JPH07160830A - Character recognizing device using neuro network - Google Patents

Abstract

PURPOSE:To provide a character recognizing device using a neuro network for which recognition accuracy is improved. CONSTITUTION:A small area neuro network 100 to be the coupled shape neuro network for every small area of a character image is composed. By performing a learning by using a normal correlation value in this network 100, a network system 100 is constructed. The intermediate layer 11 of the constructed network system 100 is defined as an input layer, an integrated neuro network 101 to be the coupled shape is constituted by an original intermediate layer (second intermediate layer) and an original output layer (second output layer) 14 and a learning is performed by utilizing also the input layer 10 of the already constructed network 100. An actual character recognition is performed in the network composed of the input layer 10, the intermediate layer 11, the second intermediate layer 13 and the second output layer 14 of the network 100.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a character recognition device for recognizing characters (including figures and symbols) using a neuro network.

[0002]

2. Description of the Related Art A conventional character recognition apparatus using a neuro network is shown in FIG. This neuro-network is called a fully connected type, and consists of three layers, an input layer 1, an intermediate layer 2 and an output layer 3, and a certain type of unit has one unit with each unit of the adjacent layer. Are connected with the weighting factor of. Then, learning is performed in advance using teacher data to build a neuronetwork system, and at the time of recognition, a normalized character image (character to be recognized) is input and a recognition result is obtained from the output layer 3. It is the one. This is an example in which the intermediate layer 2 has one stage, but there are also examples in which there are two or more stages.

In FIG. 2, "0, 1, 2, ...
"... 9" indicates an output position corresponding to a recognized numeral that outputs "1" when each of the numerals 0, 1, 2, ..., 9 is recognized.
For example, when the number “2” is recognized, the output position 2 of the output layer 3 outputs “1”, and the other output positions 0, 1, 3, ...
All 9s output "0".

The fully connected neural network learning method is as follows. (1) Prepare a learning character image. For each character image,
Teacher data (correct answer: unit number of output layer to be reacted) is set in advance. The recognition accuracy of the neural network largely depends on the prepared character image for learning, and incorrect character data settings or character images that cannot be judged by humans should not be learned. (2) Initialize the weighting factor between each unit in the network and the offset value held by each unit with a random number. (3) Select one learning character image. Character images are selected evenly for each character type using random numbers.

(4) The selected character image is input to the input layer of the neural network and sequentially calculated from the intermediate layer to the output layer using each weighting coefficient and offset value. (5) An error between the value of the output layer output in (4) and the teacher data prepared in advance is calculated, and the weighting coefficient and offset value of each layer are changed so as to reduce the error. This learning method is generally called back propagation. (6), (3) to (5) are repeated an appropriate number of times.

Thus, a fully connected neuronetwork system was constructed. Next, character recognition is performed by the following method in this constructed network system.

The character recognition method in the fully combined form is as follows. The procedure for recognizing a character image of one character by this network is shown below. (1) A size normalization process and a brightness normalization process are performed on the clipped character image. Here, the size normalization is to change the character image size so as to match the size of the input layer of this network, and the brightness normalization processing is to ensure that all the brightness values are within the range defined by the specified brightness value. To change the brightness value (a range is, for example, brightness values 0 and 255.
Between).

(2) The character image data of (1) above is
It is input to the input layer of this network, and the weighting coefficient and the offset value are used to calculate the values sequentially with the intermediate layer and the output layer. (3) Obtain the recognition result from the value of the output layer of this network and output it.

[0009]

In the fully connected neuro-network, the shape of the entire character is the object of recognition, and learning or recognition in small area units of the character is not performed. Therefore, there is a problem that it is difficult to improve the recognition accuracy.

It is an object of the present invention to provide a character recognition device using a neuro so that not only the entire shape of a character but also the shape of a character in a small area unit is learned and recognized.

[0011]

According to the present invention, there is provided a small area neural network group having at least an input layer, an intermediate layer, and an output layer, formed in each small area unit forming a character pattern; All the input layers of the network group are input layers for the entire character pattern, and all the intermediate layers connected to this input layer are the first intermediate layers for the entire character pattern, and the second intermediate layer connected to this first intermediate layer. A character recognition device using a neuron having an intermediate layer of, and an integrated network having an output layer connected to the second intermediate layer;

Further, in the present invention, in the character recognizing device using the neuro, the first neuro network system is constructed by performing learning on the basis of the teacher data by the small area neural network group, and the first neuro network system is constructed in the integrated network. Input layer consisting of all subsequent input layers, first consisting of all intermediate layers
To construct a second neural network system by performing learning based on the teacher data using the hidden layer, the unconstructed second intermediate layer, and the output layer, and the second neural network system after the construction. Discloses a character recognition device using a neuro adapted to recognize a character pattern.

Further, according to the present invention, the teacher data for each small area in the small area neural network group is a normalized correlation value corresponding to the output layer for each small area.

Further, according to the present invention, a correlation value is obtained for each small area between the learning character image and each reference character image, and the correlation value is normalized for each small area in the order of magnitude of the correlation value. The obtained normalized correlation value is used as teacher data for each small area in the small area neural network group.

[0015]

According to the present invention, by incorporating a part of the small area network into a part of the integrated network, the learning effect of the small area network can be utilized in the integrated network. Characters are recognized by learning the shape of the entire character by using the built-in part of it.

Furthermore, by using the normalized correlation value corresponding to the output layer for the teacher data in the small area network, it becomes possible to learn various character corresponding values in the small area, and the recognition accuracy of the entire character shape is improved. improves.

[0017]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS FIG. 1 is a diagram showing an embodiment of a character recognition device of the present invention. This character recognition device is used for a small area network 100.
And an integrated network 101. The small area network 100 includes an input layer 10 and an intermediate layer 11 corresponding to each division when the input layer 10 is divided into small areas.
(1), 11 (2), ... And each intermediate layer 11 (1), 1
1 (2), ... Corresponding output layers 12 (1), 12 (2),
...... and consists of.

The integrated network 101 is a network having the intermediate layers 11 (1), 11 (2), ... Of the small area network 100 as input layers, and having an intermediate layer 13 and an output layer 14 connected to the input layers.

According to the networks 100 and 101, learning is divided into learning by the network 100 and learning by the network 101.
Learning is performed at 00 to build a small area network system.

Next, the integrated network 101 is learned under the constructed small area network system 100 to construct the integrated network system. Here, the construction means that a weighting coefficient between layers in the network is determined and a neuro system path from the input layer to the output layer is created. When learning in the integrated network 101, the input layer 10 and the intermediate layers 11 (1), 11 (2), which are part of the constructed small area network system 100,
... is used as is as the input layer. Middle layer 11
Each of (1), 11 (2), ... Is used as an input layer to the intermediate layer 13.

Character recognition is performed using a network constructed by learning. In this case, the input layer 10, the intermediate layers 11 (1), 11 (2), ..., The intermediate layer 13, and the output are provided. Character recognition is performed in a network system including the layer 14. That is, the normalized character image is input to the input layer 10 and the intermediate layer 11
(1), 11 (2), ..., The recognition result is obtained through the intermediate layer 13 and the output layer 14.

As described above, according to the above-described embodiment, the present network connects the following two three-layer hierarchical networks, and becomes a four-layer hierarchical network as a whole. -Small area network: A three-layer hierarchical network with small small area units, and there are as many small areas as there are small areas. -Integrated network: A three-layer hierarchical network with each middle layer of the small area network as an input layer. The small area network has a three-layer structure only for learning, and uses only the middle layer at the time of recognition.

Next, the small area will be described. An example in which the maximum character size is 8 × 12 matrix size will be described. Naturally, this size is arbitrary in nature. The size of each small area is the same, and 15 small areas are arranged for a matrix size (pixels) of 8 × 12.

An example of this arrangement is shown in FIG. The size of each small region is set to 4 × 4 pixel size, and 1 in FIG.
5 divisions are performed. Numbers 0 to 95 shown in the figure through FIG. 3 indicate pixel positions, and FIG.
An example in which 12 pixels are regularly divided into 6 by 4 × 4 pixels, and in FIG. 3B, an example of 3 divisions for interpolating the left and right central boundary parts in the example of 6 divisions in FIG. , Fig. 3 (a)
4 shows an example of 4 segmentation for interpolating the left and right central boundary portions, and FIG. 3D shows an example of 2 segmentation for interpolating the central boundary portion of FIG. 3B. .

Here, partitioning into 15 small areas for the input layer 10 means that signals are independently output in units of these 15 small areas, and for the intermediate layer 11, 15 areas are output. Of the intermediate layers 11 (1), 11
(2), ..., 11 (15), and for the output layer 12, fifteen intermediate layers 11 (1), 11
(2), ..., 11 15 output layers 12 corresponding to (15)
(1), 12 (2), ..., 12 (15).

The learning method in the network according to the above embodiment will be described below. The overall learning procedure is shown in FIG. 4, the learning procedure in the small area network 100 is shown in FIG. 6, and the learning procedure in the integrated network 100 is shown in FIG.

(1) Learning in the small area network 100. The small area network, when inputting one character image,
All 15 small area networks are learned. The learning procedure for one network is shown below. , Prepare character images for learning, and set teacher data for each character image. For the teacher data, as shown in FIG. 4, usually, the number of the output layer to react with the character image is prepared, but the teacher data in this small area network corresponds to each output layer as shown below. The normalized interlayer value is used. The method of calculating teacher data is shown below. -A reference character image corresponding to each output layer is determined from the prepared learning character images. The reference character image refers to a character image that is normally printed without any chipping, crushing, dust, or the like. -Using the prepared learning character image and each of the reference character images (this is the dictionary of normalized correlation), the normalized correlation value in the small area of interest to the network is calculated. -This normalized correlation value is used as teacher data.

First, the weighting factor between each unit in the network and the offset value held by each unit are initialized by random numbers. , One learning character image is selected. Character images are selected evenly for each character type using random numbers. Then, the selected character image is input to the input layer of the neural network, and sequentially calculated from the intermediate layer to the output layer using each weighting coefficient and offset value. The error between the output layer value output above and the teacher data prepared in advance is calculated, and the weighting coefficient and offset value of each layer are changed so as to reduce the error. , And the above are repeated an appropriate number of times.

(2) Learning in the integrated network 101. The integrated network is subjected to the following procedure after sufficiently learning the small area network. , Prepare character images for learning, and set teacher data for each character image. The teacher data here is the number of the output layer to be reacted with the character image in the normal case. , A weighting factor between each unit in the network and an offset value held by each unit are initialized by random numbers. , One of the learning character images is selected, and the values up to the intermediate layer are calculated using the learned small area network. Character learning is selected evenly for each character type using random numbers. The value of the intermediate layer of each small area network obtained above is input to the input layer of the present integrated network, and is sequentially calculated from the intermediate layer to the output layer using each weighting factor and offset value. The error between the output layer value output above and the teacher data prepared in advance is calculated, and the weighting coefficient and offset value of each layer are changed so as to reduce the error. , And the above are repeated an appropriate number of times.

(3) Regarding the recognition method in this embodiment. This recognition procedure is shown in FIG. This network is performed by calculating up to the middle layer of the small area network, inputting them to the input layer of the integrated network, and calculating up to its output layer. The procedure for recognizing a character image of one character by this network is shown below. The size normalization processing and the brightness normalization processing are performed on the cut out character image. The small area data of the character image data is input to the input layer of each small area network, and the value of each intermediate layer is calculated using the weighting factor and the offset value. The value of the intermediate layer of each of the small area networks is input to the input layer of the integrated network, and the values of the intermediate layer and the output layer are sequentially calculated using the weighting factor and the offset value. , The recognition result is obtained from the value of the output layer of the integrated network and is output.

Next, the normalized correlation value used in the small area network will be described. Given two images A and B of 4 × 4 size as shown in FIG. 7, these two images A and B
The correlation value of is to be obtained. Here, the image A is a 4 × 4 size image of an arbitrary small area in the learning character image, and the image B is a 4 × 4 size image of an arbitrary area in the reference character image. If the correlation value between the images A and B is r, it can be obtained by the following equation.

[0032]

[Equation 1] According to this formula, r is always

[Equation 2] Becomes Here, r = 1 represents a perfect match, and r = −1 represents a perfect mismatch. A perfect match means that A and B match in all pixels, and a perfect mismatch means that A and B do not match in all pixels. The perfect match can be expressed mathematically,

[Equation 3] It means that there are α and β that satisfy (α> 0).

FIG. 8 shows an example in which the above normalized correlation values are used as teacher data. In FIG. 8, it is assumed that the learning data of the upper right quarter of the numeral “0” as shown on the left side of the figure is learning data.
On the other hand, by using '0' of various shapes as a reference image,
A correlation value is obtained between the small areas (those on the upper right side) according to (Equation 1). Then, the maximum correlation value among them (the maximum value among 0.63, 0.97, and 0.87) is set as the representative correlation value. In this example, the typical value is 0.97.

Similarly, the correlation value for the corresponding small area of various reference images for the numeral "1" is obtained, and the maximum value thereof is used as the representative correlation value. In the figure, it is -0.42. Hereinafter, representative correlation values for the numbers 2, 3, ... In the figure, numerals 2, 3, ..., 8 are omitted, and an example of the representative correlation value 0.89 in numeral 9 is shown. Next, the representative correlation value calculated by the numbers 0 to 9 is normalized to 0 to 1.
In the figure, 0.97 → 1, -0.42 → 0.00, 0.8
An example normalized to 9 → 0.64 is shown. This normalized representative correlation value is the normalized correlation value. This is used as teacher data in the small area network.

That is, according to this teacher data creating method,
For example, if the upper left part of the character has a 1/4 circular arc shape as shown in FIG. 8, the teacher data is the characters "0", "2",
Corresponding places of '3', '6', '8', '9' show high values, and corresponding places of letters '1', '4', '5', '7' show low values. Become.

As described above, by performing learning using the features of the limited small area, the feature amount is automatically extracted in the middle layer of the small area network (the neuron automatically extracts the feature in the middle layer by learning). Then, it is possible to take out (which is generally said).

This will be further described. If there are similar patterns A and B in the neuro, if learning is performed with teacher data for outputting them to another output layer, the network may be confused and may not converge. In the case of a small area network, there are many characters in which the upper left part of the character has a shape like a 1/4 circular arc as described above, and if this is learned with normal teacher data, the situation as described above will occur. The accuracy is not improved. Therefore, in order to create the teacher data according to each shape, the teacher data is created using the normalized correlation value. In actual tests, the neuron of the present invention was clearly more accurate than the fully connected neuron for characters with a lot of missing or crushed characters. The fully-combined form reads only the features of the whole character,
The network of the present invention captures the characteristics of each part of the character and comprehensively judges them for reading. From this, it is considered that the result like the test result was obtained.

Next, the arrangement and size of the small areas will be described. The arrangement of the small area network of this network must be such that the character to be recognized can be identified. The smaller the number of small areas, the faster the recognition becomes possible. However, considering the correspondence to various characters (general versatility), it is desirable to arrange the entire character pattern as follows.

Further, it is desirable that the size of the small region should be such that the feature of the smallest shape (a shape such as a straight line or a circular arc) that constitutes a character is included as much as possible.

Although an example of a numeral is shown as a character, other than that, the invention can be applied to recognition of English characters, Japanese characters, symbols, figures and the like.

[0041]

According to the present invention, learning for each small area is performed by the small area network, further learning is performed by the integrated network using the learning result, and character recognition is performed by using the learning result. This made it possible to read characters with high recognition accuracy.

[Brief description of drawings]

FIG. 1 is a diagram showing an embodiment of a character recognition device using a neural network of the present invention.

FIG. 2 is a block diagram of a character recognition device using a conventional neural network.

FIG. 3 is a diagram showing an example of small area arrangement according to the present invention.

FIG. 4 is an example diagram showing normal teacher data.

FIG. 5 is a diagram showing a learning procedure in a character recognition device using a neural network of the present invention.

FIG. 6 is a diagram showing a learning procedure in the small area neural network of the present invention.

FIG. 7 is a diagram showing a figure for obtaining a correlation value according to the present invention.

FIG. 8 is an explanatory diagram for obtaining a normalized correlation value according to the present invention.

FIG. 9 is a diagram showing a learning procedure in the integrated network of the present invention.

FIG. 10 is a diagram showing a recognition procedure in the character recognition device according to the present invention.

[Explanation of symbols]

 10 Input layer 11 (1), 11 (2), ... Intermediate layer 12 (1) Output layer 13 Intermediate layer

Claims (4)

[Claims] 1. A small area neural network group having at least an input layer, an intermediate layer, and an output layer formed in each small area unit forming a character pattern; and a character string for all input layers of each small area neural network group. An input layer for the entire pattern, the whole intermediate layer connected to the input layer being a first intermediate layer for the entire character pattern, and a second intermediate layer connected to the first intermediate layer, An integrated network having an output layer leading to a second intermediate layer;
Character recognition device using a new neuron. 2. The neuro-based character recognition device according to claim 1, wherein learning is performed by a small area neural network group based on teacher data to construct a first neuro network system, and in the integrated network, The input layer consisting of all input layers after construction, the first intermediate layer consisting of all intermediate layers, the unconstructed second intermediate layer, and the output layer are used to perform learning based on the teaching data A character recognition device using a neuro that constructs a neuro network system of and a character pattern is recognized by the second neuro network system after the construction. 3. The character recognition device using a neuron according to claim 2, wherein the teacher data for each small area in the small area neural network group is a normalized correlation value corresponding to an output layer for each small area. 4. A correlation value is obtained for each small area between the learning character image and each reference character image, and the correlation value is normalized for each small area in the order of magnitude of the correlation value. The character recognition device using a neuron according to claim 3, wherein the generalized correlation value is teacher data for each small area in the small area neural network group.