JP2595496B2 - Character classification method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a classification system for multi-character characters such as kanji, hiragana, katakana, and alphanumeric characters.

In the conventional character recognition method that reads kanji in addition to alphanumeric characters, kana, and symbols, the number of characters to be read is large, so the first step is to perform classification from the entire target character type and directly output a determination result. There is a demand for a hierarchical classification method in which a large classification is performed, and the next classification is performed by narrowing down candidates.
The large classification method in this case is relatively easy to realize first because there are many objects to be processed, and an efficient method has been desired.

The conventional classification method outputs the narrowed candidate character types based on the above-mentioned purpose, and aims to improve the efficiency of the subsequent recognition processing.
Little is known. One object of the present invention is to perform matching processing between an input character pattern and each standard character pattern and output narrowed candidate character types, but when performing the matching processing, Using the correspondence between the input character pattern and each standard character pattern obtained at the same time, a correction pattern is generated by correcting the input character pattern so as to correspond to each standard character pattern. Outputting is the second object, and in the subsequent recognition processing, the effect of improving the accuracy of recognition is obtained by using a correction pattern for each candidate character type. Further, when the efficiency of the recognition process is not required, all the character types may be output as candidate characters, and the object may be to merely improve the recognition accuracy.

SUMMARY OF THE INVENTION An object of the present invention is to provide a character classification method which can obtain the above-described effects of improving the efficiency of recognition processing and improving the degree of recognition.

Hereinafter, the present invention will be described in detail with reference to the drawings.

FIG. 1 is a block diagram showing an example of the configuration of the present invention. Reference numeral 1 denotes a local optimum matching unit which reads a quantized input character pattern as a signal 10, and sequentially reads a quantized standard character pattern for each character type as a signal 11, and sequentially reads the input character pattern and each standard character pattern. Finding a piecewise optimal correspondence between character patterns, determining whether or not to be a candidate character type based on the value of the degree of difference obtained as a result. The code indicating the character type is sent to the pattern correction unit 2 as a signal 12. The pattern correction unit 2 reads the input character pattern as a signal 10 and, at the same time, reads the information of the optimal correspondence and the code indicating the character type as a signal 12, thereby correcting the pattern for each candidate character type. Then, a code indicating the correction pattern and the candidate character type is output as a signal 13.

FIG. 2 is a block diagram showing an example of the configuration of the local optimum matching section 1. Reference numeral 3 denotes a local pattern extraction unit that inputs an input character pattern as a signal 10, generates a local pattern that does not overlap, and outputs the local pattern as a signal 131 and the position information of each local pattern as a signal 132. Reference numeral 4 denotes a shifted local pattern extraction unit which converts each standard character pattern into a signal 11
And outputs information of a plurality of shifted local patterns and positional shifts corresponding to the positional information of each of the local patterns input as the signal 132 as a signal 141. Reference numeral 5 denotes an optimum matching detection unit which receives the local pattern signal 131, the position information signal 132 of each local pattern, and a plurality of shifted local pattern signals 141 corresponding to the local patterns, and receives the input character pattern. For each of the local patterns, a local pattern that is optimally matched is detected from a plurality of shifted local patterns corresponding to the standard character pattern, and the local pattern is matched when the correspondence between the positional information of the local patterns and the local pattern is matched. A code indicating the difference between the local pattern and the difference between the input character pattern and the standard character pattern from all the differences, and a code indicating the optimal correspondence and the character type of the position information of each local pattern according to the value of the difference Is transmitted as a signal 12 or not.

The example of the configuration of the local optimum matching section 1 has been described above. Next, a logical description of the local optimum matching section 1 will be given.

FIG. 3 is a diagram for logically explaining the local pattern extraction unit 3. FIG. 3A shows an example of an input character pattern and coordinates, and 90 is an 8 × 8 binary pattern. 3 (b), (c), (d) and (e) are local patterns 901, 902, 903 and 904 of the binary pattern 90 and their coordinates. In this example, the input character pattern 90 is divided into two parts both vertically and horizontally. Each local pattern is a 4 × 4 binary pattern.
The local pattern extraction unit 3 converts the input character pattern signal 10 as shown in FIG.
It outputs the local pattern signal 131 and the coordinate-based position information signal 132 shown in (d) and (e), and can be easily realized.

FIG. 4 is a diagram for logically explaining the shifted local pattern extraction unit 4. FIG. 4A shows an example of one standard character pattern and coordinates. Reference numeral 91 denotes an 8 × 8 binary pattern. FIGS. 4 (b), (c), (d) and (e) show a local pattern obtained by dividing the standard character pattern 91 both vertically and horizontally, and a 4 × 4 binary pattern of its coordinates. For the input character pattern 90 and the standard character pattern 91, a local pattern 901
And 911 local patterns 902 and 912, local patterns 903 and 913, and local patterns 904 and 914, respectively.

4 (f), (g), (h), (i), (j),
(K), (l), and (m) show the local pattern obtained by shifting the local pattern 914 up, down, left, and right by one mesh and their coordinates.

In addition, the value of the pixel which has shifted to the outside of the pattern 91 is set to 0.

FIG. 4 shows a local pattern shifted by one mesh as an example of a shifted local pattern. However, a local pattern shifted by two meshes, a local pattern shifted by three meshes, or the like can be created according to the purpose.

Note that the local patterns 911, 912, 913, and 914 are each regarded as a local pattern shifted by 0 mesh. The shifted local pattern extraction unit 4 receives the standard character pattern signal 11 and the position information signal 132 of each local pattern of the input character pattern as shown in FIG. Since a plurality of shifted local patterns as shown in FIG. 4 are output as the signal 141, it can be easily realized.

Subsequently, a logical description of the optimum matching detection unit 5 will be given using an example. In order to obtain the optimum matching with the local pattern shifted up to one mesh between the input character pattern 91 and the local patterns 901, 902, 903 and 904 of the input character pattern 90, a plurality of corresponding local patterns 911 are shifted.
914, 9141 to 9148, a difference is determined in accordance with a predetermined evaluation method, and a combination having the smallest difference is determined as an optimal match. For example, when the exclusive OR is adopted as the evaluation method and the number of pixels whose value obtained as a result of the exclusive OR is 1 is considered as the difference value, the local pattern 90
The shifted local pattern that is optimally matched with 4 is 9146, and the difference is 0. In order to determine the correspondence of the position information, the points (2,2), (2,5), (5,
2) Then, when the coordinates of the point on the input character pattern corresponding to the point (5, 5) are determined, the shifted local pattern 9146 is obtained.
The point on the local pattern 904 of the input character pattern that optimally corresponds to the point (5,5) is the point (5,4). Similarly, FIG. 5 (a) shows a set of points corresponding to other local patterns optimally. FIG. 5 (a) shows that the point (2,2) on the standard character pattern is the point (2,3) on the input character pattern, and the point (5,2) on the standard character pattern is the point (2,3) on the input character pattern. Five,
The point (2,5) on the standard character pattern is the point (2,5) on the input character pattern, and the point (5,5) on the standard character pattern
Indicates that there is an optimal match with the point (5, 4) on the input character pattern.

FIG. 5 (b) shows the correspondence between points on the input character pattern and points on the standard character pattern. For example, points (2,3) on the input character pattern are points on the standard character pattern. (2,2)
The point (2,5) on the input character pattern corresponds to the point (2,5) on the standard character pattern.

FIG. 6 is a block diagram showing an example of the configuration of the optimum matching detection section 5. Reference numeral 50 denotes a local pattern evaluation unit which receives a local pattern signal 131 of an input character pattern, a position information signal 132 of the local pattern, and a local pattern signal 141 of a standard character pattern shifted, and corresponds to a plurality of shifted local patterns. The position information corresponding to the difference value obtained as a result of performing the evaluation is output as a signal 501. Reference numeral 51 denotes a minimum value detection unit which receives the signal 501 as an input, obtains the minimum value of the difference value for each local pattern of the input character pattern, and converts the minimum difference value into a signal 511, the minimum difference. The correspondence of the corresponding position information is output as a signal 512. Numeral 52 is a correspondence relationship storage unit, which inputs the correspondence of the position information corresponding to each local pattern of the input character pattern as a signal 512, stores it one after another, and stores the correspondence of all the position information. In the case of the ON signal, the correspondence of the position information and the candidate character type are output as the signal 12, and are not output when the control signal 514 is the OFF signal. 53 is a dissimilarity storage unit, which is initially cleared to 0, the difference value corresponding to each local pattern of the input character pattern is input as a signal 511, sequentially added, and the difference values corresponding to all the respective local patterns are obtained. When the sum of the values is obtained, the sum of the differences is output as the difference signal 513. 54 is a dissimilarity determination unit, which receives the dissimilarity signal 513 as an input, when the dissimilarity is smaller than a predetermined threshold, sets the output signal 54 to an ON signal, and when the dissimilarity is equal to or greater than the predetermined threshold, Output signal 54 is OF
Output as F signal.

The local pattern evaluation unit 50 of each component described above,
The minimum value detection unit 51, the correspondence relationship storage unit 52, the difference degree storage unit 53, and the difference degree determination unit 54 can be easily realized.

FIG. 7 is a diagram for explaining the pattern correction unit. FIG. 7A shows an input character pattern, and FIG. 7B shows a standard character pattern. It is assumed that a point X on the input character pattern corresponds to a point Y on the standard character pattern. In addition, eight points A, B, C, D, E, and
It is assumed that F, G, and H are defined. At this time, the area enclosed by the quadrilateral ABXH of the input character pattern is linearly transformed into the area enclosed by the quadrilateral ABYH, and the area enclosed by the quadrilateral BCDX is transformed into the quadrilateral XFGH by the area enclosed by the quadrilateral BCDY. The pattern correction is performed by linearly converting the area surrounded by the squares into an area surrounded by the quadrilateral YFGH and the area surrounded by the quadrilateral XDEF into an area surrounded by the quadrilateral YDEF.

FIGS. 7 (c) and 7 (d) show examples in which a finer correction is performed, and each quadrilateral Aij (i = 1,2,3,4: j
= 1,2,3,4) to the quadrilateral Bij (i = 1,2,3,4: j = 1,2,3,4)
The pattern is corrected by performing a linear conversion to. The quadrilateral linear transformation shown in this example is used in computer graphics and the like, and pattern correction can be easily executed if positional information corresponding to a two-dimensional pattern is input. Therefore, the pattern correction unit 2 can also be easily realized.

As described above, according to the present invention, it is possible to perform classification for absorbing a positional shift of an input character pattern and generate a correction pattern of the input character pattern corresponding to the candidate character type.

[Brief description of the drawings]

FIG. 1 is a block diagram showing an example of the configuration of the present invention.
In the figure, reference numeral 1 denotes a local optimum matching unit and a 2 pattern correction unit. FIG. 2 is a block diagram showing an example of the configuration of the local optimum matching section 1. In the figure, 3 is a local pattern extraction unit, 4 is a shifted local pattern extraction unit, and 5 is an optimal matching detection unit. 3 (a) to 3 (e) are diagrams for explaining the local pattern extraction unit 3, in which 90 is a binary pattern, 901, 902, 903, 904.
Are local patterns. FIGS. 4 (a) to (m) are diagrams for explaining the shifted local pattern extraction unit 4, wherein 91 is a binary pattern, 911, 91
2,913,914 is a local pattern, 9141,9142,9143,9144,9145,91
46, 9147 and 9148 are local patterns obtained by shifting the local pattern 914 by one mesh. FIGS. 5 (a) and 5 (b) are diagrams for showing correspondence of position information. FIG. 6 is a block diagram showing an example of the configuration of the optimum matching detection unit 5. In the figure, 50 is a local pattern evaluation unit, 51 is a minimum value detection unit, 52 is a correspondence relationship storage unit, 53 is a difference degree storage unit,
Reference numeral 54 denotes a difference degree determination unit. FIGS. 7A to 7D are diagrams for explaining pattern correction by the piecewise quadrilateral linear conversion performed by the pattern correction unit 2. FIG.

Claims (1)

(57) [Claims] In a method of classifying a quantized input character pattern, a character pattern is input, and a plurality of local patterns obtained by dividing the input character pattern are considered as standard characters in consideration of positional displacement. Performs appropriate matching locally with the pattern, finds the correspondence between the positional shifts in the degree of difference, determines whether or not to be a candidate character type based on the degree of difference, and if it is a candidate character type, a code indicating the character type And a local optimum matching unit that outputs the correspondence between the positional shifts, and the input character pattern, the correspondence between the positional shifts, and a code indicating the character type as input, and the correspondence between the input character patterns and the positional shifts. And a pattern correction unit that generates a correction pattern for each candidate character type and outputs the correction pattern together with the candidate character types, narrows the candidate character types, and outputs a correction pattern of the input character pattern corresponding to each candidate character type. Character classification system which is characterized the door.