JPS638514B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a simple, fast and accurate character recognition method.

In general, recognition of characters and figures is performed by extracting strokes from character and figure patterns and extracting geometric features such as the positions, lengths, and mutual relationships between the strokes. And to extract the stroke, the following 2
Two methods are widely used.

Calculate the curvature of the contour point series detected by tracing the contour of the character shape, divide the contour series using points with large curvature values as division points, and extract strokes by combining the divided series. Do this.

A character/figure pattern is thinned to form a skeleton, and the connectivity and skeleton pattern of the skeleton pattern is traced to detect sudden angle changes and extract strokes.

However, when the character/graphic pattern is large or complex, the method requires an increased amount of processing and slows down the processing speed. This method has drawbacks such as distortion of the pattern and generation of whiskers due to oxidation, which complicates subsequent processing.

Therefore, as a method to eliminate these defects, the line width W of the character/figure pattern is detected, the pattern is scanned in a desired direction, and the length l of the writing point of the pattern in the scanning direction and the line width W are calculated. A sub-pattern representing the stroke component in the scanning direction is extracted by extracting entry points that satisfy lnW (n is a constant) between ), and a method has been proposed in which the character dose in each area is calculated and recognized.

However, as shown in the example of the kanji ``木'' in Figure 1, the vertical sub-pattern of A should be at position B in the figure, but when the character is transformed as shown in C in the figure, the vertical sub-pattern is In the pattern, as shown in D, the positions of the divided regions are shifted, that is, the extracted features become unstable, making character identification difficult. In order to absorb such character deformation, subsequent processing becomes complicated, which in turn causes a decrease in processing speed.

An object of the present invention is to perform blurring processing on the feature matrix and perform character recognition using the feature matrix in order to eliminate this conventional drawback, and will be described in detail below.

FIG. 2 shows an embodiment of the character recognition device of the present invention. The optical signal input 1 is converted into a binary quantized digital electrical signal by the photoelectric converter 2 and stored in the pattern register 3 . At the same time, the line width calculation section 4 calculates the line width of the input pattern. The vertical sub-pattern extraction section 5 performs a vertical scan on the entire pattern register, and extracts a vertical sub-pattern (VSP) from the relationship between the continuous length of black bits and the line width calculated by the calculation section 4. Similarly, the horizontal sub-pattern extraction section 6 extracts the horizontal sub-pattern (HSP) by a horizontal scan, the right-diagonal sub-pattern extraction section 7 extracts the right-diagonal sub-pattern (RSP) by a 45° right-diagonal scan, and the left-diagonal sub-pattern extraction section 8 extracts a left diagonal subpattern (LSP) by scanning at a left diagonal of 45 degrees.

The character frame detection section 9 detects a character frame circumscribing the character pattern in the pattern register 3, and sends the result to the character frame division determination section 10. The character frame division determination unit 10 divides the inside of the detected character frame into an M×N area (M,
N is an integer (in this embodiment, M=N=5), and the dividing point coordinates on the X and Y axes are determined. Here, the X axis indicates the horizontal direction of the character frame, and the Y axis indicates the vertical direction.

The feature matrix extraction unit 11 extracts VSP,
Divide the character frame area on each sub-pattern register of HSP, RSP, and LSP into M×N areas, count the number of black bits Bij in each area, and use the line width W to obtain the formula (1).
Calculate the feature indicating the character line length by M×N×4
Create a dimensional feature matrix.

Lij=Bij/W (1) Then, the VSP feature matrix is the length of the character frame in the y-axis direction ΔY, the HSP feature matrix is the length in the x-axis direction ΔX, and the RSP and LSP feature matrices are √() ² + Normalization is performed in steps () ² and finally an M×N×4-dimensional normalized feature matrix is created.

The blur processing unit 12 includes the feature matrix creation unit 11
Perform blur processing on the normalized feature matrix created in . In this embodiment, blurring is performed in the direction perpendicular to the sub-pattern extraction direction using the following method. That is, for the element Mij of the horizontal feature matrix created from the horizontal sub-pattern, the element M'ij of the blur feature matrix is calculated by applying the following equation (2).

M′ij=CMij+Mi, _j-1 +Mi, _j+1 /C+2 (2) Similarly, equation (3) is for the vertical feature matrix created from the vertical subpattern, and equation (3) is for the right diagonal feature matrix created from the right diagonal subpattern. teeth
For the left diagonal feature matrix created from the left diagonal subpattern using formula (4), formula (5) is applied to calculate the element M′ij of the blur feature matrix.

M′ij=CMij+M _i-1 ,j+M _i+1 ,j/C+2 (3) M′ij=CMij+M _i-1,j+1 +M _i+1,j-1 /C+2 (4) M′ij=CMij+M _i-1,j-1 +M _i+1,j+1 /C+2 (5) However, C is a constant, and when calculating for the outermost region, it is assumed that the elements of the further outer region are equal to Mij.

That is, when p>M or q>N or p<1 or q<1 holds true for the subscripts p and q of the feature matrix element value M _pq on the right side of each of the above equations, M _pq = M _ij .

Equation (2) above is based on the horizontal feature matrix element value M _ij of interest and the element value adjacent to it in the vertical direction.
This means blurring with a weighted average value using M _i,j-1 and M _i,j+1 . Since each element value is a quantity associated with the length in each region of the horizontal sub-pattern, it can be thought of as redistributing the length of each element in the vertical direction. Similarly, in the case of a vertical feature matrix, the right diagonal is redistributed to the left diagonally, and the left diagonal is redistributed to the right diagonal. The value of C is typically 4, and in this example, C=4.

The identification unit 13 applies the distance D defined by equation (6) between the standard character mask (fm) and the blur feature matrix (fi), and selects the category name of the standard character mask such that D is the minimum value. is output to the character output 14.

D=√( _n − _i ) ² (6) As explained above, in this embodiment, since the blurring process is applied to the normalized feature matrix, there is an advantage that deformation of characters can be absorbed.

In the present invention, after creating a normalized feature matrix, blur processing is performed by applying a simpler conversion formula, and identification is performed using this blurred feature matrix, so that deformation of characters can be absorbed. A simple and accurate character recognition device can be realized.

[Brief explanation of the drawing]

FIG. 1 is an explanatory diagram of a conventional character recognition system, and FIG. 2 is a configuration diagram of an embodiment according to the present invention. 1... Optical signal input, 2... Photoelectric conversion section, 3...
Pattern register, 4... Line width calculation section, 5 to 8...
...Sub pattern extraction section, 9...Character frame detection section, 1
0...Character frame division determination unit, 11...Feature matrix extraction unit, 12...Blur processing unit, 13...Identification unit, 14...Character name output.

Claims (1)

[Claims] 1. Store a digital signal obtained by photoelectrically converting and quantizing a character figure in a pattern register as an original pattern, extract sub-patterns representing stroke components in each direction from the original pattern and store them in the sub-pattern register, The region within the character frame of the original pattern is divided into M×N regions (M and N are integers) for the sub-pattern register, and the feature amount representing the character line length in each region is normalized by the character size. In the character recognition method, a feature matrix is created using a standard character matrix, and the character shape is recognized by comparing the feature matrix with a standard character matrix. A character recognition method characterized by identifying characters and shapes using