. 7 is a block diagram of a character recognition device which is a premise of the present invention.

FIG. 8 is a diagram showing an example of a configuration of a feature amount conversion neuro.

FIG. 9 is a diagram showing an example of a feature amount vector.

FIG. 10 is a diagram showing a conventional learning method for a feature amount transforming neuro.

[Explanation of symbols]

 2,104 Feature transformation neuro 4,4 ', 132 Template 101 Scanner 102 Normalizing means 103 Feature extracting means 105 Template storing means 106 Collating means 111 Input layer 112 Intermediate layer 113 Output layer 115 Elemental neuron

Claims (4)

[Claims] 1. A feature vector extracted from a read character pattern is converted by a neural network into a standard feature vector reflecting the characteristics of each category, and a standard feature vector converted by the neural network and a template. In the template matching type character recognition method for recognizing a character by performing template matching with, by applying data for converting a given template to the same template as a teacher signal to the neural network and learning, A template matching type character recognition method by a neural network, which is characterized by converting a feature amount vector extracted from a read character pattern into a standard feature amount vector reflecting features of each category. 2. When a neural network is trained by using data for converting a given template into the same template as a teacher signal and another feature amount vector is converted by the neural network, the converted feature amount If the vector is not classified into the correct category by template matching, the template closest to the conversion result of the other feature amount vector by the neural network from the templates of the correct category corresponding to the other feature amount vector. Retraining data is generated by selecting, and the other feature vector is used as the input signal of the neural network, and the template closest to the conversion result of the other feature vector is used as the output signal to regenerate the neural network. To learn The template matching type character recognition method by the neural network according to claim 1. 3. A neural network is trained by using, as teacher signals, data for converting a given template into the same template, and the feature vector of the template is converted by the learned neural network and converted. The template matching type character recognition method by a neural network according to claim 1, wherein the template is used as a template for performing template matching. 4. The neural network is relearned using another feature amount vector as an input signal of the neural network, and a template closest to the conversion result of the other feature amount vector is used as an output signal. The template matching type character by a neural network according to claim 2, wherein the feature quantity vector is converted by the re-learned neural network and the converted template is used as a template for performing template matching. Recognition method.