JPH07109612B2 - Image processing method - Google Patents

Description

The present invention relates to an image processing method for extracting features from pattern data.

[Summary of Disclosure] The present specification and the drawings show an image processing method for extracting strokes in a predetermined direction from a graphic pattern, and particularly in a pattern feature extraction apparatus, an original graphic pattern is parallel to a direction deviated from the predetermined direction by a predetermined angle. By using a two-dimensional pattern obtained by performing a predetermined logical operation between the figure pattern obtained by moving and the pixels of the original figure pattern as the stroke of the original figure pattern in the predetermined direction, a small processing procedure and a small device Disclosed is a technique for reducing the influence of character line width variation and character position variation of an original figure on a scale, and extracting a feature of a pattern.

[Prior Art] Conventionally, when performing identification determination processing in, for example, an optical character recognition device by reading and recognizing fixed characters such as printed characters,
A so-called template matching method has been widely adopted in which an input figure and a standard character figure stored in a dictionary memory are superposed and a standard character having the maximum degree of overlap is output as a recognition result. However, the template matching method naturally requires accurate alignment between the input figure pattern and the dictionary figure pattern, and if the character line width changes, the degree of overlapping of figures is greatly affected, so the input figure is obtained. Therefore, it was necessary to carefully binarize the original figure. Further, since the two-dimensional figures are matched with each other, it is necessary to have the dictionary pattern as a two-dimensional pattern, which causes a drawback that the dictionary memory capacity increases.

In order to make up for the above-mentioned drawback, the dictionary pattern is not just a binary template, but between the black pixel area and the white pixel area,
There is a case where a transition region that allows either white or black is provided and a three-valued template is used, but the dictionary memory capacity is further increased, which is inappropriate.

Further, in order to reduce the influence of the line width variation of the input characters and to incorporate the structural features of the figure, the strokes in a specific direction are extracted from the given figure pattern, and the mutual relationship between the extracted strokes is used to extract the strokes. It is necessary to find the feature quantity of the figure. In that case, for example, in the case of a character recognition device, the following stroke extraction method is generally widely used. It is a method of tracing a contour of a character figure to detect and extract a predetermined stroke from the curvature, inclination and the like of the detected contour point series, or the character figure as shown in FIGS. 2 (a) and 2 (b). In this method, a contour component having a specific inclination is detected and extracted by scanning with a 3 × 3 differential filter.

However, in the former method, when the character / graphic pattern becomes large or becomes complicated, the processing amount of contour tracking increases,
Further, there is a drawback when the stroke is detected for the first time by further performing processing for obtaining the inclination, the length, and the connection relationship on the detected contour point series. Further, the latter method has a drawback that it takes a lot of time to scan the filter, and the processing time increases dramatically as the image size increases.

[Problems to be Solved by the Invention] The above-mentioned problems of the conventional example can be attributed to the problem of how to efficiently extract the characteristics of the graphic even if the input graphic varies. Therefore, an object of the present invention is to provide an image processing method capable of reducing the influence of character line width variation and character position variation with a small processing procedure and a small device scale.

[Means for Solving Problems] The configuration of the image processing method of the present invention for achieving the above-mentioned object is to create a figure pattern by shifting an original figure pattern in a predetermined direction, and to shift the original figure pattern and the shift. Pattern information is generated by logically operating the figure pattern, and the number of black pixels obtained by counting the black pixels of the pattern information is set as characteristic data representing the characteristic of the original figure pattern in the predetermined direction, The feature data is derived in a plurality of directions, and the original figure pattern is recognized using the feature data obtained in each of the plurality of directions and direction information representing the direction. .

[Operation] In the image processing method of the above configuration, the number of black pixels in the pattern information obtained by logically operating the original figure pattern and the figure pattern obtained by shifting the original figure pattern is used as the characteristic data. Further, the characteristic data is extracted in a plurality of directions.

Embodiments Embodiments of the present invention will be described in more detail below with reference to the accompanying drawings.

When the present invention is applied to a pattern feature extraction apparatus, it becomes as shown in FIG. 1, for example. The pattern feature extracting apparatus of the embodiment shown in FIG. 1 comprises a parallel moving means 101 for moving the original figure pattern 100 in a direction parallel to the stroke direction to be extracted by a predetermined angle, and a predetermined logic between pixels. It is composed of a logical operation means 102 for performing an operation.

In the structure shown in FIG. 1, the pattern parallel moving means 10
Stroke information to be extracted is obtained by performing a predetermined logical operation, for example, exclusive OR operation, on the pixels of the pattern 103 obtained by 1 and the original figure pattern 100.

FIG. 3 shows an example of the configuration when the present invention is applied to a character recognition device. The character image printed on the form 1, which is the original image, is illuminated by an appropriate light source, and is imaged on the photoelectric converter 3 by the lens 2. The analog data output from the photoelectric converter 3 is converted into a digital signal by the A / D converter 4 and converted into a binary signal by the binarizing circuit 5 and input to the character extracting circuit 6. The character cutout circuit 6 cuts out a partial image corresponding to one character on the form and outputs the partial image to the normalization circuit 7. The normalization circuit 7 sends an image obtained by enlarging or reducing this one-character image to a reference size to the feature extraction circuit 8; however, when the character size on the form is unified in advance to one type, The normalization circuit 7 is not necessary. The feature extraction circuit 8 extracts a feature amount, which will be described later, from the input character image, and the character recognition unit 12 performs character recognition. The character recognition unit 12 calculates the degree of difference with the dictionary pattern in the dictionary memory 10 by using the degree-of-difference calculation circuit 9 for example, and
The identification circuit detects the dictionary pattern that minimizes the difference and outputs it as the recognition result.

Next, the processing procedure from the feature extraction of the character graphic, which is the main focus of the present embodiment, to the determination of the determination will be described in detail. The process from the feature extraction to the discrimination determination corresponds to the feature extraction circuit 8 to the discrimination circuit 11 in FIG. First, the feature extraction circuit 8 performs character stroke extraction using the following means.

FIGS. 4 to 9 (a) and (b) show the results obtained by inputting the input character graphic 22 to the feature extraction circuit 8 and performing the process according to the feature extraction step by step. is there. Fig. 4-22
The input pattern is like a character box 21, for example 8 bits x
It is assumed that the normalization circuit 7 has normalized so that the size is 12 bits. It should be noted that FIG. 4 shows the number “1” of the form 1 in FIG. 2, the shaded area is a black pixel (bit value = 1), and the blank area is a white pixel (bit value = 0).
The processing procedure of feature extraction will be described below step by step with reference to the drawings.

First, as shown in FIG. 4, the coordinates x and y axes are taken, and each pixel in the pattern is set to M (x, y). However, (x = 1,2, ..., 8,
y = 1,2, ..., 12). That is, M (x, y) = 1 (x, y) is a black pixel 0 (x, y) is a white pixel (1).

5 (a) and 5 (b) show the pixel pattern M (x,
y) in the upper left (-45 degrees) and upper right (+45 degrees) directions (5th
The pattern is translated by one pixel in the direction of the arrow in FIGS.

6 (a) and 6 (b) are respectively FIG. 5 (a) and FIG. 4,
The patterns of FIG. 5 (b) and FIG. 4 are overlapped so that the character frames coincide with each other, an exhaustive OR is taken for each pixel, and the result is further translated in the right direction by one pixel. is there. That is, the pixels of FIGS. 6 (a) and 6 (b) are replaced by E1 (x,
y), E2 (x, y), E1 (x, y) = M (x-1, y) M (x, y-1) E2 (x, y) = M (x, y) M It can be expressed as (x-1, y + 1) (2). Here, it is a symbol that represents exhaustive disjunction.
As can be seen from the patterns 41 and 42 of FIGS. 6A and 6B, the contour component of the character pattern was used to obtain the pattern of FIG. 5A in FIG. 6A. The result is that the component parallel to the direction of the -45 degree translation is eliminated (because the original pattern has less -45 degree component), and the component perpendicular thereto (that is, +45 degree component) is emphasized. . On the other hand, in FIG. 6 (b), the parallel movement direction (-
The component parallel to (45 degrees) is erased.

Next, FIG. 7 is obtained by overlapping the character frames of the patterns of FIGS. 6 (a) and 6 (b) so as to overlap each other and taking the logical product (AND) of the pixels. The pattern 51 of FIG. 7 obtained in this way is oblique (±
It is said that contour lines other than (45 degrees) have been extracted.

Finally, the pattern 51 of FIG. 7 and FIG. 6 (a), and FIG. 7 and FIG. 6 (b) are overlapped with each other so that the character frames are overlapped with each other. As a result, patterns 61 and 62 shown in FIGS. 8A and 8B are obtained. 8th
When the diagrams (a) and (b) are written as F1 (x, y) and F2 (x, y), respectively, F1 (x, y) = E1 (x, y) (E1 (x, y) x E2 ( x, y)) F2 (x, y) = E2 (x, y) (E1 (x, y) × E2 (x, y)) (3) Here, the symbol x indicates an AND operation. By performing the above operation, F1 (x, y) in FIG. 8 (a) and F2 (x, y) in FIG. 8 (b) can be obtained. 61 and 62 correspond to a stroke component that forms an inclination of +45 degrees with the x-axis in F1 (x, y) and a stroke component that forms an inclination of −45 degrees with the x-axis in F2 (x, y). .

9 (a) and 9 (b) are the results of extracting the stroke (71) parallel to the y-axis and the stroke (72) parallel to the x-axis by the same method as described above. That is, first, the original pattern 22
Is parallel-moved by one pixel in parallel with the x-axis to obtain a pattern similar to that in FIG. 5 (a), and similarly, parallelly moved by one pixel in parallel with the y-axis to obtain a pattern similar to that in FIG. 5 (b). obtain. A pattern similar to that shown in FIG. 7 is obtained by taking the exclusive OR of the two patterns thus obtained with the pixel of the original pattern 22 and further taking the logical product. The pattern similar to that shown in FIG. 7 should be that the ± 45 ° component is extracted. Therefore, when exclusive OR is performed by the same method as that used to obtain the patterns of FIGS. 7 (a) and 7 (b), patterns 71 and 72 shown in FIGS. 9 (a) and 9 (b) are obtained. A stroke component parallel to the y axis and the x axis is obtained.

As described above, according to the feature extraction circuit of the present embodiment, with respect to the two-dimensional character pattern, the shift operation is performed several times and the simple logical operation is performed so that the +45 degree direction, the −45 degree direction, and the x axis of the x axis are changed. Strokes in four directions in the parallel direction and the direction parallel to the y-axis (that is, strokes in eight directions of 0 degree, 45 degrees, 90 degrees, 135 degrees, 180 degrees, 225 degrees, 270 degrees, 315 degrees). Can be extracted in the form of the density of black pixels in the place where
And (b) and patterns 61, 62, 7 of FIGS. 9 (a), (b).
The feature amount of the character pattern can be obtained by counting 1,72 black pixel distributions.

FIG. 10 shows an example of means for obtaining the feature amount of the character pattern from the stroke extraction image. In the example of FIG. 10, the pattern 61 in which strokes forming 45 degrees with the x-axis of FIG. 8 (a) are extracted is converted into a frequency distribution 81 in the y-axis direction, and the y-axis is 82,8.
It is divided into three equal areas such as 3,84 and the total value of the frequencies in each area is counted. If the frequency distribution is 81, this count value is "7", "3", respectively for each area of 82, 83, 84.
It becomes "1". When the same processing is performed on the three images 62, 71, 72 of FIGS. 8 (b), 9 (a), and (b), each count value is "2" in FIG. 8 (b). , "1", "0", Fig. 9 (a)
"5", "11", "6", and "3", "3", "2" in FIG. 9 (b). Since twelve numerical sequences are obtained for the above four images, a 12-dimensional vector in which the numerical sequences are arranged in order can be used as the feature amount of the input pattern (hereinafter referred to as the input vector).

In the dissimilarity detection circuit 9 of FIG. 3, the stroke is previously vectorized for the reference character and the vector is stored in the dictionary memory 10 by the same method as that for obtaining the input vector. The degree of difference from the vector stored in the dictionary memory 10 (hereinafter referred to as a dictionary vector) is obtained by a known method such as the Euclidean distance between vectors or the absolute value distance. The discrimination circuit 11 selects the smallest one (least squares method) from the above-obtained dissimilarities for all the character categories stored in the dictionary memory, and outputs the corresponding dictionary category name as the recognition result.

FIG. 11 is an example of the configuration of the feature extraction circuit of the present invention. 20
A pattern for one character after binarization processing and normalization processing is input to the 1 pattern register. 20 from this pattern
While controlling it with the W / R controller of No. 2, an image in which a predetermined number of bits are translated is generated in the pattern registers 203 to 206. The pattern register is, for example, RAM (Random Acce
If configured with ss Memory), the W / R controller 202 can be configured with something like an address counter.

The patterns stored in the pattern registers 201 and 203 to 206 are input to the logical operation circuit 207, and (2) to
After performing a predetermined logical operation process such as equation (3),
Output to the counter of 9 to 212. That is, the logical operation circuit 207
The output from the counter is counted by the counters 209 to 212 controlled by the controller 208 for the number of pixels in the specific partial area, and the count value is stored in the memory 213. That is, memory 2
Each element of the 12-dimensional vector described above is stored in 13.

As described above, when the features of the embodiments according to the present invention are enumerated, the strokes of a character pattern can be efficiently extracted as a pixel distribution with a simple device configuration. Further, since the presence or absence of strokes is expressed by the pixel density, it is possible to easily vectorize the pattern by counting the frequency. Therefore, since the dictionary pattern to be compared may be stored as a vector instead of as a two-dimensional pattern, the storage capacity required for storage can be small. Further, since the contour line of the character is extracted as a stroke, it is possible to reduce the variation of the character line width, that is, the reduction of the recognition rate due to the variation of the binarization threshold, and it is possible to incorporate the structural feature of the character. . Also, since sufficient recognition ability can be obtained without increasing the number of region divisions when vectorizing the pattern,
It is also possible to reduce a reduction in recognition rate due to a character position change.

Next, several modified examples of the embodiment will be described below. First
FIG. 12 shows another example of the feature extraction circuit 8. Comparing FIG. 12 with FIG. 11, in FIG. 12, the output of the pattern register 301 is transferred to the pattern registers 304 to 307 via the AND gate in order to move the data from the pattern register 301 in parallel by the control of the counter 302. input. Further, the logical operation circuit 207 is specifically configured by an AND gate and an EX-OR (exclusive OR) gate, and the counters 209 to 212 in FIG. 11 are omitted.

Furthermore, in FIGS. 11 and 12, a counter and a pattern register are used for parallel movement of the image, but instead of these, a shift register having bits for the number of pixels required for parallel movement is used in real time. It is also possible to use a method such as transferring data to. In the feature extraction circuit, strokes are extracted in four directions (substantially eight directions) in the above method, but in some cases, four directions or less may be used as long as a sufficient recognition rate can be obtained even in four directions or less. Absent. On the contrary, the parallel movement amount is set to 1 in the x direction in order to increase it in four or more directions.
If there are 2 pixels in the y direction, or 2 pixels in the x direction and 1 pixel in the y direction, then 30 °, 60 °, 120 °, 150 °, 210
The stroke detection direction can be increased, such as degrees, 240 degrees, 300 degrees, and 330 degrees. Further, the number of divided regions shown in FIG. 10 is not limited to three, and the number of divided regions may be increased if the recognition accuracy is improved.

EFFECTS OF THE INVENTION As described above, the image processing method of the present invention uses, as the characteristic data, the number of black pixels in the pattern information obtained by logically operating the original figure pattern and the figure pattern obtained by shifting the original figure pattern. It is adopted. By using such feature data, so-called stepwise determination can be performed, and the recognition accuracy for various patterns becomes high.

Further, by extracting the characteristic data in a plurality of directions, even if there is a variation in the line width or position of the metaphor original drawing pattern, the influence of the variation on the pattern recognition can be minimized and the recognition accuracy can be improved. .

[Brief description of drawings]

FIG. 1 is a basic configuration diagram of an embodiment according to the present invention, and FIGS. 2 (a) and 2 (b) are 3 × 3 differential filters used for contour line detection in a conventional example, and FIG. Is applied to a character recognition device, a configuration example of the character recognition device, FIG. 4, FIG. 5 (a), (b), FIG. 6 (a),
(B), FIG. 7, FIG. 8 (a), (b) and FIG. 9 (a), (b) are diagrams showing, in order, images processed for feature extraction in the embodiment. , FIG. 10 is a diagram for explaining the vectorization of the features of the extracted pattern, FIG. 11 is a diagram showing an example of the circuit configuration of the feature extraction circuit, and FIG. 12 is a modification of the circuit configuration of the feature extraction circuit. It is a figure. In the figure 1 …… Form (original image), 2 …… Lens, 3 …… Photoelectric converter, 4 …… A / D converter, 5 …… Binarization circuit, 6 …… Character cutout circuit, 7 …… Normalization circuit, 8 ... Feature extraction circuit, 9 ... Dissimilarity calculation circuit, 10 ... Dictionary memory, 11 ...
Identification circuit, 12 ... Recognition circuit, 21 ... Character frame, 31,32,41,4
2,51,61,62,71,72 …… Extraction pattern, 81 …… Density frequency,
82-84 …… Divided area, 201,203-206 …… Pattern register, 301,303-311 …… Pattern register, 202 …… W / R controller, 207 …… Logic operation circuit, 208 …… Controller, 209–212 …… Counter , 213 ... memory, 302 ... counter.

Claims (2)

[Claims] 1. A graphic pattern is created by shifting an original graphic pattern in a predetermined direction, logical operation is performed on the original graphic pattern and the shifted graphic pattern to generate pattern information, and a black pixel of the pattern information is calculated. The number of black pixels obtained by counting is used as characteristic data representing the characteristic of the original figure pattern in the predetermined direction, the characteristic data is derived in a plurality of directions, and obtained in each of the plurality of directions. An image processing method, characterized in that the original figure pattern is recognized using characteristic data and direction information indicating the direction thereof. 2. The counting of the number of black pixels used as the characteristic data is performed within a predetermined area of the pattern information, and in the recognition step, the original figure pattern is recognized including area information representing the predetermined area. The image processing method according to claim 1, wherein