JP4404276B2 - Image information processing apparatus, image information processing method, and recording medium - Google Patents

Description

[0001]
BACKGROUND OF THE INVENTION
The present invention relates to an image information processing apparatus for recognizing and detecting defect information such as foreign matter generated on a semiconductor wafer, a printed circuit board, or the like.
[0002]
[Prior art]
Defects and foreign matter that are generated or mixed in the manufacturing process of printed circuit boards or semiconductor wafers cause defective products. Therefore, it is necessary to detect the defect / foreign matter as soon as possible.
[0003]
2. Description of the Related Art Conventionally, there has been an image information processing apparatus that detects, recognizes, or determines a defect or a foreign substance as described above in a photograph of a printed board or an image input by photographing a semiconductor wafer with an optical microscope or the like. Most of these methods are digital image processing that handles images digitally, and image processing is performed using a computer.
[0004]
Conventionally, defect detection using image processing is mainly known as (a) comparison with a golden device, (b) comparison with CAD data, (c) die-to-die comparison, (d) FET method, and the like.
[0005]
(A) In comparison with a golden device, an image of an ideal semiconductor wafer (hereinafter referred to as a golden device) having no defect is prepared in advance, and the image of this golden device is compared with the image to be inspected to detect defects. To detect. Comparing images means processing for creating a difference image of pixel values for each corresponding pixel of two images. When there is no defect in the image to be inspected, the image is the same as the image data of the golden device, so that the entire screen becomes a completely flat image having a pixel value of 0 by the difference processing. However, if there is a defect or foreign matter, pixels with values other than 0 appear intensively in that portion. Therefore, by extracting a group of pixels having values other than 0 and obtaining the size and the center of gravity, it is possible to detect a defect or a foreign object.
[0006]
(B) In the comparison with the CAD data, the CAD image data designed for the device to be inspected is compared with the image data of the manufactured wafer.
[0007]
As a result, defects and foreign matter can be detected as in (a) comparison with the golden device.
[0008]
(C) In die-to-die comparison, image data of adjacent dies are compared using the fact that a plurality of the same chips (hereinafter referred to as dies) are arranged on one wafer. As a result, defects and foreign matter can be detected as in (a) and (b).
[0009]
(D) In the FET method, the defect generated on the wafer is localized, whereas the defect is detected by removing the wiring pattern by utilizing the property that the wiring pattern is periodic. That is, the input wafer pattern image is Fourier-transformed using a two-dimensional FET (fast Fourier transform), etc., and a specific spatial frequency component corresponding to the wiring pattern is removed using a band stop filter or the like in the spatial frequency domain. By doing so, defects can be detected.
[0010]
[Problems to be solved by the invention]
(A) When using comparison with a golden device, it is necessary to prepare an image of the golden device in advance. To create an image of a golden device, you must first find a sample that is free of defects. The task of finding this defect-free sample must be done carefully with the human eye. In addition, the patterns of semiconductor wafers are very diverse due to the current trend of high-mix low-volume production, and there are many changes. It is necessary to create a golden device image for each of these varieties or design changes, which requires considerable effort.
[0011]
(B) Comparison with CAD data is efficient because it is not necessary to search for a golden device as in (a). Recently, as semiconductor chips have been highly integrated, a conventional optical microscope cannot be used to obtain a detailed image of a wafer pattern, and a scanning electron microscope (SEM) has been used. ing. The SEM image has more noise than the optical microscope image. Therefore, when a difference image between CAD data without noise and an SEM image is created, a large amount of noise other than defects is detected.
[0012]
Further, in both methods (a) and (b), when performing image comparison, it is necessary to perform accurate alignment first.
[0013]
(C) In die-to-die comparison, adjacent dies are compared with each other, so that it is basically unnecessary to perform alignment. However, there is a drawback that it cannot be detected if there is a similar defect in the same place by chance. Similarly to (b), when an SEM image is applied, noise is detected.
[0014]
(D) The FET method can detect defects even from an SEM image. However, this method requires a very long processing time. For example, when a two-dimensional Fourier transform is applied to a digital image composed of 512 × 512 pixels, approximately 1.31 × (10 8) additions / subtractions and 8.39 × (10 8) multiplication are required. It becomes. Further, complicated wiring patterns such as MPU cannot be accurately removed by filtering in the spatial frequency domain.
[0015]
The first object of the present invention is to detect a defect even from an unclear SEM image in which the image comparison methods (a), (b), and (c) are greatly affected by noise and cannot detect the defect accurately. It is to provide an image information processing apparatus capable of performing the above.
[0016]
A second object of the present invention is to provide image information capable of detecting a defect from a single image to be inspected without using a golden device or CAD data required in the image comparison methods (a) and (b). It is to provide a processing device.
[0017]
A third object of the present invention is to provide an image information processing apparatus capable of detecting a defect from an inspection object image at high speed without using a large amount of processing time as in the (d) FET method.
[0018]
[Means for Solving the Problems]
  The present inventionIs a two-dimensional wavelet transform means for performing two-dimensional wavelet transform on input digital image data having circular defect image data, and a vertical line detection component and a horizontal line detection component obtained by the two-dimensional wavelet transform. Threshold processing is applied to create a binary image of vertical line detection components and horizontal line detection components, and a square is detected from the binary image obtained by the binarization processing means. Square Hough transforming means for applying the square Hough transform to obtain the position and size of the object.
[0019]
The two-dimensional wavelet transform means separates a wiring pattern having only vertical or horizontal line components and defects having components in various directions. Then, binarization processing means binarizes the detected components of the vertical and horizontal lines to obtain a binary image, and the square Hough transform means applies a square Hough transform to the binary image. The defect is detected as a square. A circular defect is also detected as a square.
[0020]
Therefore, defects can be detected from a single inspection target image without using a golden device or CAD data, and a large amount of processing time is not required as in the FET method (two-dimensional wavelet transform takes less time than the FET method). .
[0021]
  The present inventionImage information processing deviceIsAnd binarization processing means for detecting specific graphic information by applying threshold processing to the parameter space obtained by the square Hough conversion means.You may do.
[0022]
  The present inventionImage information processing deviceIsA labeling means for grouping the active pixels adjacent to each other in the binary image of the parameter space obtained by binarizing the square Hough transformation result, and the center of gravity of each label obtained by the labeling process Centroid calculating means for obtaining coordinatesYou may do.
[0023]
  The present inventionImage information processing deviceIs, It is equipped with a noise removal means for performing local noise removal on the input digital imageYou may do.
[0024]
  The present inventionImage information processing deviceIsIsolated point removal means for removing isolated active pixels from the binary image of the vertical line detection component and horizontal line detection component after the binarization processing means is applied to the two-dimensional wavelet transform result. HaveYou may do.
[0025]
  The present inventionImage information processing deviceIsThe square Hough conversion means performs Hough conversion for detecting the xy coordinates of the center of the square and the length of one side.You may do.
[0026]
  The present inventionImage information processing deviceIsThe square Hough transform means is a means for Hough transforming the two binary images corresponding to the vertical line detection component and the horizontal line detection component of the two-dimensional wavelet transform into the same parameter space.You may do.
[0027]
  The present inventionImage information processing deviceIsThe square Hough transform means is a means for performing a Hough transform on only one of the two binary images corresponding to the vertical line detection component and the horizontal line detection component of the two-dimensional wavelet transform.You may do.
[0028]
  The present inventionImage information processing deviceIs, Energy calculating means for calculating the energy of the vertical line detection component and the horizontal line detection component of the two-dimensional wavelet transform, and a square Hough transform means, the binary image corresponding to the component having the lower energy Perform square Hough transform onYou may do.
[0029]
  The present inventionImage information processing deviceIsThe threshold value of the binarization processing means for the parameter space by the square Hough transform changes in accordance with the length parameter of one side of the square.You may do.
[0030]
  The present inventionImage information processing deviceIs, Equipped with a duplicate detection and removal means for removing duplicates when a plurality of objects overlap in positionYou may do.
[0031]
  The present inventionIs a two-dimensional wavelet transform process for performing two-dimensional wavelet transform on input digital image data having circular defect image data, and a vertical line detection component and a horizontal line detection component obtained by the two-dimensional wavelet transform process. A binarization processing step for creating a binary image of the vertical line detection component and the horizontal line detection component by applying threshold processing to the binary image obtained by the binarization processing step A square Hough transform step of obtaining a position and a size of the object by applying a square Hough transform for detecting a square.
[0032]
  The present inventionIs a two-dimensional wavelet transform process for performing two-dimensional wavelet transform on input digital image data having circular defect image data, and a vertical line detection component and a horizontal line detection component obtained by the two-dimensional wavelet transform process. A binarization process for creating a binary image of the vertical line detection component and the horizontal line detection component by applying a threshold processing to the binary image obtained by the binarization process. A computer-readable recording medium storing a program for causing a computer to execute a square Hough transform process for obtaining the position and size of the object by applying a square Hough transform to detect the object It is.
[0033]
DETAILED DESCRIPTION OF THE INVENTION
The present invention can be broadly divided into a two-dimensional wavelet transform means for separating defect components from a wiring pattern of a wafer pattern image, and a Hough transform means for detecting the position and size of a defect from a binary image from which the defect components are separated. In order to improve the defect detection performance by the above means, other means are supplementarily used.
[0034]
(A) Square Hough conversion means
For the Hough transform, see “Basics of Image Recognition [ΙΙ]” (Mori, Sakakura, Ohm) pp. 3-19, the Hough conversion process will be briefly described. The Hough transform is image processing for detecting a specific figure from a binary image. Here, a method for detecting a square will be described.
[0035]
In the xy plane, as shown in FIG. 4, a square inclined 45 degrees is expressed by the following equation, where (a, b) is the center coordinate of the square and r is the length of one side of the square.
[0036]
| X−a | + | y−b | = r / √2 (1)
That is, one square on the xy plane is represented by one point in the (a, b, r) parameter space. The Hough transformation converts the xy plane into the parameter space. In this case, the transformation into the (a, b, r) space indicating the square features of the xy plane. ) Called conversion. The square Hough transform is realized by the following equation obtained by modifying equation (1).
[0037]
b = ± (r / √2− | x−a |) + y (2)
A specific example of a square Hough transform will be described with reference to FIG. Each observation point α (x1, y1), β (x2, y2), and γ (x3, y3) in FIG. 4A is mapped to the (a, b, r) space by equation (2). The r axis is orthogonal to the ab plane. FIG. 4B shows an ab plane sliced at a certain r0 (here r0 = 24). For each plane at r, a single Hough curve (square) is obtained from one observation point. If the observation point is on a square whose half distance (hereinafter referred to as radius) is r0, the square (Hough) curves intersect at a point (a0, b0, r0). That is, when a square exists on the xy plane, it can be detected as an intersection of many Hough curves in the parameter space. At the intersection, the luminance at the intersection increases in proportion to the number of intersecting Hough curves. Therefore, a square on the xy plane can be detected by finding a parameter having a large luminance value in the parameter space. In FIG. 4B, only the parameter (a0, b0, 24) has a luminance value of 3, and the others are luminance values of 2 or less. By detecting a large luminance value (a0, b0, 24), a square having a radius (24) and a center (a0, b0) at three points on the original xy plane is detected.
[0038]
Note that depending on the square Hough transform, a circle in the binary image can be detected as a square circumscribing the circle. Digital image processing targets image data, that is, discrete data. As shown in FIG. 16, a square having the same central coordinates intersects or touches a circle having a radius r (hereinafter referred to as “intersect”) is a section where the square radius r is r0 / √2 or more and r0 or less. is there. In this section, when the square circumscribes the circle, that is, when r = r0, it is sufficient to prove that the section where the circle and one side of the square intersect is the longest. This is proved using mathematical induction.
[0039]
(1) When r = r0
Since the radius of the circle varies discretely, the radius next to the radius r0 is r0 + 1. Since the r = r0 + 1 square does not touch the circle with the radius r0, the r = r0 square has a longer section where it intersects the circle than the r = r0 + 1 square. That is, G (r0)> G (r0 + 1), where G (r) is a section where a square with a radius r intersects with a circle with one side r0.
[0040]
(2) When r = ri
Assuming that G (ri) <G (ri + 1), consider the case of r = ri−1. The relationship between a circle and a square when r = ri is as shown in FIG. Since the circle and the square intersect at eight places shown in FIG. 15, G (ri) = 8 (α (ri) −β (ri)).
[0041]
α (ri)²= R0²-Ri²
β (ri)²= R0²-(Ri + 1)²
Therefore, α (ri)²-Β (ri)²= 2ri + 1
Further, when a similar study is performed on α (ri−1) and β (ri−1) corresponding to r = ri−1,
(Α (ri-1))²-(Β (ri-1))²= 2ri-1
It becomes. Therefore, taking the difference of squares and organizing
α (ri)²-Β (ri)²-{(Α (ri-1))²-(Β (ri-1))²} = 2> 0
Therefore,
α (ri)²-Β (ri)²> (Α (ri-1))²-(Β (ri-1))²
It becomes.
[0042]
From α (ri)> 0, β (ri)> 0, α (ri) -β (ri)> 0, α (ri-1) -β (ri-1)> 0,
α (ri) -β (ri)> (α (ri-1))-(β (ri-1))
It can be.
[0043]
G (ri) = 8 (α (ri) −β (ri))
Since G (ri−1) = 8 (α (ri−1) −β (ri−1)),
G (ri)> G (ri-1).
[0044]
(3) From (1) and (2),
... <G (r0-2) <G (r0-1) <G (r0)> G (r0 + 1) = G (r0 + 2) = ...
When G (r0) is the maximum value, that is, when r = r0, the section where the circle and the square intersect is the longest.
[0045]
That is, when a square Hough transform is performed on this circle, the most Hough curves intersect in the ab parameter space of r = r0 to obtain a peak value. Therefore, when a square Hough transform is applied to a circle with a radius r0, a radius r0, that is, a square circumscribing the circle is detected. For example, FIG. 16B shows the result of applying the square Hough transform to the binary image of the particle defect on the semiconductor wafer as shown in FIG. The peak value is taken at r = 78 and detected as a square. Even when the circle Hough transform is applied to the same binary image, the circle of r = 78 is detected, and it can be seen that the square Hough transform correctly detects the circle.
[0046]
(B) Wavelet transform means
The wavelet transform is described in detail in CHUI, “An Introduction to WAVELETS” Academic Press, 1922. Here, the wavelet transform process will be briefly described.
[0047]
The wavelet transform for image data is a two-dimensional wavelet transform, which is realized by combining the one-dimensional wavelet transform for the x-axis direction and the one-dimensional wavelet transform for the y-axis direction. explain. There are many basis functions used when performing wavelet transform, but here we will explain using the Harr Wavelet, which has the simplest structure. For other wavelet basis functions, the output information is almost the same except that the shape of the function is different. The wavelet transform is composed of two orthogonal functions, a scaling function and a wavelet function. The scaling function is a function that outputs data smoothing information (= low-pass information), and the wavelet function is a function that outputs detailed data information (= high-pass information). In the case of Harr Wavelet, the scaling function is g0 = g1 = 1/2, the wavelet function is h0 = 1/2, and h1 = −1 / 2.
[0048]
When there are input signals s0 to s15, Harr Wavelet converted output signals t0 to t15 are as follows.
[0049]
t0 = g0 · s0 + g1 · s1 t8 = h0 · s0 + h1 · s1
t1 = g0 · s2 + g1 · s3, t9 = h0 · s2 + h1 · s3
t2 = g0 · s4 + g1 · s5, t10 = h0 · s4 + h1 · s5
...
t7 = g0 · s14 + g1 · s15, t15 = h0 · s14 + h1 · s15
Try inputting a specific signal to this Harr Wavelet transform. For example, the following input signal s is given.
[0050]
s (n) = {0,0,0,0,0,0,0,2,2,2,2,2,2,2,2,2}
(3)
This signal changes greatly from 0 to 2 in one place. A place where such a signal greatly changes is called an edge. As in (3), an edge with a large signal value is called an edge rising edge, and an edge with a small signal value is called an edge falling edge. When the signal of (3) is Harr Wavelet transformed, the following result t is obtained.
[0051]
N → 0, 1, 2, 3, 4, 5, 6, 7; 8, 9, 10, 11, 12, 13, 14, 15
t (N) = {0, 0, 0, 1, 2, 2, 2, 2; 0, 0, 0, -1, 0, 0, 0, 0} (4)
(Low-pass component) (High-pass component)
The result of wavelet transform is called a wavelet coefficient. An edge of the input signal s is detected in the wavelet coefficient t (11) of the high pass component. In this way, the wavelet transform can detect the edge component of the input signal.
[0052]
This wavelet transform can also be applied to two-dimensional image data such as an SEM image of a wafer pattern. Specific examples in which the wavelet transform is applied to image data are shown in FIGS. In this example, the original image in FIG. 5A is 512 × 512 digital data. The image is first subjected to a one-dimensional wavelet transform in the x-axis direction. That is, the wavelet transform for 512 signals in the x-axis (horizontal axis) direction is repeated 512 times in the y-axis (vertical axis) direction. Thereby, the image of FIG. 5B is obtained. FIG. 5B shows the original image of FIG. 5A divided vertically into two pieces, in which low-pass information (L) is stored on the left side and high-pass information (H) is stored on the right side.
[0053]
This time, the same wavelet transform is applied to this FIG. 5B in the y-axis direction. Thereby, the image of FIG. 6 is obtained. FIG. 6 is an image in which FIG. 5B is vertically divided into two, and low-pass information (L) is stored on the upper side and high-pass information (H) is stored on the lower side. Accordingly, in FIG. 6, the original image of FIG. 5A is divided into four parts, low-pass information (LL component) in both the x and y axes in the upper left, and information combining the high pass information in the x axis and the low pass information in the y axis in the upper right. (HL component), information that combines low-pass information in the x-axis direction and high-pass information in the y-axis direction (LH component) is stored in the lower left, and high-pass information (HH component) in both the x and y axes is stored in the lower right. It will be. That is, the upper right part is a Y-axis direction line component (vertical line component) existing in the original image, the lower left part is an X axis direction line component (horizontal line component) existing in the original image, and the lower right part. Indicates a line component in an oblique direction. In addition, here, the conversion process is first performed in the x-axis direction and then the conversion process is performed in the y-axis direction to obtain FIG. 6, but the final two-dimensional wavelet can be obtained even if the x and y conversion order is changed. The completely same image of FIG. 6 is obtained after conversion.
[0054]
When the two-dimensional wavelet transform having the above properties is applied to a semiconductor wafer image having a granular defect, the defect and the background wiring pattern can be separated. FIG. 7 is a simulation of a semiconductor wafer image. Horizontal lines running horizontally correspond to wiring patterns, and black circles in the center correspond to defects. Thus, the semiconductor wafer has the wiring patterns aligned in the horizontal direction or the vertical direction. When the two-dimensional wavelet transform is applied to such an image, the result of FIG. 8 is obtained. When attention is paid to the vertical line detection component and horizontal line detection component, the wiring pattern appears only in the LH component for detecting the lower left horizontal line. On the other hand, defect components having various direction components are detected as an upper right HL component and a lower left LH component. As described above, when the two-dimensional wavelet transform is used, it is possible to separate a wiring pattern having only a vertical line or horizontal line component and a defect having components in various directions.
[0055]
By combining the two-dimensional wavelet transform described above and the Hough transform, it is possible to detect the position and size of a defect present in the semiconductor wafer image.
[0056]
Hereinafter, embodiments of the present invention will be described with reference to the drawings. First, the configuration of the image information processing apparatus according to the embodiment of the present invention will be described with reference to FIG. The computer 101 includes a CPU 31 that performs input / output and calculation, a cache 32 that stores frequently used instructions and data and speeds up processing, a floating-point arithmetic unit 33, a RAM 34 that stores user inputs and data, and a ROM 36 that stores system programs. A display device 35 for displaying a user's selection menu and calculation results, a keyboard for inputting parameters and commands, an input device 37 for a pointing device (such as a mouse), and a magnetic disk 38 for storing the calculation results. These devices are connected by a CPU bus 30.
[0057]
A DUT (device under test) 41 such as a semiconductor wafer to be inspected is loaded / unloaded onto a stage 45 by a loader / unloader 46. The loaded DUT is input as an image by an image input device 42 such as an electron microscope, converted into digital data by an A / D converter 43, stored in a frame memory 44, or transferred to a RAM 34 on a computer. This digital image is, for example, a grayscale image of 256 gradations with a resolution of 512 × 512 pixels. The frame memory 44 is connected to the CPU bus 30 via the I / O bus 40. The operation of the control device 47 is controlled by the stage 45 and the loader / unloader 46. The control device 47 is connected to the computer 101 via the I / O bus 40.
[0058]
The wavelet transform unit 51, the Hough transform unit 53, the binarization processing unit 54, the labeling processing unit 55, the noise removal unit 56, the isolated point removal unit 57, the centroid calculation unit 58, and the duplicate detection removal unit 59 are included in the frame memory 61. , 63, 64, 65, 66, 67, 68, 69 are connected to the CPU bus 30. The energy calculating means 52 is connected to the frame memory 61 and the CPU bus 30. It should be noted that the wavelet transform unit 51, the energy calculation unit 52, the square Hough transform unit 53, the binarization processing unit 54, the labeling processing unit 55, the noise removal unit 56, the isolated point removal unit 57, the centroid calculation unit 58, and the duplicate detection removal unit The operation 59 is described in the description of the operation of the embodiment.
[0059]
Next, the operation of the first embodiment will be described with reference to FIG. 1A. The image data stored in the frame memory 44 or the RAM 34 is used as input image data, which is first transferred to the frame memory 61, and the two-dimensional wavelet transform is performed by applying the wavelet transform means 51 (S1). A vertical line detection component and a horizontal line detection component are obtained by two-dimensional wavelet transform. These components have wavelet coefficient values having a large absolute value in the edge portion, and the wavelet coefficients in other portions have values of 0 or close to 0.
[0060]
Next, the absolute value of the wavelet coefficient is transferred to the frame memory 64, and binarization processing is performed by the binarization processing means 54 (S2). That is, the binarization processing means 54 performs binarization processing by giving appropriate threshold values to the vertical line detection component and the horizontal line detection component, and sets the coefficient value of the edge portion as 1 and the other values as 0. Create a value image. Here, the threshold value is set to ± 5 empirically.
[0061]
Finally, the binary image is transferred to the frame memory 63, and the square Hough transformation is performed by the square Hough transformation means 53 (S3). The parameter space is three-dimensional (a, b, r). Note that the position and size of a defect can be automatically specified by adding a process after the first embodiment.
[0062]
According to the first embodiment, the following effects can be obtained. Although the absolute value of the wavelet coefficient at the edge portion is large, the sign may be positive or negative. Since the negative coefficient is 0 in the normal binarization process, for example, when normal threshold processing is performed on the vertical line detection component in the upper right of FIG. 8, only half of the edge is detected as shown in FIG. However, the Hough transform can be detected even if the target graphic is missing. In contrast to FIG. 9, when the Hough transform is used, in the parameter space, the brightness of the parameter corresponding to the defect is increased in the case of a square component existing in the original image, that is, a wafer pattern image. By observing the parameter space, the position and size of the defect can be specified.
[0063]
The operation of the second embodiment will be described with reference to FIG. 1B. The parameter space data after the Hough transform is sent to the frame memory 64, and the binarization processing means 54 performs binarization processing (S4). Note that the threshold for binarization is changed by r. When a square with a center (a0, b0) and a radius r0 exists in the xy plane, the brightness of the parameter (a0, b0, r0) is 8r0. For example, the threshold value is set to 4r0. When set, the brightness of the parameter exceeds the threshold even if half of the square with radius r0 is missing. When threshold processing is performed with this threshold value, one group of luminance 1, that is, white pixels appears.
[0064]
The operation of the third embodiment will be described with reference to FIG. 1C. The result of the binarization process (S4) is transferred to the frame memory 65. Then, the labeling processing unit 65 performs a labeling process for grouping pixels having a luminance value of 1 adjacent to the result of the binarization process (S4) (S5). In the parameter space after the Hough transform, the luminance value becomes a maximum value in the parameter in which the detected Taisho figure exists, but the luminance value of the parameter adjacent to it also increases. Therefore, in the result of the binarization process, the parameter of the luminance value 1 exists in the vicinity of the parameter corresponding to the figure to be detected (for example, a square). By the labeling process (S5), it is regarded as a feature parameter for one figure in a portion where the parameter of luminance value 1 is concentrated.
[0065]
Then, the a-coordinate and b-coordinate of the parameter belonging to each label obtained are sent to the frame memory 68, and the centroid calculating means 58 calculates the average value (a ′, b ′) of the a-coordinate and b-coordinate of the parameter as the centroid. (S6). Thereby, the position and size of the figure to be detected can be automatically specified from the input image.
[0066]
The operation of the fourth embodiment will be described with reference to FIG. 2A. Prior to the two-dimensional wavelet transform (S1), the input digital image is sent to the frame memory 66, and noise is removed by a noise removing means 56 such as a median filter.
[0067]
If spike noise is present in the input image, it is detected as an edge by wavelet transform. If a large number of noises other than the detection target graphic are subjected to the Hough transform, those other than the detection target are erroneously detected. By applying a noise removal filter to the original image, erroneous detection can be prevented. The Hough transform increases the processing time in proportion to the number of pixels having a pixel value of 1 in the binary image. Since the number of wavelet coefficients of the pixel value 1 generated by the wavelet transform unit + binarization processing unit is reduced by the noise removing unit 56, the processing speed is also improved.
[0068]
The operation of the fifth embodiment will be described with reference to FIG. 2B. In order to prevent erroneous detection and improve processing speed, the image after wavelet transform means + binarization processing means is sent to the frame memory 67, and the isolated points are removed by the isolated points removing means 57 (S2A). This is an image processing means in which a binary image having a pixel value of 1 and all adjacent pixels having a pixel value of 0 is regarded as an isolated point, and the pixel value of the pixel is replaced with 0.
[0069]
In the binary image after wavelet transform, there are adjacent wavelet coefficients having a coefficient value of 1 corresponding to the figure to be detected. Therefore, even if the coefficient value is 1, if all the surroundings are coefficient value 0, this coefficient is regarded as noise and is set to coefficient value 0. Thereby, prevention of erroneous detection and improvement of processing speed are realized.
[0070]
The operation of the sixth embodiment will be described with reference to FIG. 2C. The binarization process (S2) in the fifth embodiment is replaced with an absolute value binarization process (S2). In consideration of the fact that the coefficient value of the edge part may become negative due to the wavelet transform, the binarization process is performed on the absolute value of the wavelet coefficient instead of the normal binarization process. .
[0071]
When this absolute value binarization processing is applied to the vertical line detection component in the upper right of FIG. 8, the entire edge can be detected as shown in FIG. By applying the Hough transform to this, the brightness of the parameter corresponding to the defect is increased and the detection is facilitated as compared to using the normal binarization process.
[0072]
The operation of the seventh embodiment will be described with reference to FIG. 3A. This is different from the fifth embodiment in the square Hough transform (S3). That is, the square Hough transform means 53 performs Hough transform on both the vertical line detection component and the binary image corresponding to the horizontal line detection component of the wavelet transform result sent to the frame memory 63, but the two Hough transforms are converted into the same parameter space. Overlay and convert. Since the figure to be detected appears in both the vertical line component and the horizontal line component, it can be reliably detected by this method.
[0073]
The operation of the eighth embodiment will also be described with reference to FIG. 3A. This is different from the fifth embodiment in the square Hough transform (S3). That is, the square Hough transform unit 53 performs Hough transform on one of the binary images corresponding to the vertical line detection component and the horizontal line detection component of the wavelet transform result sent to the frame memory 63.
[0074]
As shown in FIGS. 7 and 8, the wiring pattern of the semiconductor wafer image is aligned in the horizontal or vertical direction, and appears only in either the vertical line component or the horizontal line component. By performing the Hough transformation only on the component where the wiring pattern does not appear, only the detection target can be Hough transformed and can be reliably detected. A wiring pattern is not erroneously detected as a detection target.
[0075]
The operation of the ninth embodiment will also be described with reference to FIG. 3B. In this embodiment, means for automatically determining whether to perform Hough transform on either the vertical line component or the horizontal line component is added to the eighth embodiment. That is, the energy calculating means 52 calculates energy based on the wavelet coefficient sent to the frame memory 61 (S1A), and the direction in which the energy is small is regarded as the direction in which no wiring pattern is detected, and only that direction is Hough transformed. .
[0076]
In FIG. 8, only the wiring pattern + defect appears in the horizontal line detection component at the lower right of the screen, and only the defect appears in the vertical line detection component at the upper right of the screen. The absolute value of the wavelet coefficient is large in the part where the edge is detected, and the other parts have coefficient values close to zero. Accordingly, the degree of detection of the edge component of the vertical line component and the horizontal line component is quantified and compared, and if the degree of detection is smaller, it is regarded as the component in which the wiring pattern is not detected, and it is subjected to Hough transform, whereby FIG. In the case of, only the vertical line component at the upper right of the screen can be Hough transformed to detect the figure to be detected. As a method for quantifying the degree of edge detection, the energy of the wavelet coefficient is defined. This means that each wavelet coefficient value is W (s, t), the number of two-dimensional wavelet coefficients is M in the x-axis direction, and the y-axis. It is defined by equation (5) as N in the direction.
[0077]
E = (1 / MN) Σ_{s = 1} ^MΣ_{t = 1} ^NW (s, t) (5)
The operation of the tenth embodiment will be described with reference to FIG. 3A. This is different from the fifth embodiment in binarization processing (S4). That is, the threshold value of the binarization process (S4) is variable depending on the magnitude of r. For example, it is assumed that a square C1 with a center (a, b) and r = 10 and a square C2 with a center (c, d) and r = 100 exist in the original image. In the parameter space, the luminance value of the parameter (a, b, 10) corresponding to C1 is 4 × 10 = 40. The luminance value of the parameter (c, d, 100) corresponding to C2 is approximately 4 × 100 = 400. In this way, the luminance value of the parameter changes greatly depending on r. By changing the binarization threshold in proportion to r, a square of any size can be detected.
[0078]
The operation of the eleventh embodiment will be described with reference to FIG. 3C. This is because there is overlap detection removal (S7) after the center of gravity calculation (S6) in the ninth embodiment. In this case, when a plurality of squares are detected by the centroid calculation (S6), it is checked whether or not these squares overlap on the xy plane.
[0079]
For example, there may be a case where a noise component is added to a partial arc of a square existing in the xy plane and a square slightly smaller than that is detected. It is determined that it is a false detection to detect a plurality of defects at the same position, and only the larger square is detected, and even if a small square overlapping therewith is detected, it is excluded. Specifically, when multiple squares are detected, a check is made to see if the center of the smaller square is not inside the larger square, and if it is inside the square, it is removed. This is realized.
[0080]
In the embodiment of the present invention, a dedicated frame memory is provided for each processing means. However, as shown in FIG. 12, the frame memory 61 can be shared to save memory and improve efficiency. Is possible.
[0081]
Further, as shown in FIG. 13, it is also possible to replace the devices 41 to 47 with a general-purpose electron microscope system 102, and replace each means 51 to 59 and a frame memory attached to them with a computer 101 such as a workstation.
[0082]
【The invention's effect】
Claims 1 to 3 and claims 6 to 11 correspond to the second object of the present invention. As a result, it is possible to detect a defect even without a golden device or CAD data required in the conventional image comparison method.
[0083]
Claims 4 to 5 and claims 7 to 11 correspond to the first object of the present invention. In the case of an image such as the SEM image described in the embodiment where the pixel value changes with time, the conventional image comparison method detects a large amount of noise. On the other hand, in the present invention, it is possible to detect a defect with almost no noise.
[0084]
Claims 1, 4, 5, 8 and 9 correspond to the third object of the present invention. In the case of the conventional FET method, 1.31 × (10 to the 8th power) addition and subtraction and 8.39 × (10 to the 8th power) times in order to remove the periodic wiring pattern from the 512 × 512 pixel image. However, in the case where the square Hough transform of the present invention is replaced with the circular Hough transform, addition and subtraction of about 5.5 × (10 5) times and 1.1 are performed by introducing wavelet transform means. Processing can be performed by multiplying x (10 6) times. All other image processings except for the Hough transform have a calculation amount of 10 7 or less. Even in the FET method, in order to automatically detect the position and size of an image, processing such as Hough transform is required, so the present invention is not disadvantageous.
[0085]
The circle Hough transform includes a square process and a square root extraction process that are not necessary for the square Hough transform. In the present invention, the Hough transform is limited to the square Hough transform, so that the calculation can be performed faster than the circular Hough transform. For example, in the C compiler of the UltraSparc S-7 / 300U workstation (170 MHz), the circular Hough conversion takes 4650 msec, whereas the square Hough conversion only takes 3441 msec. Therefore, the calculation can be performed at higher speed than the FET method.
[0086]
The calculation amount of the Hough transform is proportional to the number of pixels of the binary image to be processed. Therefore, the amount of calculation is smaller when the number of pixels to be subjected to the Hough transform is smaller. Therefore, the calculation speed can be further improved by performing the noise removal and isolated point removal processing described in claims 4 and 5.
[0087]
As described above in detail, the present invention is capable of producing a defect by composing an apparatus that combines image processing centering on Hough transform and wavelet transform on an image having complicated background information such as a semiconductor wafer pattern. It is possible to automatically detect the position and size of the image, and the effect is great both in terms of defect detection performance and processing speed.
[Brief description of the drawings]
FIG. 1 is a flowchart showing operations of first, second, and third embodiments, respectively.
FIGS. 2A and 2B are flowcharts showing the operations of the fourth, fifth and sixth embodiments, respectively.
FIG. 3 is a flowchart showing an operation of the seventh, eighth and tenth embodiments, B is an operation of the ninth embodiment, and C is an operation of the eleventh embodiment.
FIG. 4 is an explanatory diagram of square Hough transform, where A is a square in (x, y) image space, and B is a Hough curve in (a, b, r) parameter space.
FIG. 5A is an original image of a sample “house”, and B is an image obtained by performing wavelet transform in the x-axis direction on the image of A.
6 is an image obtained by performing wavelet transform in the y-axis direction on the image of FIG. 5B.
FIG. 7 is a sample image simulating a semiconductor wafer image.
8 is an image obtained by performing two-dimensional wavelet transform on the image of FIG.
9 is an image obtained by applying binarization processing to the vertical line detection component (upper right portion) of FIG.
10 is an image obtained by applying an absolute value binarization process to the vertical line detection component (upper right portion) in FIG. 8;
FIG. 11 is a block diagram showing a configuration of an image information processing apparatus according to an embodiment of the present invention.
FIG. 12 is a block diagram showing another configuration of the image information processing apparatus according to the embodiment of the present invention.
FIG. 13 is a block diagram showing still another configuration of the image information processing apparatus according to the embodiment of the present invention.
FIG. 14 is a diagram showing the relationship between a circle and a square when the radius is r0.
FIG. 15 is a diagram illustrating a relationship between a circle and a square when the radius is r = ri.
FIG. 16 is a diagram illustrating a binary image of particle defects on a semiconductor wafer and a square Hough transform result thereof.
[Explanation of symbols]
51 Wavelet transform means
52 Energy calculation means
53 Hough conversion means
54 Binarization processing means
55 Labeling processing means
56 Noise removal means
57 Isolated point removal means
58 Center of gravity calculation means
59 Duplicate detection removal means

Claims (8)

Two-dimensional wavelet transforming means for performing two-dimensional wavelet transform on input digital image data having circular defect image data;
A binarization process for creating a binary image of the vertical line detection component and the horizontal line detection component by applying threshold processing to the vertical line detection component and the horizontal line detection component obtained by the two-dimensional wavelet transform Means,
The binary image obtained by the binarization processing means is applied with a square Hough transform that detects the xy coordinates of the center of the square and the length of one side to circumscribe the circular defect. A square Hough conversion means for detecting the circular defect as a square and determining the position and size of the circular defect ;
An image information processing apparatus. The image information processing apparatus according to claim 1 ,
An image information processing apparatus comprising noise removal means for performing local noise removal on an input digital image. The image information processing apparatus according to claim 2 ,
Isolated point removing means for removing isolated active pixels from the binary image of the vertical line detection component and the horizontal line detection component after applying binarization processing means to the two-dimensional wavelet transform result Image information processing apparatus. The image information processing apparatus according to any one of claims 1 to 3 ,
Square Hough transform means is image information that performs Hough transform on only one of the two binary images corresponding to the vertical line detection component and the horizontal line detection component of two-dimensional wavelet transform. Processing equipment. The image information processing apparatus according to claim 4 .
Energy calculating means for calculating the energy of the vertical line detection component and the horizontal line detection component of the two-dimensional wavelet transform is provided, and a square Hough transform means applies the binary image corresponding to the component having the lower energy to the binary image. An image information processing apparatus that performs square Hough transformation on the image. The image information processing apparatus according to any one of claims 1 to 5 ,
An image information processing apparatus provided with an overlap detection and removal means for removing an overlap when a plurality of the circular defects overlap in position. A two-dimensional wavelet transform process for performing two-dimensional wavelet transform on input digital image data having circular defect image data;
Binarization that creates a binary image of the vertical line detection component and the horizontal line detection component by applying threshold processing to the vertical line detection component and horizontal line detection component obtained by the two-dimensional wavelet transform step. Processing steps;
The binary image obtained by the binarization process is applied with a square Hough transform that detects the xy coordinates of the center of the square and the length of one side to circumscribe the circular defect. A square Hough conversion step of detecting the circular defect as a square and determining the position and size of the circular defect ;
An image information processing method comprising: Two-dimensional wavelet transform processing for two-dimensional wavelet transform on input digital image data having circular defect image data;
Binary processing for creating a binary image of the vertical line detection component and the horizontal line detection component by applying threshold processing to the vertical line detection component and horizontal line detection component obtained by the two-dimensional wavelet transform process Processing,
A square circumscribing the circular defect is applied to the binary image obtained by the binarization process by applying a square Hough transform that detects the xy coordinates of the center of the square and the length of one side. A square Hough transform process for detecting the circular defect and determining the position and size of the circular defect ;
A computer-readable recording medium on which a program for causing a computer to execute is stored.

Images