JPH0545990B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION (Technical Field) The present invention relates to a fast and accurate feature extraction method.

(Background Art) Conventional character/figure recognition devices often employ a method of extracting strokes from character/figure patterns and recognizing them using the positions, lengths, and mutual relationships between strokes of the extracted strokes. ing. This type of device (1) calculates the curvature of a contour point series detected by tracing the contour of a character figure, divides the contour series using points with a large value of curvature as dividing points; Extract the strokes by combining the series, or
or (2) perform thinning processing on the character figure pattern to create a skeleton, trace the connectivity and skeleton pattern, detect sudden angle changes, etc., extract strokes, and extract strokes from the extracted strokes. Character shapes were recognized by extracting geometric features of strokes. However, in method (1), when the character/graphic pattern becomes large or complex, the amount of processing increases, resulting in a decrease in processing speed. Furthermore, there are problems such as pattern distortion and whiskers due to thinning, and subsequent processing becomes complicated.

(Problems to be solved by the invention) An object of the present invention is to improve these drawbacks by extracting a sub-pattern representing a stroke component in a desired direction of a character figure pattern, and scanning the sub-pattern from a point on the side of a character circumscribing frame. Detect the positions of all character lines on the scanning line, normalize the distance between the scanning start point and the character line by the size of the character circumscribing frame in the scanning direction, and extract the side-by-side array of the sum of N powers. It is characterized by dividing and extracting a group of feature vectors, and its purpose is to provide a fast and stable character recognition method.

(Structure and operation of the invention) FIG. 1 shows an embodiment of the character recognition device of the invention, FIG. 2 shows an example of a sub-pattern, and FIG. 3 shows an example of feature extraction.

In FIG. 1, 1 is a photoelectric conversion section, 2 is a pattern register, 3 is a line width calculation section, 4 is a sub-pattern extraction section,
5 is a character frame detection section, 6 is a feature extraction section, 7 is a feature matrix extraction section, 8 is an identification section, and 9 is a character name output.

In the operation of this embodiment, characters on a form set in a reading mechanism are converted into a binary quantized digital electrical signal by a photoelectric converter 1, and stored in a pattern register 2. At the same time, the line width (W) of the input pattern is calculated in the line width calculating section 3. The sub-pattern extraction unit 4 performs a vertical scan on the entire pattern register, and extracts a vertical sub-pattern (VSP) based on the relationship between the continuous length of black bits and the line width calculated by the line width calculation unit 3.
Extract. Similarly, a horizontal subpattern (HSP) is extracted by a horizontal scan, a right diagonal subpattern (RSP) is extracted by a 45° rightward scan, and a left diagonal subpattern (LSP) is extracted by a 45° leftward scan. Figure 2 shows an example of the original pattern and each sub-pattern, where a is the original pattern, b is the vertical sub-pattern (VSP), c is the horizontal sub-pattern (HSP), d is the right diagonal sub-pattern (RSP), and e is the left diagonal. It is a subpattern (LSP). The character frame detection unit 5 detects a rectangular frame (hereinafter referred to as a character frame) circumscribing a character figure pattern in a pattern register, and determines a position for defining the character frame in a two-dimensional plane defined by the pattern register. The coordinates are sent to the feature extraction unit 6. In the following description, a coordinate system will be used in which the origin is at the lower left of the character frame, the X axis is in the horizontal direction, and the Y axis is in the vertical direction. The feature extraction unit 6 first starts horizontal scanning of the vertical sub-pattern from a point P (0, y) on the left side, which is the vertical side of the four sides forming the character frame, and detects the change from a white point to a black point. A value obtained by detecting all points and normalizing the distance between the detected change point and the point P on the vertical side where the scanning started, that is, the difference in the X coordinate, using the length of the character frame in the X direction as a normalization constant. to the Nth power (N is a constant, in this example, N=
The value of 2) is calculated for all the detected change points, and the total sum is stored in the array V _l (y). However, the white dots represent the character background parts, and the black dots represent the character line parts. Also, equation (1) is the above-mentioned V _l (y)
is expressed by the formula, and △X _k indicates the distance between each change point and the scanning start point on the side of the character frame,
k=1, . . . , K, K represents the number of detected change points. Further, ΔX in equation (1) is the length of the character frame in the horizontal direction, and C is an integer conversion constant, and in this embodiment, C=50.

V _l (y) = _K 〓 ^k=1 {△x _k ／△X×C} ² ...(1) Perform the above process starting from all points on the two vertical sides of the character frame. , for the vertical subpattern, an array V _l (i) created by starting horizontal scanning from a point on the left side of the character frame, an array V _r (i) created by starting horizontal scanning from a point on the right side of the character frame )
Extract. However, i=0, . . . , YT, YT is the Y coordinate of the upper side of the character frame. Through similar processing, vertical scanning is performed from all points on the two horizontal sides of the character frame for the horizontal sub-pattern, right diagonal sub-pattern, and left diagonal sub-pattern, and the array H _b (j ), H _t (j), array R _b (j), R _t (j) for the right diagonal sub-pattern, array L b (j), L _t ( _j ) for the left diagonal sub-pattern
Extract. However, j=0, . . . , XR, XR is the X coordinate of the right side of the character frame. The subscript b of the array to be extracted for the horizontal subpattern, right diagonal subpattern, and left diagonal subpattern indicates that the starting point is a point on the horizontal bottom side of the character frame, and t indicates the starting point is a point on the horizontal top side of the character frame. represents something that has been. or,
H _b (j), H _t (j), R _b (j), R _t (j), L _b (j),
L _t
When extracting (j), the vertical length ΔY of the character frame was used as a normalization constant for the distance between the scanning start point and the point where the white point changes to the black point. The feature matrix extraction unit 7 uses the eight types of arrays extracted by the feature extraction unit 6, divides each array into M pieces (M is a constant, in this example, M=7), and divides each array into M pieces (M is a constant, in this example, M=7), and An M×8-dimensional feature matrix F is created by calculating the average value of the array values within the division unit.
Extract (m, n). However, m=1,...,M,
n=1,...,8. The identification unit 8 calculates the distance (D) shown by equation (2) between the feature matrix extracted by the feature matrix extraction unit 7 and the standard character mask f(m, n) written in the same format. The category name of the standard character mask that gives the smallest distance is output to the character name output 9.

D=√{(,)−(,)} ² ...(2) (Effect of the invention) As explained above, the feature matrix extracted by the feature matrix extraction unit of this embodiment is based on the strokes of the character figure pattern. It expresses the position, length, direction, etc. of characters, and expresses the characteristics unique to characters. Furthermore, as can be observed in the example shown in FIG. 3, in which two types of character patterns with similar shapes and the arrangement extracted by the feature extraction unit are graphically expressed, local differences in the characters are sufficiently reflected in the arrangement. Therefore, recognition accuracy can be improved. In addition, when creating each array in the feature extraction section, the distance between the point on the side of the character frame where scanning started and the character line is normalized by the size of the character frame in the scanning direction, so it is easy to use for handwritten characters. Since it is possible to absorb variations in font size due to unique differences in writers, highly accurate recognition is possible. Furthermore, since feature extraction from character/graphic patterns is achieved through a simple scanning process, advanced recognition is possible, and the apparatus has the advantage of being able to be miniaturized.

The present invention scans vertically or horizontally a sub-pattern in which stroke components in each direction are extracted from a character figure pattern, and calculates a value obtained by normalizing the distance between a character frame side and a character line by the size of the character frame in the scanning direction. Since it is characterized by the N-th power sum of can.

[Brief explanation of the drawing]

FIG. 1 is an embodiment of the character recognition device of the present invention, FIGS. 2 a to 2e are diagrams showing examples of sub-patterns, and FIG. 3 is a diagram showing examples of characteristics. 1...Photoelectric conversion unit, 2...Pattern register,
3...Line width calculation section, 4...Sub pattern extraction section,
5...Character frame detection unit, 6...Feature extraction unit, 7...
Feature matrix extraction unit, 8...Identification unit, 9...Character name output.

Claims (1)

[Claims] 1 Scan the character/figure pattern in the desired direction,
The sub-patterns are extracted in multiple directions by detecting the cross-section of the character line in the scanning direction and extracting a cross-section whose cross-sectional length is sufficiently longer than the character line width of the character/figure pattern. , scan in a predetermined direction using a point on the side of the character circumscribing frame of the character figure pattern as a starting point, detect the positions of all character lines on the scanning line, and detect the point on the side from which scanning started, and N, which is a value obtained by normalizing the distance to the detected character line by the size of the character circumscribing frame in the predetermined direction.
The process of extracting the sum of products (N is a constant) is performed for each of the sub-patterns, and for each scanning line, with all points on at least two of the four sides of the character circumscribing frame as the starting point. Then, for each sub-pattern, extract the array of the sum of N-th power for each scanning line with the side of the character circumscribing frame as a unit, and divide each of the extracted sum-of-N arrays into M pieces (M is a constant). By dividing and calculating the average value of the N-th power sum in each division unit and extracting it as an M-dimensional feature vector from each array, a group of feature vectors for the character figure pattern is extracted. A character recognition method characterized by recognizing characters by comparing a group of feature vectors with a dictionary expressed in the same format.