JPH0367381A - Character recognition device - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

DETAILED DESCRIPTION OF THE INVENTION Field of the Invention The present invention relates to a character recognition device suitable for recognizing handwritten or printed character patterns composed of two-dimensional dot matrices such as those read from an image scanner or the like. It can be used as an input device for handwritten or printed characters into a computer.

Conventional technology A conventional example of a method for extracting features from a character pattern consisting of a two-dimensional dot matrix read from an image scanner etc. is a 3x3 dot matrix as shown in Figures 5 and 6. There is a method to extract features consisting of patterns. In this method, there are 3×3 feature extraction regions A, B, ...I shown in FIG. 5(b) for the 5×5 input surface shown in FIG. 5(a). , each corresponds to the pixel at the numbered position on the input surface. The pattern in this feature extraction area is compared with the feature pattern shown in FIG. 6 to check whether the feature shown in FIG. 6 exists. Then, if each feature exists, the value of l, and if it does not exist, the value of O is associated and input to the neural network, and learning is performed by pack propagation. In Figure 6, 38 is a pixel where no pattern exists;
9 represents a pixel where a pattern exists.

Problems to be Solved by the Invention When reading character patterns from an image scanner, etc.
A discretization error occurs because a character pattern, which is a continuous quantity, is read with a finite number of pixels. Pixels of reading devices such as image scanners are usually made of photoelectric conversion elements such as photodiodes.

Then, when a character pattern does not exist on a pixel, an output corresponding to the area of the character pattern that occupies that pixel is obtained. If it is above the threshold value, it is defined as 1 and binarization is performed. In this way, character patterns are discretized, but even if the character pattern is the same, even a slight shift in its position will result in a large change in the dot pattern after reading. For example, “1” as shown in FIG. 7(a)
Consider the case where a character pattern of 5×5 pixels is read and binarized. FIG. 7(b) shows the reading results, showing that 1 is output for the five pixels marked with diagonal lines, and O is output for the remaining pixels. FIG. 7(C) shows the reading result when the same character pattern is shifted slightly to the right, and the result is different from FIG. 7(b). In this way, even with the same character pattern, the dot pattern output by the reading device varies greatly depending on the positional relationship with the pixels of the reading device. This obviously poses a problem in recognizing the character pattern. To avoid this, increase the resolution of the reading device and read the character pattern using many fine pixels. However, in that case, the size of the feature becomes larger than 3×3 dots, and the feature cannot be extracted using the conventional method. In addition, the conventional feature extraction method that simply increases the number of dots may not extract features well or may extract completely different features when the size of characters, line thickness, or shape changes. There is sex. In particular, when recognizing handwritten characters, it is necessary to extract features that are less affected by changes in character size, line thickness, and shape in order to increase the recognition rate.

The purpose of the present invention is to provide character recognition that performs feature extraction with little influence from changes in character size, line thickness, and shape in recognition of handwritten or printed character patterns composed of dot matrix read from an image scanner or the like. The goal is to provide equipment.

Means for Solving the Problems In order to achieve the above object, the character recognition device of the present invention employs the following feature extraction method.

(1) The input screen, which holds handwritten or printed character patterns read from an image scanner or the like as a two-dimensional dot matrix, is divided into multiple regions and features are extracted within each region. These regions may overlap each other.

(2) Each cell of the feature extraction surface is scanned within each region, or each cell of the feature extraction surface is arranged in parallel throughout the region.

(3) Cells on the feature extraction surface consist of dot matrix,
Each dot has a weight value. A negative value is a weight for a dot for which a pattern must not exist in order to determine that this feature exists, and a positive value is a weight for a dot for which a pattern must exist.

0 is the weight for dots that may or may not have a pattern.

(4) To determine whether a feature exists or not, multiply the value of each dot on the input surface by the weight of the cell dot on the feature extraction surface, and then calculate the sum over all dots on the cell. This is done by comparing the value with a threshold value certified for each cell.

Effects According to the means described above, it is possible to provide a feature extraction method that is less affected by changes in character size, line thickness, and shape.

Example An embodiment of the present invention will be described below with reference to the drawings.

FIG. 1 is a diagram showing an input surface and a feature extraction surface, and FIG. 2 is a diagram showing an example of cells on the feature extraction surface.

In FIG. 1, numeral 1 indicates a character pattern input surface. Handwritten or printed character patterns to be recognized are read from an image scanner or the like. The input surface is a matrix of dots corresponding to the two-dimensionally arranged pixels of the character pattern, and each dot holds the character pattern by having an O or positive value. For example, if the read character pattern is binarized, dots for pixels where the character pattern does not exist have a value of 0, and dots for pixels where the character pattern does exist have a value of 1. The character pattern may not be binarized, and each dot may have a positive value corresponding to the gradation, for example. 2 in FIG. 1 is one of the regions in which the input surface is divided into a plurality of regions. Each divided region becomes a unit for extracting features. In Figure 1, 1 with no overlap
Although an example is shown in which the image is divided into six areas, these areas may have portions that overlap with each other. Further, the number and size of areas can be freely set. 3 is an enlarged area of 2. In this example, this area has a size of 16 dots x 16 dots. 4 is a feature extraction surface, and 5 is one cell of the feature extraction surface. In Figure 1, the feature extraction surface has 16 cells. One cell on the feature extraction surface corresponds to one feature to be extracted. In other words, there are as many cells as the number of features to be extracted on the feature extraction surface. Either all the cells on the feature extraction surface 4 are scanned vertically and horizontally in 3, row by column, or each cell is arranged in parallel in 3 so that each feature exists in 3. Find out if it does.

FIG. 2 is an example of cells on a feature extraction surface for extracting features of a size of 12×12 dots. Figure 2 (a) is a cell for extracting features at the top of the vertical bar, Figure 2 (b) is a cell for extracting features that are bent, Figure 2 (C
) is an example of cells for extracting intersecting features. Each cell on the feature extraction plane is a dot matrix with a size equal to the size of the feature to be extracted. That is, one dot on the cell corresponds to one dot on the input surface. Each dot has a weight value. In the example of the cells on the feature extraction surface in FIG. 2, the weights have three values -1, 0, and 1. 6 in Figure 2 is a dot with a weight of 1,
7 is a dot with a weight of O, and 8 is a dot with a weight of -1. A negative value is a weight for a dot for which a pattern must not exist in order to determine that this feature exists, and a positive value is a weight for a dot for which a pattern must exist. O is the weight for dots that may or may not have a pattern. In this way, by setting the weight of O and recognizing dots that may or may not have a pattern when extracting features, feature extraction is possible with less influence from changes in font size, line thickness, and shape. You will be able to do this. In Figure 2, 3 of the feature extraction plane
Although only one example of cells is shown, in order to perform character recognition, features are extracted using cells corresponding to other features that are considered necessary for character recognition. In addition, the weight value of each dot of the cell on the feature extraction surface is −
There are three values, 1, 0, and 1, but the value may be set multiple times. By increasing the number of values and appropriately setting the threshold, more flexible feature extraction becomes possible. Furthermore, in the example of FIG. 2, the size of the cells on the feature extraction surface is 12×12 dots, but this size only needs to be larger than the number of dots corresponding to the line width of the read character. However, the larger this size is, the more margin there is for changes in the thickness and shape of the characters, but if it is too large, the negatively weighted portions will come into contact with the surrounding patterns, making it impossible to extract features.

Hereinafter, the feature extraction method in the character recognition device of the present invention will be explained step by step.

Handwritten or printed character patterns to be recognized that are read from an image scanner or the like are held as a matrix of dots on the input surface. Each dot has an O or positive value, O indicates that there is no pattern for that dot,
Positive numbers indicate a pattern. The input surface is divided into several areas. The respective regions may have portions that overlap with each other. Feature extraction is performed for each region. In one region, each cell on the feature extraction plane is scanned vertically and horizontally in every row and column, or each cell is arranged in parallel in the region. Then, it is determined whether the feature exists at each location. A determination as to whether a feature exists is made as follows.

The value of each dot on the input surface is multiplied by the weight of the corresponding dot of the cell on the feature extraction surface, and the sum is calculated for all dots on the cell.

The value is then compared with a threshold value set for each cell, and if it is larger than the threshold value, it is determined that the feature exists, and if it is smaller than the threshold value, it is determined that the feature does not exist. The cell outputs its results. All the features contained in a region can be extracted by scanning all the cells on the feature extraction surface in one region, or by making the above judgment for each cell arranged in parallel in the region. Extract. By performing this operation on all areas on the input screen, the features of the entire human cover pattern are extracted.

FIG. 3 is a block diagram of character recognition using a three-layer hierarchical neural network, and FIG. 4 is a flowchart.

In FIG. 3, 1 is an input surface, and 4 is a feature extraction surface.

9 is an input layer of a neural network having a three-layer hierarchical structure, 10 is an intermediate layer, and 11 is an output layer. 12 in Figure 3
represents a neuron. Furthermore, arrows between neurons represent connections between the neurons. Each connection has a strength, and when a signal propagates through each connection, the strength of that connection is applied as a weight to the signal. Figure 3 shows a three-layer neural network with four neurons in the input layer, three neurons in the hidden layer, and four neurons in the output layer.
The number of layers and the number of neurons in each layer can be freely set.

A method for character recognition using a hierarchical neural network will be described below with reference to FIG.

Features are extracted from handwritten or printed character patterns read by an image scanner or the like using the feature extraction method described above. The results are output from each cell on the feature extraction surface. This information is input to each neuron in the input layer of a hierarchical neural network. When human input is applied to a neural network, signals propagate through the neural network, and recognition results are output to each neuron in the output layer of the neural network. If the recognition result output from the neural network is incorrect, learning is repeated until the correct recognition result is output. During learning, the strength of connections between neurons is changed according to the pack propagation learning rule by providing a desired output as a teacher signal.

The above describes a method for character recognition using a hierarchical neural network, but the features extracted using the feature extraction method of the character recognition device of the present invention are classified based on conditions, as is done in artificial intelligence. It can also be used in other character recognition methods, including the method of determining characters by using

Effects of the Invention According to the present invention, changes in character size, line thickness, and shape have little effect on recognition of handwritten or printed character patterns composed of two-dimensional dot matrix read from an image scanner or the like. A method of feature extraction can be provided.

[Brief explanation of drawings]

Figure 1 is a diagram of an input screen and cells according to an embodiment of the present invention, Figure 2 is a diagram of an example of cells, Figure 3 is a block diagram of a character recognition device using a neural network with a three-layer hierarchical structure, and Figure 4 is a flowchart of character recognition, Figure 5 is a diagram of the input screen and feature extraction area of a conventional character recognition device, Figure 6 is a diagram of the characteristic pattern of a conventional character recognition device, and Figure 7 is a diagram of the conventional character recognition device. It is a figure showing the problem of a recognition device. 1...Input surface, 2...Regions obtained by dividing the input surface, 4...Feature extraction surface, 5...Cells of feature extraction surface, 6. ...A dot with a weight of 1, 7...A dot with a weight of O, 8...
・Dot with a weight of 1, 9...Input layer of a 3-layer hierarchical neural network, 10...
Middle layer of a three-layer hierarchical neural network, 11
...Output layer of 3-layer hierarchical neural network, 12...Neuron, ■3-37...
... Pixel for character reading, 38.40...
Pixels with no pattern, 39, 41... pixels with a pattern.

Claims (3)

[Claims] (1) An input surface that holds a character pattern by dots having information arranged in a two-dimensional matrix, the input surface being divided into a plurality of regions,
A number of dots equal to or less than the number of dots in the row and column directions of the area are arranged in the row and column directions, respectively, and each of the dots is weighted with a positive, 0, or negative value. What is claimed is: 1. A character recognition device comprising cells marked with , and means for scanning the area by moving the cells in the row and column directions. (2) An input surface that holds a character pattern by dots having information arranged in a two-dimensional matrix, the input surface being divided into a plurality of regions,
A number of dots equal to or less than the number of dots in the row and column directions of the area are arranged in the row and column directions, respectively, and each of the dots is weighted with a positive, 0, or negative value. A plurality of cells are arranged in parallel to cover the area so that the cells correspond to positions shifted by one dot in the row or column direction within the area. A character recognition device characterized by: (3) The character recognition device according to claim 1 or 2, characterized in that it has a plurality of types of cells having different ways of arranging dots having positive, 0, and negative values.