JP2827628B2 - Pattern deformation method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a pattern deformation method for deforming a directional feature pattern in character recognition in accordance with a mapping function for horizontal and vertical deformation.

[0002]

2. Description of the Related Art In a feature extraction process for character recognition, the validity of a directional feature is determined, for example, by referring to the IEICE Transactions (D), Vo.
l. J65-D, No. 5, pp. 550-557 (May 1982), Saito, Yamada, Yamamoto: Analysis of Handwritten Kanji by Direction Pattern Matching Method ". In such a method using directional features, classification processing is often performed by superimposing feature patterns after feature extraction, and thus normalization of character pattern size is indispensable processing. Conventionally, a method of performing linear deformation has been used, but it has been shown that a non-linear normalization method for correcting a positional displacement of a stroke of a character pattern is effective for superposition classification processing. Such a nonlinear normalization method includes, for example, “Transactions of the Institute of Electronics and Communication Engineers (D), Vol.
D, No. 11, pp. 1379-1383 (Showa 59
November), Yamada, Saito, Yamamoto: Linear density equalization-nonlinear normalization method for correlation method-
No. 29988, “Pattern normalization method” and the like. This normalization method is one of pattern deformation methods for deforming a character pattern so that the stroke intervals of the character are spatially uniform.

Conventionally, normalization processing has been performed on an input binary pattern, and directional features have been extracted from the normalized binary pattern. In this case, the image to be transformed is an input binary pattern. FIG. 4 and FIG.
This will be described in 2. FIG. 4 is a diagram for explaining the normalization by the mapping function in the horizontal direction and the vertical direction. The (i, j) plane in FIG. 4 holds the position information on the original pattern,
It is assumed that the (I, J) plane holds position information on the deformation pattern. Here, as an example for explanation, the original pattern is 2 × 2 pixels, the modified pattern is 4 × 4 pixels, and the length of one side of the pixel is 1. However, the number of pixels of the original pattern and the modified pattern is It is not limited to. One square small area in FIG. 4 corresponds to one pixel, and the coordinates are considered as real numbers. Also, (i,
I) The curve I = φ (i) on the plane and the curve J = ψ (j) on the (j, J) plane are horizontal and vertical mapping functions. By these two mapping functions, a square small area having two points (0, 0) and (1, 1) corresponding to one pixel on the (i, j) plane as vertices is represented by a two-dimensional area on the (I, J) plane. The point (0,
0) and (I1, J1) are mapped to a rectangular area having vertices. However, since the rectangular area and each pixel on the (I, J) plane do not correspond one-to-one, the pixel value of the deformation pattern is not uniquely determined. FIG. 12 is a diagram for explaining the normalization of the input binary pattern. FIG.
An input binary pattern having 1 as a pixel value, and FIG. 12B is a modified pattern. As a method of determining the pixel value of the deformed pattern in this case, each pixel value on the (I, J) plane is inversely mapped on the (i, j) plane at the center point of the square small area representing the pixel. The pixel value was set equal to the pixel value of the square small area on the (i, j) plane including the mapping point. When the input binary pattern is deformed, there is not much problem even if the line width of the area of the pixel value 1 which is the character line portion changes, and conversely, since the continuity is required, the above method is used.

However, in the method of extracting the direction feature after performing the normalization, there is a problem that the shape of the character pattern is changed by the normalization process, and the feature of the direction of the character pattern is not preserved. As a method for solving this problem, there is Japanese Patent Application Laid-Open No. 1-111287 "feature extraction method". In this method, a mapping function and a directional feature for nonlinear normalization are independently obtained in advance from an input binary pattern, and then the directional feature is normalized using the mapping function. Thereby, the displacement of the stroke is corrected, and the directional characteristics of the character pattern are preserved.

[0005]

When the pixel to be transformed is an input binary pattern, the pattern transformation is realized by a method of obtaining the pixel value of the transformation pattern from the pixel value of the corresponding pixel by inverse mapping of the coordinate values. However, in the feature extraction method in which the mapping function and the directional feature for nonlinear normalization are independently obtained from the input binary pattern, and then the directional feature is normalized using the mapping function, the directional feature is normalized. Because of the pattern, the following problems occur in the same modification method as in the related art. The directional feature pattern is a line figure having a line width of 1 in which a feature appears only in the contour portion because the directional feature pattern is obtained by local directionality of the contour portion of the character pattern. If the pattern is normalized by the above-described deformation method, the line width may be doubled or tripled due to the inclination of the mapping function, and sometimes the line width may be lost. Would. Since the transformed direction feature pattern is directly reflected on the feature pattern used for the classification process, the recognition ability is reduced as a result.

SUMMARY OF THE INVENTION An object of the present invention is to provide a pattern deformation method which realizes normalization for correcting a positional shift of a directional feature without largely changing a feature amount indicating a degree of direction in a local region of the directional feature pattern. Is to do.

[0007]

According to a first aspect of the present invention, a multi-valued directional feature pattern having a component representing a directional property corresponding to each pixel of an image as a component in accordance with a horizontal and vertical transformation mapping function. In the pattern deformation method for deformation, a direction feature pattern information storage unit that stores information of a direction feature pattern, a mapping function storage unit that stores a mapping function for horizontal and vertical deformation, and a direction feature pattern information storage unit. Pixel midline calculation means for reading information of the direction characteristic pattern and obtaining information representing a midline passing through the center of each pixel and parallel to the direction of the direction characteristic; A line mapping means for obtaining information representing a mapping line obtained by mapping a line on a deformation pattern by a horizontal and vertical deformation mapping function obtained from the mapping function storage means; An intersection coordinate calculating means for obtaining a line group perpendicular to the direction of the midline before mapping on the above, and obtaining coordinates of an intersection of the line group and the mapping line obtained by the midline mapping means; and A pixel value determining means for determining a pixel value of a deformed pattern from the intersection coordinates obtained by the calculating means and a pixel value of the directional characteristic pattern obtained from the directional characteristic pattern information storing means; and a deformation obtained by the pixel value determining means. And a modified pattern storage unit for storing a pattern.

According to a second aspect of the present invention, there is provided a multi-valued directional feature pattern having a component representing a directional property corresponding to each pixel of an image as a component.
In a pattern deformation method for deforming according to a horizontal and vertical transformation function, a direction feature pattern information storage means for storing information of a direction feature pattern, and a mapping function storage for storing a horizontal and vertical transformation function. Means, pixel direction coordinate information calculating means for reading direction characteristic pattern information from the direction characteristic pattern information storage means, and obtaining the coordinates of the center point of each pixel; and the central point of each pixel obtained by the pixel center point coordinate calculating means. A center point mapping means for calculating coordinates of a mapping point mapped on a deformation pattern by a horizontal and vertical deformation mapping function obtained from the mapping function storage means, and a direction obtained from the direction feature pattern information storage means. Weights corresponding to the inclinations of the horizontal and vertical transformation mapping functions obtained by the mapping function storage means are assigned to the pixel values of the feature pattern. A mapping pixel value calculating means for obtaining a multiplied mapping pixel value; a mapping point coordinate obtained by the center point mapping means; and a pixel for determining a pixel value of a deformation pattern from the mapping pixel value obtained by the mapping pixel value calculating means. Value determining means;
And a modified pattern storage unit for storing the modified pattern obtained by the pixel value determining unit.

According to a third aspect of the present invention, a multi-valued directional feature pattern having a component representing a directional property corresponding to each pixel of an image as a component,
In a pattern deformation method for deforming according to a horizontal and vertical transformation function, a direction feature pattern information storage means for storing information of a direction feature pattern, and a mapping function storage for storing a horizontal and vertical transformation function. Means, pixel direction coordinate information calculating means for reading direction characteristic pattern information from the direction characteristic pattern information storage means, and obtaining the coordinates of the center point of each pixel; and the central point of each pixel obtained by the pixel center point coordinate calculating means. The center point mapping means for obtaining the coordinates of a mapping point mapped on a deformation pattern with the horizontal and vertical deformation mapping functions obtained from the mapping function storage means, and the mapping point coordinates obtained by the center point mapping means. And a pixel value determining means for determining a pixel value of a deformation pattern from a pixel value of the direction characteristic pattern obtained from the direction characteristic pattern information storage means. Characterized by comprising a deformation pattern storage means for storing a deformation pattern obtained by the pixel value determining means.

[0010]

The processing procedure for pattern deformation according to the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram showing a processing procedure of the pattern deformation method of the first invention. FIG. 5 is a diagram for explaining the pattern deformation according to the first invention, and is a case of a horizontal component. The same procedure can be used for other directions.

The direction feature pattern information storage means 1 in FIG.
It stores the pixel value indicating the degree of the direction of the multi-valued direction feature pattern, the coordinates of the pixel, and the direction of the direction feature indicated by the pixel. FIG. 5A shows an example of a direction feature pattern, in which two pixels have pixel values of A and B, respectively,
It is. The direction feature is generally obtained from a local pattern of the outline of the character line, and is represented by multiple values according to the direction. Photo
The image function storage means 2 stores a horizontal mapping function I = φ (i)
And information of a vertical mapping function J = ψ (j) is stored.
Things. This horizontal and vertical mapping function
Is the same as the mapping function used in the prior art.

First, the pixel middle line calculation means 3 of FIG. 1 reads from the directional feature pattern information storage means 1 the pixel value and coordinates of each pixel of the directional feature pattern and information on the direction of the directional feature represented by the pixel. Is obtained, which represents a center line that passes through the center of and is parallel to the direction of the direction feature. In FIG. 5A, for a pixel having pixel values A and B, a horizontal middle line 51 passing through the center of the pixel is drawn. The equation of the middle line 51 is j = 1.5. For a pixel of 0, even if it is mapped, it will be 0, and its explanation is omitted.

Next, the midline mapping means 4 transforms the midline of each pixel obtained by the pixel midline calculation means 3 with the horizontal and vertical deformation mapping functions obtained from the mapping function storage means 2. Information representing the mapping line mapped above is obtained. A dotted line 52 in FIG. 5B is a diagram obtained by mapping the middle line 51 with the horizontal and vertical mapping functions shown in FIG. The equation of the mapping line 52 is J = ψ (1.5).

Then, the intersection coordinate calculating means 5 calculates the midline mapping.
The mapping line obtained by means 4 and before mapping on the deformation pattern
Find the coordinates of the intersection with the line group perpendicular to the direction of the midline
You. As shown in FIG. 5 (c), the vertical direction
, And an intersection 54 with the mapping line 52 is obtained.
This line group 53 may have any number of lines,
Figure 8 shows a typical example when sampling in the direction
You. FIGS. 8A and 8B are horizontal directions, and FIGS.
(D) is a vertical direction, and FIGS. 8 (e) and (f) are right-upward directions.
8 (g) and 8 (h) are perpendicular to the right-downward direction.
It is an example of a line group of a direction. The more intersections, the finer the intersection
Can be calculated very well, but on the contrary, it takes time,
And speed. In FIG. 5 (c), FIG.
Using the line group, four intersections are obtained. Line group 53 equation
Is represented by I = n + 0.5 (n = 0, 1, 2, 3)
And the coordinates of the mapping line 52 and the intersection 54 are (n + 0.5,ψ
(1.5)) (n = 0, 1, 2, 3).

Finally, the pixel value determination means 6 determines the pixel value of the deformation pattern from the intersection coordinates obtained by the intersection coordinate calculation means 5 and the pixel value of the direction feature pattern obtained from the direction feature pattern information storage means 1. I do. In FIG. 5D, the pixel value of the pixel on the directional feature pattern including the point when the intersection is inversely mapped on the directional feature pattern is entered in the pixel of the deformation pattern including each intersection 54. In FIG. 5D, since only one intersection exists in one pixel, the pixel value of the corresponding directional feature pattern becomes the pixel value of the deformation pattern as it is. However, when multiple intersections are obtained for one pixel, the pixel value corresponding to each intersection is
The pixel value of the final deformation pattern is determined by the average, the maximum value, the addition, and the like.

According to the above procedure, the directional feature pattern can be transformed using the mapping functions in the horizontal and vertical directions. In this procedure, one pixel of the directional feature pattern is mapped as a line having a width of 1 on the deformed pattern, and the transformed directional feature pattern having a line width of 1 is stably obtained without being affected by the inclination of the mapping function. Can be. In the first invention, the pixels are linearly mapped, so that at the same time as the position correction,
The correction is made so that the deviation of the characteristic value caused by the variation of the shape of various input patterns is also eliminated. The same applies to the vertical direction, the upward direction, and the downward direction.

FIG. 2 is a block diagram showing a processing procedure of the pattern deformation method of the second invention. Direction feature pattern information storage means 1, mapping function storage means 2, pixel value determination means 6
And the modified pattern storage means 7 are the same as the respective components of the pattern modified system of FIG. FIG. 6 is a diagram for explaining a pattern deformation according to the second invention.
Similarly to the case of the horizontal component.

First, the pixel center point coordinate calculating means 8 reads information on the pixel value and coordinates of each pixel of the directional feature pattern from the directional feature pattern information storage means 1 and obtains the coordinates of the center point of each pixel. In FIG. 6A, two pixels have pixel values A and B, respectively. For example, the coordinates of the center point 61 of the pixel having the pixel value A are (0.5, 1.5).

Next, the center point mapping means 9 sets the pixel center point
The center point of each pixel obtained by the target calculation means 8 is described by a mapping function notation.
For horizontal and vertical deformation obtained by storage means 2
Calculate the coordinates of the mapping point mapped on the deformation pattern using the mapping function
You. FIG. 6 (b) is a mapping of the center point 61 of FIG. 6 (a).
And its coordinates are (φ (0.5),ψ
(1.5)).

Further, the mapping pixel value calculation means 10 adds the horizontal and vertical transformation mapping functions obtained from the mapping function storage means 2 to each pixel value of the directional feature pattern obtained from the directional feature pattern information storage means 1. Is obtained by multiplying a mapping pixel value by weighting according to the inclination of the pixel. For example, FIG.
Is the weight applied to the mapping point 62 in φ (1) −φ
(0) = I ₁ −0 = 1.8. The weights in the upward and downward directions can be realized by calculating an amount representing the inclination in the oblique direction from the mapping function in the horizontal direction and the vertical direction, for example, by using the theorem of three squares, or by a simple average. It may be approximated.

Finally, the pixel value determining means 6 determines the pixel value of the deformation pattern from the center point coordinates obtained by the pixel center point coordinate calculating means 8 and the mapped pixel values obtained by the mapped pixel value calculating means 10. . The deformation pattern is as shown in FIG.

In the first invention, the pixels are mapped linearly, whereas in the second invention, the lines are gathered and approximated to one pixel. The same applies to the vertical direction, the upward direction, and the downward direction.

FIG. 3 is a block diagram showing a processing procedure of the pattern deformation method of the third invention. Direction feature pattern information storage means 1, mapping function storage means 2, pixel center point coordinate calculation means 8, center point mapping means 9, pixel value determination means 6,
The modified pattern storage means 7 is the same as each component of the pattern modified system of FIG. FIG. 7 is a diagram for explaining the pattern deformation according to the third invention, and shows a case of a horizontal component. As in FIG. 6, FIG. 7A shows the direction feature pattern, and FIG. 7B shows the mapping of the center point.

The difference from the second aspect is that the pixel value input by the pixel value determining means 6 is a pixel value obtained from the direction feature pattern information storage means 1. Therefore, the pixel value of the direction feature pattern is directly applied to the pixel on the deformation pattern including the mapped center point without weighting the pixel value of the direction feature pattern. As a result, the pattern shown in FIG.

The third aspect of the present invention corresponds to the fact that only the position of the directional feature is corrected and the feature is mapped as it is.
The same applies to the vertical direction, the upward direction, and the downward direction, and there is no difference in processing depending on the direction.

[0026]

DESCRIPTION OF THE PREFERRED EMBODIMENTS In the first invention, the characteristics of the deformation pattern of the upward-sloping component and the downward-sloping component are different depending on how to take a line group for finding the intersection. I'll give it.

One is that, as shown in FIGS. 8 (e) and 8 (g), the line group passes through the center point of the pixel of the deformation pattern, so that the pixel values to be mapped can maintain 4-connection. Features. FIG. 9 is a diagram for explaining the pattern deformation in the case of the component rising in the rightward direction, and FIG.
Is used. As in FIG. 5, FIG. 9A shows the direction feature pattern, FIG. 9B shows the mapping of the middle line, and FIG. 9C shows the intersection.
When the group of lines in FIG. 8E is used, since the mapped center line always has an intersection even if the pixel is slightly touched, quadruple connection is guaranteed and the image is mapped as shown in FIG. 9D.

The other is, as shown in FIGS. 8 (f) and 8 (h), a center line mapped by passing a line group through the midpoint of four sides of a pixel represented by a square of a deformed pattern. Is a straight line, it is characterized in that eight connections can be maintained in that range. FIG. 10 is a diagram for explaining the pattern deformation in the case of the component in the upward-sloping direction similarly to FIG. 9, but FIG. 8F is used for the line group. As in FIG. 9, FIG. 10A shows the direction feature pattern, FIG. 10B shows the mapping of the middle line, and FIG. 10C shows the intersection. In the case where the line group of FIG. 8F is used, if an intersection occurs at pixels obliquely adjacent to each other, if the mapping line is a straight line, no intersection occurs between pixels connected in the horizontal and vertical directions during the straight line. In the range corresponding to the mapping of one pixel of the direction feature pattern, 8-connection is guaranteed, and the mapping is performed as shown in FIG. In the figure, when a plurality of intersections are obtained for one pixel, an average is used.

Next, an embodiment applied to an actual character pattern is shown in FIG. FIG. 11A shows an example of an input character pattern, which is a binary pattern captured on a device memory by an input device such as a scanner. The black background is 1 and the white background is 0. FIG. 11B shows a direction feature pattern created by calculating feature amounts in the horizontal direction, the vertical direction, the right-up direction, and the right-down direction corresponding to each pixel from the input binary pattern of FIG. The aforementioned document “Transactions of the Institute of Electronics and Communication Engineers (D), Vol. J65-D, No. 5, pp. 550”
-557 (May 1982), Saito, Yamada, Yamamoto: Analysis of Handwritten Kanji by Direction Pattern Matching Method "and the like. The pixel value is multi-valued.

FIG. 11 (c) shows the conventional technique shown in FIG.
This is a case where the method of normalizing the input binary pattern described with reference to FIG. 2 is directly applied to the normalization of the direction feature pattern in FIG. In this method, there is a portion where the line width of the direction feature is partially large. Even with the same character type, handwritten characters vary widely, and the mapping function changes accordingly. Therefore, the line width of the direction feature varies depending on the character,
It is inconvenient to use this feature for classification.

FIG. 11D shows the first invention in FIG.
This is a case where the method is applied to the normalization of the direction feature pattern of (b). In this method, the line width is changed to 1. Even for handwritten characters, patterns of the same character type are likely to be transformed into close patterns.

FIG. 11E shows a case where the second or third invention is applied to the normalization of the direction feature pattern shown in FIG. 11B. In the second invention and the third invention, the mapped position is the same, but the pixel value is different. In this scheme, the lines are discontinuous. However, in practice, this directional feature pattern is not used directly in the classification process, and a compression process of, for example, combining about 64 pixels into one element is performed to reduce the number of dimensions, so that sufficient continuity is not maintained. There are good results.

[0033]

The results of simple classification experiments using the conventional method (the method for obtaining FIG. 11 (c)) and the first to third inventions are shown below. After transforming the directional feature pattern using each method, 8 × 8 pixels were combined into one element and subjected to a blurring process and a vector normalization process to obtain a feature vector used for classification. In the classification, calculate the city block distance,
The closest pattern was the result. For the data, we used the JIS first-level handwritten kanji database provided by the Electronic Technology Research Institute,
A reference pattern serving as a dictionary was created by averaging the feature vectors of 0 sets, and 10 sets were selected from even numbers to perform a classification experiment. The table shows the average classification rate of 30,360 characters.

[Table 1]

From the results shown in the table, it can be seen that the first to third inventions are much more effective than the conventional method. The recognition performance is improved in the order of the third, second, and first inventions.
The second aspect of the present invention has a feature that it can be executed at a high speed because the processing is simpler, though the accuracy is slightly lower than that of the first aspect. Also, the first invention maps linearly, and the second invention maps by weighting so that the deviation of the characteristic values caused by the variation of the shape of the input pattern is corrected. In the third invention, the direction feature obtained from the input pattern is mapped as it is. Therefore, the third aspect of the present invention is characterized in that although the accuracy is slightly lower than those of the first and second aspects, the processing can be executed at a higher speed.

As described above, the pattern deformation method of the present invention
It can be applied to the normalization of direction feature patterns in character recognition,
Furthermore, it is possible to realize normalization for correcting misregistration of directional features without greatly changing the feature amount indicating the degree of direction in a local region of the directional feature pattern, which is greatly useful for improving the performance of character recognition. .

[Brief description of the drawings]

FIG. 1 is a block diagram showing a processing procedure of a pattern deformation method according to the first invention.

FIG. 2 is a block diagram showing a processing procedure of a pattern deformation method according to the second invention.

FIG. 3 is a block diagram showing a processing procedure of a pattern deformation method according to a third invention.

FIG. 4 is a diagram for explaining normalization by a mapping function in a horizontal direction and a vertical direction.

FIG. 5 is a view for explaining pattern deformation according to the first invention.

FIG. 6 is a view for explaining pattern deformation according to the second invention.

FIG. 7 is a view for explaining pattern deformation according to the third invention.

FIG. 8 is a diagram showing an example of a group of lines for obtaining an intersection in the first invention.

FIG. 9 is a diagram for explaining a pattern deformation of a rightward upward component according to the first invention.

FIG. 10 is a diagram for explaining a pattern deformation of a rightward upward component according to the first invention.

FIG. 11 is a diagram showing an example in which the conventional method and the present invention are applied to specific characters.

FIG. 12 is a view for explaining pattern deformation of an input binary pattern.

[Explanation of symbols]

 Reference Signs List 1 direction feature pattern information storage means 2 mapping function storage means 3 pixel midline calculation means 4 midline mapping means 5 intersection coordinate calculation means 6 pixel value determination means 7 modified pattern storage means 8 pixel center point coordinate calculation means 9 center point mapping means DESCRIPTION OF SYMBOLS 10 Mapping pixel value calculation means 51 Middle line 52 Middle line mapping line 53 Line group perpendicular to the direction of the middle line 54 Intersection 61 Center point 62 Center point mapping point

Continuation of front page (56) References JP-A-64-29988 (JP, A) JP-A-3-85686 (JP, A) "Handwritten Kanji Recognition by Parallelization of Normalization and Feature Extraction", Electronic Information Communication Proc. Of the JSCE Autumn Meeting (6), D-235, pp. 6-237, published on August 15, 1991 "On the consistency between nonlinear normalization and feature extraction in handwritten kanji recognition", IEICE Technical Research Report, Vol. 91, No. 317, PRU91-85, p. 55-62, issued on November 14, 1991 (58) Fields investigated (Int. Cl. ⁶ , DB name) G06K 9/62 640 JICST file (JOIS) Patent file (PATOLIS)

Claims (3)

(57) [Claims] 1. A pattern deformation method for deforming a multi-valued directional feature pattern having a feature amount representing a directional property corresponding to each pixel of an image in accordance with a horizontal and vertical deformation mapping function. Directional feature pattern information storage means for storing the information of the directional feature pattern, and a mapping function storage means for storing the mapping function for horizontal and vertical deformation, reading information of the directional feature pattern from the directional feature pattern information storage means, Pixel midline calculation means for obtaining information representing a midline parallel to the direction of the direction feature passing through the center of the pixel, and obtaining the midline of each pixel obtained by the pixel midline calculation means from the mapping function storage means. A midline mapping means for obtaining information representing a mapping line mapped on the deformation pattern by a horizontal and vertical deformation mapping function, and a middle line before mapping on the deformation pattern. Intersection coordinate calculation means for obtaining a line group perpendicular to the intersection line, and obtaining the coordinates of the intersection of the line group and the line group obtained by the midline mapping means, Intersection coordinates obtained by the intersection coordinate calculation means, A pixel value determining unit that determines a pixel value of the modified pattern from a pixel value of the directional characteristic pattern obtained from the directional characteristic pattern information storing unit; and a modified pattern storage unit that stores the modified pattern obtained by the pixel value determining unit. A pattern deformation method characterized by becoming. 2. A pattern deformation method for deforming a multi-valued directional feature pattern having a feature amount representing a directional property corresponding to each pixel of an image in accordance with a horizontal and vertical transformation mapping function. Direction feature pattern information storage means for storing the information of the horizontal direction and the vertical direction, and a mapping function storage means for storing the mapping function for transformation in the horizontal direction and the vertical direction; and reading the direction feature pattern information from the direction feature pattern information storage means, Pixel center point coordinate calculation means for obtaining the coordinates of the center point, and the center point of each pixel obtained by the pixel center point coordinate calculation means is calculated by a horizontal and vertical transformation mapping function obtained from the mapping function storage means. A center point mapping means for obtaining coordinates of a mapping point mapped on a deformation pattern; and a pixel value of a direction feature pattern obtained from the direction feature pattern information storage means. Mapping pixel value calculation means for obtaining a mapping pixel value weighted according to the inclination of the horizontal and vertical transformation mapping functions obtained by the mapping function storage means; and a mapping point obtained by the center point mapping means. From the coordinates and the mapped pixel value obtained by the mapped pixel value calculating means, a pixel value determining means for determining the pixel value of the deformed pattern, and a deformed pattern storage means for storing the deformed pattern obtained by the pixel value determining means. A pattern deformation method characterized by becoming. 3. A pattern deformation method for deforming a multi-valued directional feature pattern having a feature amount representing a directionality corresponding to each pixel of an image in accordance with a horizontal and vertical transformation mapping function. Direction feature pattern information storage means for storing the information of the horizontal direction and the vertical direction, and a mapping function storage means for storing the mapping function for transformation in the horizontal direction and the vertical direction; and reading the direction feature pattern information from the direction feature pattern information storage means, Pixel center point coordinate calculation means for obtaining the coordinates of the center point, and the center point of each pixel obtained by the pixel center point coordinate calculation means is calculated by a horizontal and vertical transformation mapping function obtained from the mapping function storage means. Center point mapping means for obtaining coordinates of a mapping point mapped on a deformed pattern; mapping point coordinates obtained by the center point mapping means; and directional feature pattern information. A pixel value determining unit that determines a pixel value of the deformation pattern from a pixel value of the direction feature pattern obtained by the storage unit; and a deformation pattern storage unit that stores the deformation pattern obtained by the pixel value determination unit. Pattern deformation method.