JPH07110843A - Character recognizing device - Google Patents

Abstract

PURPOSE:To obtain a character code included in a highly reliable category to which an input character belongs as a character candidate. CONSTITUTION:A distance calculating means 101 calculates a distance value between an input character pattern and the reference pattern of each category in a character dictionary 102. A reliability evaluating means 103 is constituted of a neural net, previously learns relation between the distance value and reliability, and when the distance value of the input character pattern is inputted at the time of recognizing the relation, a reliability evaluating value to which each category belongs is outputted based upon the learned result. A teacher signal preparing means 104 prepares a teacher signal for learning the means 103 while changing it in accordance with the distance value order of the input category. A character candidate determining means 105 outputs a character code with a reliability evaluation value for a category whose reliability evaluation value is more than a previously determined threshold as a character candidate.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION The present invention compares an input character pattern with a standard pattern of each category stored in a character dictionary, and outputs a character code of a highly reliable category to which an input character belongs as a character candidate. The present invention relates to a character recognition device, and more particularly, to a character recognition device that outputs a character code with a reliability evaluation value belonging to each category.

[0002]

2. Description of the Related Art In character recognition by a pattern matching method, the distance between an input character pattern and a standard pattern of each category in a character dictionary is usually calculated, and the character candidates are ranked in descending order of distance (for example, Shinichiro Hashimoto). : Introduction to character recognition, pp. 32-36, Ohmsha (1982)). That is, the category with the smallest distance is set as the first character candidate. However, the method using such a distance value rank is the same as the first method.
There is a problem that the reliability of the correct answer is different depending on the case even for the character candidates. The relationship between the distance value and the reliability is complicated, and just adding the distance value does not indicate the reliability.

As a method of solving the above problem, a method of learning the relationship between the distance value and the reliability by using a neural net can be considered (for example, DE Rumelhart, J. et al.
L. McCelland and the PDP Research Group:
Parallel Distributed Processing, The MIT pre
ss (1986)). In this case, the input layer has the same number of units as the standard pattern, and the distance from the standard pattern is input. Units corresponding to each category are arranged in the output layer, and the reliability evaluation value in which the input category belongs to the corresponding category is output. At the time of learning, 1.0 is given as a teacher signal to the unit corresponding to the input category, and 0.0 is given to the other units. At this time, the category with the smallest distance may not match the input category, but it is expected that the recognition rate will be improved by correctly teaching the input category. However, in most cases, the category with the smallest distance becomes the input category for the same distance value input. Therefore, this method has a problem that a mixture is likely to occur when the category with the smallest distance does not match the input category and when the category with the smallest distance becomes the input category with the same distance value input.

[0004]

SUMMARY OF THE INVENTION The present invention has been made to solve the above problems, and an object of the present invention is to neuralize the relationship between the distance between an input pattern and a standard pattern and the reliability belonging to each category. We provide a character recognition device that can output highly reliable character candidates from distance value input by learning using the net and learning without adversely affecting other categories even if the minimum distance category does not match the input category. To do.

[0005]

A character recognition apparatus of the present invention comprises a character dictionary storing standard patterns of each category,
A distance calculating means for calculating a distance value between the input character pattern and the standard pattern of each category in the character dictionary, and a neural network in advance are used to determine the relationship between the distance value between the input character pattern and the standard pattern and the reliability. A reliability evaluation unit that learns and inputs the distance value calculated by the distance calculation unit and outputs a reliability evaluation value in which the input character pattern belongs to each category; and the distance when the reliability evaluation unit learns. By inputting the distance value calculated by the calculating means and the instruction of the input category, the category other than the instructed input category is constant, the input category is changed according to the distance value rank of the category, and the reliability is changed. A teacher signal creating means for creating a teacher signal for learning by the evaluating means and a reliability evaluation value output by the reliability evaluating means are input, and the reliability evaluation value is stored. For category is equal to or higher than a threshold value that defines beforehand, composed of a character candidate determination means for outputting designated by the reliability evaluation value the character code as a character candidate.

[0006]

The distance calculation means calculates the distance between the input character pattern and the standard pattern in the character dictionary. At the time of learning, the reliability evaluation means learns by the distance value and the teacher signal by the teacher signal creation means. The teacher signal creating means creates a teacher signal according to the distance value rank of the input category. When a new input character pattern is given at the time of recognition, the distance value is similarly calculated by the distance calculating means and input to the reliability evaluating means. The reliability evaluation means outputs the reliability evaluation value belonging to each category based on the learning result. This is input to the character candidate selection means, and the character code with the evaluation value is output as the character candidate in the descending order of the evaluation value for the category that outputs an appropriate threshold value or more.

According to the present invention, the relationship between the distance between the input pattern and the standard pattern and the reliability is learned by using a neural network, and a highly reliable character candidate can be output from the input of the distance value. Further, when the category with the smallest distance does not match the input category, learning can be performed without adversely affecting the other categories.

[0008]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described in detail below with reference to the drawings.

FIG. 1 is a block diagram showing an embodiment of the present invention. In the figure, an input character pattern is obtained, for example, by extracting character features such as PDC features in the form of a vector from an input character image (eg, Ogita, Naito, Masuda: Identification of handwritten Chinese characters by outer-direction contribution feature). , Theoretical theory (D), vol.J66-D, No.10, p
p. 1185-1192 (1983)). Character dictionary 10
In No. 2, several hundreds of character dictionary samples are collected per category to extract PDC features, and an average value for each category is created as a standard pattern.

The distance calculating means 101 calculates the distance value between the input character pattern and the standard pattern of each category stored in the character dictionary 102. That is, the input character pattern is x = (x ₁ , x ₂ , ..., X _n ) ... (1) The standard pattern of category j (j = 1, 2, ..., J) is m _j = (m _j1 , m _j2 , ..., _{M jn} ) (2), the distance l _j between the input character pattern and the standard pattern of the category j is calculated by using Euclidean distance as shown in Equation 1.

[0011]

[Equation 1]

The reliability evaluation means 103 is composed of a three-layer neural network as shown in FIG. In FIG. 2, the input layer has as many units as the number of standard patterns, and the output layer has units corresponding to each category. In the input layer,
The distance value l _{j from} the standard pattern obtained by the distance calculation means 101 and an instruction of what the input category is are input. The unit U _{j in the} output layer corresponds to the category j and outputs the reliability evaluation value in which the input character pattern belongs to the category j.

The operation of the reliability evaluation means 103 is as follows depending on learning and recognition.

At the time of learning, the distance value calculated by the distance calculation means 101 and the teacher signal generated by the teacher signal generation means 104 described later are input to the reliability evaluation means 103. Here, as one of the teacher signal generation methods, there is a method in which the unit corresponding to the input category is set to 1.0 and the other units are set to 0.0. If learning is performed using this teacher signal, the relationship between the distance value and the reliability is learned.
However, using this teacher signal creation method, as described above, there is confusion when the category with the smallest distance does not match the input category and when the category with the smallest distance becomes the input category with similar distance value input. There is a problem that it is easy. Therefore, when the category is not the minimum distance value, the teacher signal given to the input category is reduced as the distance value rank of the input sample decreases. That is, the teacher signal given when the distance value rank is i is y _i (i
= 1, 2, ..., J), y ₁ = 1> y ₂ > y ₃ ... (4). This can reduce the effect of the above confusion.

In the embodiment, the teacher signal creating means 104 is
The teacher signal y _i given to the unit U _{j according} to the distance value rank of the input category is created as shown in Expression 2 from the output of the distance calculating means 101 and the instruction of the input category (instructed by the user).

[0016]

[Equation 2]

FIG. 3 shows an algorithm for creating a teacher signal by the teacher signal creating means 104. In this way, when the input category is j, but the distance value l _j is not the minimum, the teacher signal y _i is made smaller than 1.0 to improve the learning effect.

At the time of recognition, when a new input character pattern is given, the distance from the standard pattern is calculated by the distance calculation means 101 as in the case of learning, and the reliability evaluation means 1
It is input to 03. The reliability evaluation means 103 outputs a reliability evaluation value belonging to each category based on the learning result.

From the output of the reliability evaluation means 103, the character candidate determination means 105 is a higher-ranked character candidate in the order in which a large value is output in the category that outputs a value equal to or greater than a predetermined threshold value a, Outputs character code with reliability evaluation value. If there is no category in the order in which the value equal to or greater than the threshold value a is output, there is no character candidate and it is rejected.

[0020]

As described above, in the character recognition device of the present invention, the relationship between the distance between the input character pattern and the standard pattern and the reliability belonging to each category is learned using a neural network, and the distance is learned. Even if the minimum category does not match the input category, there is an effect that the learning can be performed without adversely affecting the other categories and the character code of the highly reliable category to which the input character belongs can be output as a character candidate.

[Brief description of drawings]

FIG. 1 is a block diagram showing an embodiment of the present invention.

FIG. 2 is a diagram showing a configuration example of a neural network used in the reliability evaluation means of FIG.

FIG. 3 is a diagram showing an example of a teacher signal creation algorithm in the teacher signal creation means of FIG.

[Explanation of symbols]

 101 distance calculation means 102 character dictionary 103 reliability evaluation means 104 teacher signal generation means 105 character candidate determination means

Claims (1)

[Claims] 1. A character dictionary that stores a standard pattern of each category, a distance calculation unit that calculates a distance value between an input character pattern and a standard pattern of each category in the character dictionary, and a neural network in advance, The relationship between the distance value between the input character pattern and the standard pattern and the reliability is learned, and the distance value calculated by the distance calculation means is input.
A reliability evaluation unit that outputs a reliability evaluation value in which an input character pattern belongs to each category, and at the time of learning of the reliability evaluation unit, input a distance value calculated by the distance calculation unit and an input category instruction, A teacher signal creation unit that creates a teacher signal for learning of the reliability evaluation unit by changing the input category to a category other than the designated input category, and changing the input category according to the distance value rank of the category. Input the reliability evaluation value output by the reliability evaluation means,
A character candidate determining means for outputting the category of which the reliability evaluation value is equal to or higher than a predetermined threshold value, with the character code as a character candidate, and outputting the reliability evaluation value with the reliability evaluation value. Recognition device.