JPH02240791A - Character recognizing device - Google Patents

Abstract

PURPOSE:To recognize a character pattern having a positional deviation and a deformation by extracting a feature from an inputted character pattern, inputting its feature and allowing a neural network to execute learning. CONSTITUTION:A character is read by a photodiode array 28, and from character data, a feature is extracted by a feature extracting layer 29, inputted to an intermediate layer 30 of a neural net and from an output layer 31, a result of recognition is outputted. When this result of recognition is different from an input character, a correct character recognition result is inputted as an educator signal 32. Subsequently, learning is executed by varying weight of the neural net by a back propagation method so that a difference of the educator signal and the result of recognition of the neural net becomes small, and learning is repeated until a correct recognition is executed. In such a way, a character recognition which reduces a positional deviation and a deformation of an input character pattern by a feature extracting surface placed on the whole input surface can be executed.

Description

DETAILED DESCRIPTION OF THE INVENTION Field of Industrial Application This invention relates to a character recognition device, and can be used, for example, as an input device for handwritten characters into a computer.

Conventional technology Conventionally, when character recognition is performed using a neural network, an input character pattern is divided into m rows and n columns of pixels and input to the neural network, and then correct character recognition is performed using mxn pixel data. Learning is performed using methods such as backpropagation to ensure that

For example, as shown in FIG. 5, a hierarchical network consisting of an input layer, a middle layer, and an output layer is used. The input layer is 64 (
=8x8) pixels 34, the number of intermediate layer units is eight, and the output layer is composed of four units 35. Each unit of the input layer is connected to all hidden layers, and each unit of the hidden layer is connected to all output layers. Each unit 35 of the intermediate layer and output layer is composed of neurons using a sigmoid function as a threshold numerical value.

Problems to be Solved by the Invention For example, consider a case where an input pattern is divided into 5×5 pixels as shown in FIG. 6, and learning is performed using these as input. Since this input can be regarded as 25 data strings, even if the input pattern is the same, if the position of one row is shifted, the data string becomes a completely different pattern string and cannot be recognized as shown in FIG. A similar problem occurs even if the input cover pattern is deformed, so it is extremely difficult to recognize character patterns such as handwritten characters that are misaligned or deformed using this method.

An object of the present invention is to provide a character recognition device for recognizing character patterns that are misaligned or deformed.

Means for Solving the Problems The character recognition device of the present invention extracts features included in an input character pattern and configures a neural network that performs learning using the features as input. The feature extraction surface is arranged over the entire input surface by shifting the feature extraction surface one row and one column at a time.

Effect: According to the above-described means, character recognition can be performed by reducing positional deviation and deformation of an input character pattern using the feature extraction surface arranged over the entire input surface.

Embodiment FIG. 1 shows a feature extraction surface of an embodiment of a character recognition device to which the present invention is applied. In this example, there are 3 x 3 feature extraction planes A, B,...■ shown in Figure 1 (b) for the 5 x 5 input plane shown in Figure 1 (a). , each corresponds to the pixel at the numbered position on the input surface. If a pattern like that shown in FIG. 2 exists in this feature extraction surface, a value of "1" is associated, and if it does not exist, a value of "0" is input to the neural network, and learning is performed by backpropagation. In FIG. 2, 26 represents a pixel where no pattern exists, and 27 represents a pixel where a pattern exists.

FIG. 3 shows a block diagram of the character recognition device of the present invention. Characters are read by a photodiode array 28, features are extracted from the character data by a feature extraction layer 29, inputted to an intermediate layer 30 of a neural network, and a recognition result is output from an output layer 31. If this recognition result differs from the input character, the correct character recognition result is input as the teacher signal 32. Learning is performed by changing the weights of the neural network using the backpropagation method so as to reduce the difference between the teacher signal and the recognition result of the neural network, and the learning is repeated until correct recognition is achieved. Character data may be input by handwriting using a tablet instead of by optical reading. The method of the present invention is applied to the feature extraction layer 29. A flow chart of the operation of this character recognition device is shown in FIG.

In addition, in the embodiment of the present invention, the feature extraction surface is placed over the entire input surface in order to take advantage of the high speed provided by parallel processing, but in order to simplify the circuit, the connection between the feature extraction surface and the input surface is Features may be extracted by sequentially scanning the entire input surface while changing the input area.

The features extracted on the feature extraction plane are composed of 3 x 3 pixels, so there are 29 possible combinations, but among these, the features that may actually appear in the character pattern are used to distinguish between characters. Only useful things should be used for extraction. In the embodiment of the present invention, for example, 27 features as shown in FIG. 2 are extracted. In this example 3×3
Although features of pixels are extracted, it may be square or rectangular, even if there are many pixels such as 4×3 or 5×5.

The feature extraction plane may be configured by fixing connections so as to output an output corresponding to the features shown in Fig. 2, for example, but this part may also be configured as a neural network and output corresponding to the features shown in Fig. 2 may be configured. It is also possible to perform learning by backpropagation in advance so as to output an output and use it as a feature extraction surface.

Effects of the Invention As described above, according to the present invention, character patterns with misalignment or deformation can be recognized by extracting features from input character patterns and using the features as input for learning by a neural network. I can do it.

[Brief explanation of drawings]

Figures 1(a) and 1(b) show the character input surface ((a)) and feature extraction surface ((b)) of a character recognition device according to an embodiment of the present invention.
), FIG. 2 is a diagram of feature patterns extracted by the character recognition device of the present invention, FIG. 3 is a block diagram of the character recognition device of the present invention, and FIG. 4 is a diagram of character recognition by the character recognition device of the present invention. 5 is a flowchart, FIG. 5 is a diagram of a conventional character recognition device using a neural network, and FIGS. 6 and 7 are diagrams showing examples of character reading by the conventional character recognition device. 1 to 25... Character reading pixels, 26, 36.4
0...Pixel where pattern exists, 27, 37.
39...Non-existent pixel, 28...Photodiode array, 29...Feature extraction layer,
30... Middle layer, 31... Output layer, 3
2... Teacher signal, 33... Learning control unit using back propagation method, 34... Input layer, 35... Neuron. Name of agent: Patent attorney Shigetaka Awano and one other person

Claims (1)

[Claims] A character recognition device comprising a neural network, extracting features from an input character pattern, and causing the neural network to learn the types of the features by backpropagation.