JPH02240790A - Character recognizing device - Google Patents

Abstract

PURPOSE:To execute learning of a relative position relation between each feature, as well by using a detecting position of a feature of a character as an input to a neural network. CONSTITUTION:A character is read by a photodiode array 3, a feature is extracted from character data, and a detecting position of its feature is converted into a numeric value with regard to each direction of the vertical and the left and right directions by a part of a numeric value converting part 5 of a feature detecting position. This value is inputted to an intermediate layer 6 of a neural net and a result of recognition is outputted from an output layer 7. When this result of recognition is different from an input character, a correct character name is inputted as an educator signal 8. 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 which has the same feature but in which a position of a feature is different can also be discriminated, and a character recognition which scarcely depends on a misalignment and a deformation of the character 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, in a character recognition device having a neural network that performs learning by backpropagation.

Conventional technology A character recognition device configured with a neural network that performs learning by backpropagation has been known for a long time. For example, after dividing an input character pattern into n rows and n columns of pixels and inputting them into a neural network, Learning is performed using methods such as backpropagation so that correct character recognition can be performed from pixel data. As shown in FIG. 6, a hierarchical network consisting of an input layer, a middle layer, and an output layer is used. The input layer is 64 (=8X8
) units 10, each corresponding to an optical sensor (for example an array of photodiodes) for reading characters. The number of units in the intermediate layer is eight, and the output layer is composed of four units. Each unit of the input layer is connected to every intermediate layer, and each unit 11 of the intermediate layer is connected to every output layer. Each unit in the intermediate layer and output layer is composed of neurons that use a sigmoid function as a threshold value.

Problems to be Solved by the Invention In the character recognition device constructed using the above-mentioned neural network, an input character is decomposed into pixels on a surface (input surface) composed of, for example, a sequence of n×m dots. Therefore, although the input data shown in FIGS. 7(a) and 7(b) have the same shape, the arrangement of dots is completely different. For example, if the pixels on the input screen where the character pattern exists ,, if we consider non-existent pixels to be 0, then Figure 7 (a
), then from the top left to the right, 00000 01000
The sequence is 010000100o 0OO00, but in Figure 7(b) it is ooooo oooio oo
ot.

The sequence is completely different from 00010 00000. Therefore, it is difficult for conventional character recognition devices to recognize characters that are misaligned or deformed.

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 has a part that extracts features from an input character matrix and a part that detects the position of the features, and receives the type of the feature and its detected position as input. In the neural network that performs learning, the detected position of the feature is decomposed into each direction (up, down, left, and right), and the relative position in each direction is converted into a numerical value and expressed.

Effect: According to the means described above, it is possible to distinguish between characters that have the same features but different positions of the features, and moreover, it is possible to recognize characters that are not easily dependent on positional displacement or deformation of the characters.

Embodiment FIG. 1 shows an input screen of an embodiment of the character recognition device of the present invention. The figure exemplarily shows an input screen with 5 rows and 5 columns and the character "4" input there.Characteristics within the input screen (
For example, the features shown in FIG. 2) are extracted, and the detected position is multiplied by a function as shown in FIG. 3 in each of the up, down, left, and right directions and expressed as an analog quantity between 0 and 1. Third
The function in the figure matches the length of the horizontal axis with the width of the characters in FIG. When the function values at the positions where the feature exists are read from FIG. 3, the values are as shown in FIG. 2, respectively. Character recognition can be performed by inputting these numerical values into a three-layer neural network and having it learn.

FIG. 4 shows a block diagram of the character recognition device of the present invention. Characters are read by a photodiode array 3, features are extracted from the character data, and the detected position of the feature is digitized in each of the up, down, left, and right directions by a feature detection position digitizing section 5. This value is input to the intermediate layer 6 of the neural network, and the recognition result is output from the output layer 7. If this recognition result differs from the input character, the teacher signal 8 indicates the correct character name.
Enter as . Learning is performed by changing the weights of the neural network using the pack propagation method so as to reduce the difference between the teacher signal and the recognition results of the neural network, and the learning is repeated until correct recognition is achieved. A flowchart of the operation of this character recognition device is shown in FIG.

The feature extraction layer of the embodiment shown in Fig. 2 extracts features formed by three layers and three pixels, but instead, a feature extraction method used in conventional AI (artificial intelligence) may also be used. good.

Further, the function shown in FIG. 3 used to quantify the position information does not need to be a Zsigmond function, but may be any monotonically increasing function, and a straight line may be used.

Although the present invention has been described above based on the embodiments, it goes without saying that the present invention is not limited to the above embodiments, and can be modified in various ways without departing from the spirit thereof. For example, in the embodiment shown in FIG. 1, the number of pixels on the character input surface is 1.
A large number such as 6x16 or 24x24 may be used, and handwriting input using a tablet may be used for the character reading part.

Effects of the Invention According to the present invention, by using the detected positions of character features as input to a neural network, it is possible to learn the relative positional relationship between each feature, and it is possible to avoid deformation and positional deviation of characters. A strong character recognition device is obtained.

[Brief explanation of drawings]

FIG. 1 is a diagram of an example of readable characters according to the present invention, FIG. 2 is a diagram showing features extracted for the characters in FIG. 1 and their position information,
FIG. 3 is a diagram of a filter function used for feature position detection according to the present invention, and FIG. 4 is a configuration diagram of a character recognition device according to the present invention.
FIG. 5 is a flowchart of character recognition according to the present invention, FIG. 6 is a block diagram of a conventional character recognition device, and FIG. 7 is an explanatory diagram of problems in the conventional character recognition method. 1.13... Pixel where a character pattern exists, 2
．． 12...Pixel where no character pattern exists, 1
0... Input layer pixel, 11... Neuron. Name of agent: Patent attorney Shigetaka Awano and one other person Figure ((IJ pressure direction (b) right direction left and right + -m-University pattern child position III series 2-
：A diagram of the history of patterns and kotois

Claims (1)

[Claims] It has a part that extracts character features and a part that detects the position of the feature, and receives the feature and its detected position as input,
A character recognition device comprising a neural network that performs learning by backpropagation.