JPH04340246A - Method of extracting pattern feature - Google Patents

Abstract

PURPOSE:To extract the feature of a pattern as a whole easily and in a short time by a method wherein small regions are formed of the unit number of picture elements attaching importance to individual picture elements constituting a two-dimensional image data and a position belonging to the pattern in each small region is judged on the basis of the differentiated value of the image data by the individual picture elements in the small region. CONSTITUTION:A two-dimensional image data 1 is formed by arranging many picture elements 2 in a matrix shape. The individual picture elements 2 display the shade, e.g., in 11 scales which are indicated by integers from 0 to 10. Small regions 4 are constituted of nine picture elements 2, 3 (row) X 3 (column). Respective four differentiated values are computed individually for every picture element 2 constituting each small region 4 by using individual parameters in a total of four directions in the longitudinal direction, the +45 deg. direction, the transverse direction and the -45 deg. direction. When the magnitude relationship of the four differentiated values is invenstigated, it is possible to judge in which direction the region is tilted. The position of the small region 4 with reference to patterns 3 can be judged from the tilted direction of the image data of all the picture elements 2 constituting the small region 4.

Description

[Detailed description of the invention]

0001

[Field of Industrial Application] The present invention relates to a pattern feature extraction method, and in particular, the present invention relates to a method for extracting pattern features, and in particular, calculates a pattern in two-dimensional image data by determining the tilt direction of the image data for each local area, and extracting the pattern from the image data tilt direction of each local area. This invention relates to a pattern feature extraction method that extracts the entire pattern.

0002

2. Description of the Related Art For example, in a semiconductor mask or printed circuit board inspection process in a semiconductor manufacturing factory, it is necessary to inspect whether the shape of the manufactured semiconductor mask or printed circuit board matches a design pattern. To perform this inspection process automatically, first, the semiconductor mask or printed circuit board to be measured is read with an imager to obtain two-dimensional image data. Then, a pattern to be measured is extracted from the two-dimensional image data and compared and contrasted with a reference pattern preset in the storage unit.

In the conventional pattern feature extraction method,
This was done using a pattern matching method. That is, as shown in FIG. 6, the pre-expected (a) edge portions, (b) corner portions, (c) uniform portions, etc. of the pattern are shown in FIG. (d) In each of the other parts, a large number of reference unit patterns are stored in advance in the storage section. Then, the read two-dimensional image data is divided into, for example, square unit image data having the same shape, and the pattern of each divided unit image data is determined.
The respective reference unit patterns are compared and contrasted in order to find out which reference unit pattern each square unit image data corresponds to.

[0004] Then, by replacing all the unit image data with the reference unit pattern, the entire pattern can be obtained from the read two-dimensional image data. Then, it is only necessary to check whether the obtained overall pattern matches the stored reference pattern.

[0005]

[Problems to be Solved by the Invention] However, the above-described pattern feature extraction method still has the following problems.

That is, as shown in FIG. 6, there are a huge number of types of reference unit patterns. For example, it is necessary to prepare separate reference unit patterns depending on the number of pixels in the vertical and horizontal directions, inversion of black and white, and deviation in relative position. Further, as shown in FIG. 7, it is necessary to prepare multiple types of reference unit patterns even if the pattern is the same. As described above, when the number of reference unit patterns to be compared increases, the number of times each measured unit image data must be compared becomes enormous, resulting in a significant decrease in processing efficiency.

Furthermore, the read two-dimensional image data is
When processing information as digitized data, it can be easily compared using each standard unit pattern shown in Fig. 6, but for example,
The image data of each pixel constituting the two-dimensional image data may be composed not of binary data but of multi-valued data indicating gradations of light and shade. In such cases, there are problems that cannot be easily compared and contrasted.

The present invention has been made in view of the above circumstances, and by determining the position to which each local area belongs to the pattern based on the differential value of pixel data of each pixel in the local area, even if each local area is It is an object of the present invention to provide a pattern feature extraction method that can easily extract the features of an entire pattern in a short time even if pixel data of a pixel is multivalued data representing shading.

[0009]

[Means for Solving the Problems] In order to solve the above problems, in the pattern feature extraction method of the present invention, pixel data of adjacent surrounding pixels for each pixel constituting a local area consisting of a unit number of pixels. The differential value in each direction is calculated using a differential parameter from the level difference of The position to which the corresponding local area belongs to the pattern is determined from the pixel data inclination direction of the pixel, and the characteristics of the pattern are extracted from the position to which each local area belongs.

0010

[Operation] According to the pattern feature extraction method configured as described above, the differential value in each direction for each pixel constituting a local area consisting of a unit number of pixels is determined. Then, the slope direction of the pixel data of one pixel is determined from the differential values in each direction. That is, this pixel data tilt direction is a value indicating in which direction the pixel data of the relevant pixel is tilted in relation to surrounding pixel data. Furthermore, it is determined to which position in the pattern the local area belongs based on the combination of the tilt directions of each pixel.

[0011] Therefore, for example, if this local area belongs to a vertical edge portion of a pattern, the direction in which the local areas of this vertical edge portion are continuous becomes the pattern boundary direction. Thus, by two-dimensionally combining all the local regions, two-dimensional features of the pattern can be extracted.

0012

DESCRIPTION OF THE PREFERRED EMBODIMENTS A pattern feature extraction method according to an embodiment of the present invention will be described below with reference to the drawings.

FIG. 1 is a diagram showing the format of two-dimensional image data to which the pattern feature extraction method of the embodiment is applied. This two-dimensional image data 1 is obtained by arranging a large number of pixels 2 in a matrix, and each pixel 2 is capable of displaying 11 levels of shading represented by integers from 0 to 10, for example. Therefore, the pattern 3 in this two-dimensional image data 1 is represented by each pixel 2 in 11 levels of gradation.

As shown in the figure, a local area 4 is composed of a total of nine pixels 2, three in the vertical direction and three in the horizontal direction. In FIG. 1, this local area 4 is located near the edge of pattern 3 (position A), near the corner (position B), near the diagonal edge (position C), and inside or outside of pattern 3. This shows the state where the image is located in the uniform part (position D). Next, a method for calculating the differential value of each pixel data will be explained using FIG. 2.

FIG. 2(a) shows nine pixels 2 belonging to a local area 4 near the edge shown at position A in FIG. FIG. 3 is a diagram showing pixel data shown in levels. For example, the five pixels 2 at the bottom are completely located within the pattern 3, so the pixel data is 0. Conversely, the top five pixels 2
is completely outside pattern 3, so the pixel data is 1
It is 0. The three horizontal rows in the center are located at the border of pattern 3,
The level of pixel data increases from bottom to top from 3 to 5 to 7.

Further, FIG. 2(b) shows a vertical differential operator 5a. FIG. 2(f) is a diagram showing a differential value a obtained by spatially differentiating the pixel data of each pixel 2 in the vertical direction using the differential operator 5a. For example, the differential value a indicating [4] in the center of FIG. 2(f) is
] The difference [2] between the pixel data of [7] and the pixel data of upper [7]
] is the value obtained by adding the difference [2] between the pixel data of the center [5] and the pixel data of the bottom [3].

In this way, in the vertical differentiation, the differential value a of each pixel 2 is calculated by adding the difference between the pixel data of the pixel data to be differentiated and the pixel data of each pixel 2 adjacent above and below.
are calculated in order. In this way, the calculation process of each differential value a of the nine pixels 2 constituting the local area 4 is performed.

Similarly, FIG. 2(d) shows a lateral differential operator 5c. FIG. 2(h) is a diagram showing the differential value c obtained by spatially differentiating the pixel data of each pixel 2 in the horizontal direction using the differential operator 5b. That is, in this horizontal and vertical direction differentiation, the differential value c of each pixel 2 is calculated in order by adding the difference between the pixel data of the pixel data to be differentiated and the pixel data of each pixel 2 adjacent to the left and right. In this example, as shown in FIG. 2(a), the pixel data does not change at all in the horizontal direction, so each differential value c is all [0]
It is.

Further, FIG. 2(c) shows the differential operator 5b in the +45° direction. FIG. 2(g) uses this differential operator 5b to convert the pixel data of each pixel 2 by +45°.
It is a figure which shows the differential value b spatially differentiated in the oblique direction. That is, in this +45° diagonal differentiation, the differential b of each pixel 2 is sequentially calculated by adding the difference between the pixel data to be differentiated and the pixel data of each pixel 2 adjacent in the +45° diagonal direction. I will calculate it.

Furthermore, FIG. 2(e) shows a differential operator 5d in the -45° direction. FIG. 2(i) uses this differential operator 5d to calculate the pixel data of each pixel 2 by −45
It is a figure which shows the differential value d spatially differentiated in the oblique direction of degree. That is, in this -45° diagonal differentiation, the differential d of each pixel 2 is obtained by adding the difference between the pixel data to be differentiated and the pixel data of each pixel 2 adjacent in the -45° diagonal direction. Calculate in order.

In this way, each differential parameter 5a in a total of four directions: vertical direction, +45° direction, horizontal direction, -45° direction
~5d, four differential values a, b, c, and d were calculated for each of the nine pixels 2 constituting each local area 4.

When the calculation process of the differential values a, b, c, d in each direction of each pixel 2 in each local area 4 is completed, the slope direction of the pixel data is determined for each pixel 2 according to the algorithm shown in FIG. do.

When the flowchart is started, first, the differential values a, b, c, and d of the corresponding pixel 2 in each direction are calculated as described above. Next, in P (program step) 1, it is checked whether the maximum value of the absolute values of the four differential values a to d is one. If the maximum value is one, the direction corresponding to the corresponding differential value is set as the tilt direction of the pixel data of this pixel 2 at P2. Then, the process proceeds to P3, and if the corresponding differential value is a (-) value, the inclination direction determined in P2 is reversed by 180 degrees.

In P1, if there are two or more maximum values of the absolute values of the four differential values a to d, conversely, check whether the minimum value of each absolute value is one. . If the minimum value is one, the 90° direction corresponding to the corresponding differential value is set as the tilt direction of the pixel data of this pixel 2 in P5. and,
Proceeding to P3, if the corresponding differential value is a (-) value, the inclination direction determined in P5 is reversed by 180°.

Further, in P4, if there are two or more minimum values of the absolute values of the four differential values a to d, it is checked in P6 whether all the absolute values are 0 or not. If it is 0,
This pixel 2 is assumed to exist in an internal or external uniform area far away from the boundary of the pattern 3, and is determined to be in an indeterminate tilt direction (uniform area) at P7. Further, if an absolute value other than 0 exists in P6, it is determined that the tilt direction is other than ±45°.

In this way, the four differential values a, b, c, d
By examining the magnitude relationship of the pixel data of each pixel 2, in comparison with each pixel adjacent to this pixel 2,
The direction in which it is inclined is uniquely determined.

For example, the center pixel data [
5], the differential values a, b, b, and d of pixel 2 are 4, 4, 0, and -4 as shown in the figure. As it is, this pixel 2
The direction of inclination of the pixel data is not determined, but according to the algorithm shown in FIG.
That is, the vertical direction is the tilt direction, which coincides with position A of the two-dimensional image data 1 in FIG.

FIG. 4 shows the pixel data of each pixel 2 belonging to each local area 4 at each position A to D in FIG. 9 is a diagram illustrating the relationship between nine pixels 2 in the inclination direction D indicated by an arrow. FIG.

As for the pixel data near the edge (position A) shown in FIG. 4(a), as shown in FIG. It is the downward (vertical) direction that exists. Also, Figure 4(c)
For pixel data near the corner (position B) shown in
As shown in FIG. 4D, the inclination direction of the pixel data of each pixel 2 is directed toward the lower left where the corner of the pattern 3 exists as a whole.

For the pixel data of the uniform portion (position D) shown in FIG. 4(d), as shown in FIG. 4(e), each pixel 2
The inclination directions of the pixel data are all undefined and directions are undefined. Furthermore, near the oblique edge shown in FIG. 4(f) (position C
), as shown in Figure 4(g),
The inclination direction of the pixel data of each pixel 2 is all directed toward the lower right direction where the edge of the pattern 3 exists.

When the process of determining each inclination direction D of the pixel data of each pixel 2 in each local area 4 is completed, the position to which each local area 4 belongs to the pattern 3 is determined according to the algorithm shown in FIG.

When the flow chart starts, in Q1, the data gradient direction D of all N pixels 2 constituting the local area 4 is determined as described above. In Q2, all N pieces are shown in Fig. 4(e).
If the direction is undefined as shown in , it is determined in Q3 that the corresponding local area 4 is located in the uniform part.

Further, if the inclination direction D of the pixel data of all pixels 2 is in the same direction in Q4, the process proceeds to Q5. In Q5, if the direction is the vertical direction shown in FIG. 4(b), it is determined to be the vertical edge position. Furthermore, if it is in the horizontal direction in Q6, it is determined to be a horizontal edge position. Further, if the direction is neither vertical nor horizontal, it is determined that the direction is diagonal as shown in FIG. 4(g), and the position is set as the diagonal edge position.

Further, if it is determined in Q7 that there are two types of inclination directions D of the N pixels 2, the process proceeds to Q8. and,
If one of the inclination directions D is indeterminate in Q8, it is determined that the inclination direction D is located at the end of the edge portion. In Q9, if the two types of inclination directions D come into contact with and separate from each other in four directions, 45
It is determined that the direction is diagonal other than °.

[0035] Also, in Q7, three or more types of inclination directions D
exists, and if the two types of inclination directions D do not separate from each other in Q9, it is determined that this local area 4 belongs to another part such as a corner part.

In this way, the position of this local area 4 with respect to the pattern 3 can be determined from the slope direction D of the pixel data of all the pixels 2 constituting the local area 4.
It can be determined each time.

[0037] Once the positions to which the pattern 3 belongs in all the local areas 4 of the number of pixels formed around all the pixels 2 constituting the secondary image data 1 shown in FIG. By connecting the local subregions 4 of the affiliated positions, the characteristics of the pattern 3 can be extracted.

According to the pattern feature extraction method configured as described above, the inclination direction D of pixel data for each pixel 2 constituting the two-dimensional image data 1 is determined according to a predetermined algorithm shown in FIG. Furthermore, the local sub-area 4 centered on each pixel 2 is formed using the respective tilt directions D of the unit number of pixels 2 constituting the local sub-area 4 according to a predetermined algorithm shown in FIG. It is automatically determined to which position the pattern 3 belongs. Then, the characteristics of the pattern 3 are extracted based on the position to which each local sub-region 4 belongs.

In this way, since the features of pattern 3 included in two-dimensional image data 1 are automatically extracted according to a certain algorithm, a large number of reference unit patterns are extracted as shown in FIGS. 6 and 7. Compared to the conventional method of preparing unit image data in the storage unit and checking by trial and error which reference unit pattern the retrieved unit image data matches, processing efficiency can be significantly improved. Furthermore, the required storage capacity can be significantly reduced compared to the case where a large number of reference unit patterns are prepared in the storage section.

Furthermore, the pixel data of each pixel 2 is differentiated using differential parameters 5a to 5d, and the tilt direction D of the corresponding pixel 2 is calculated from each differential value a to d. If the slope direction D is calculated using this method, even if the pixel data of each pixel 2 is not binary data of only black and white, but multivalued gradation data indicating light and shade, the same processing procedure can be used. It can be executed with , and the features of pattern 3 can be easily extracted.

Note that the present invention is not limited to the embodiments described above. In the method of the embodiment, local area 4 is
Although it is formed with a total of 9 pixels 2 of ×3, it may be formed with a total of 16 pixels 2 of 4 × 4, or furthermore, it may be formed with a total of 25 pixels 2 of 5 × 5. Of course.

Furthermore, in the method of the embodiment, FIG.
Although the differential parameters 5a to 5d shown in ) to (e) are used, other differential operators such as Sobel may also be used. Further, in the implementation method, the directions of spatial differentiation are limited to four types, but spatial differentiation may be performed in eight directions, for example. In this way, by increasing the conditions in the spatial differential direction, the features of pattern 3 can be extracted in more detail.

[0043]

Effects of the Invention As explained above, according to the pattern feature extraction method of the present invention, a local sub-region is formed by a unit number of pixels centered on each pixel constituting two-dimensional image data, and this local sub-region is The position to which each local region belongs to the pattern is determined based on the differential value of the pixel data of each pixel. Therefore, even if the pixel data of each pixel is multivalued data representing shading, the characteristics of the entire pattern can be extracted easily and in a short time.

[Brief explanation of drawings]

FIG. 1 is a diagram showing two-dimensional image data to which a pattern feature extraction method according to an embodiment of the present invention is applied;

[Figure 2
] A diagram showing the relationship between the differential parameters and each differential value used in the method of the example,

[Fig. 3] Flowchart for determining the slope direction of pixel data of each pixel used in the method of the embodiment,

FIG. 4 is a diagram showing the inclination direction of each pixel forming the local area determined according to the flowchart in FIG. 3;

[Figure 5] Flowchart for determining the belonging position for each local area pattern used in the method of the embodiment,

[Figure 6]
A diagram showing each standard unit pattern adopted in the conventional method,

FIG. 7 is a diagram showing details of each reference unit pattern adopted in the conventional method.

[Explanation of symbols]

1... Two-dimensional image data, 2... Pixel, 3... Pattern, 4...
Local area, 5a to 5e... Differential parameter, a to d... Differential value, D... Inclination direction.

Claims (1)

[Claims] Claim 1. In a pattern feature extraction method for extracting pattern features in two-dimensional image data consisting of a plurality of pixels, for each pixel constituting a local area consisting of a unit number of pixels, pixels of adjacent surrounding pixels are extracted. Calculate the differential value in each direction from the data level difference using the differential parameter,
The slope direction of pixel data is determined for each pixel based on the calculated differential value in each direction, and the position of the corresponding local area with respect to the pattern is determined from the pixel data slope direction of all pixels constituting each local area. A pattern feature extraction method characterized in that the features of the pattern are extracted from the belonging position of each local area.