TW201512649A - Method of chip detects inspecting, system therefor, and computer program product thereof - Google Patents

Abstract

The invention relates to a hybrid-based inspection approach to detect defect in chip. For high intensity defect, simple threshold method is used to detect candidate defect which contains defect, bright spots, weak bright spots and noise by the characterization of high intensity. Then morphology operations are used to filter out small candidate defect which is noise. Then, support vector machine is used to classified high intensity defect. For low intensity defect, a boundary defect detection algorithm is presented to detect low intensity defect. The result shows that the invention can efficiently detect most defects.

Description

Method for detecting wafer image and its system and computer program product

The invention relates to a method, a system and a computer program product for detecting a wafer image defect, in particular to a method for detecting a wafer image used in semiconductor wafer inspection, a system thereof and a computer program product.

With the rapid development of semiconductor process technology, most of the wafer industry relies on manual methods for flaw detection. However, the manual detection method has low detection efficiency and high personnel cost, and it is also caused by factors such as negligence and misjudgment. The correct rate will vary and it is difficult to ensure the quality of the product. And the production speed of the product is getting faster and faster, and it is not enough to load it only by manual inspection.

Therefore, the invention of the optical detecting instrument replaces the naked eye detection, and the image analysis of the captured wafer image is performed by a computer, thereby reducing the time and labor required for performing wafer inspection. At present, various optical detection methods are applied to the detection of semiconductor germanium, and most of the methods require first positioning the wafer to be tested, and then capturing the image of the wafer, and then The image analysis is performed by a general image processing algorithm, and the abnormality of the wafer image is found through image analysis, thereby achieving the function of detecting the wafer defect. For a complete wafer defect detection process, the processing requires a large number of processes, of course, It takes a lot of processing time. In addition, various wafer defect detection devices or methods are currently unable to quickly classify the wafers to be tested.

In recent years, many scholars have proposed different methods to detect flaws on wafers, which are roughly divided into five categories: intensity-based methods, gradient-based methods, frequency-based methods, machine learning-based methods, and hybrid-based hybrids. Methods. The advantages, disadvantages, and practices of the five flaw detection methods are as follows: The strength-based flaw detection method is simple and easy to implement, but it is not easy to detect and cause errors when the intensity of the flaw is not very variable or the image quality is poor. Test results. The gradient-based chirp detection method can effectively detect flaws in low-contrast images, but once the gradient of the texture in the wafer is similar to the gradient change of the tantalum, the tantalum cannot be effectively separated from the wafer. The frequency-based chirp detection method utilizes the regularity of the existence of the wafer texture, and removes the background in the frequency domain to find the flaws in the image. However, when the texture of the tantalum changes irregularly, such methods cannot effectively detect the flaws. . The detection method based on machine learning is to use the training data to obtain the characteristics of the flaw through the learning method without setting any parameters, and then use these characteristics to judge whether there is flaw in the test image, but it is necessary in the training phase. Updating the weights in an iterative manner is more time consuming and time consuming. Hybrid based detection method for different types of sputum It is detected by different kinds of methods, but the calculation amount is relatively large and the program is complicated. With the development of technology, the design of wafers is more and more complicated. For a more complex wafer surface, the effect is limited by using only a single method, but different types of defects are generated in different wafer manufacturing processes. This makes traditional detection methods difficult to analyze. There is currently no single method that can detect high-intensity defects and low-intensity defects of wafers quickly and efficiently.

The invention provides a method for detecting a wafer image, a system thereof and a computer program product for simultaneously detecting high-intensity defects and low-intensity defects of the wafer, and more accurately detecting the wafer defect.

In view of the above, an embodiment of the present invention provides a method for detecting a wafer image defect, comprising the following steps: a first step of capturing a wafer image. In the second step, the wafer image is corrected in the horizontal direction to reduce the error caused by the rotation error. In the third step, a binarization module is used to extract a candidate of high intensity in the wafer image. In the fourth step, the two-phase heteromorphism module is used to remove a small flaw. In a fifth step, a machine learning module is used to classify at least one spot and at least one high intensity defect in the wafer image by a support vector machine. In the sixth step, since the low-intensity 瑕疵 usually occurs at the edge portion of the wafer, the Hough conversion module is used to convert and find the wafer boundary, and the wafer image is divided into the wafer portion and the background portion by the average value of the wafer image intensity to calculate an energy map. . The seventh step is to fill the outline of the wafer image by the type method. The fine ridges in the wafer image are removed and used to obtain a complete wafer boundary profile. And in the eighth step, the Hough transform module is used to find a straight line representing the boundary of the wafer, and the difference between the straight line and the complete wafer boundary contour is calculated to detect the low intensity 瑕疵.

In addition, the two-phase heteromorphism module of the foregoing embodiment includes smoothing the contour of the image by using the disconnection operation and eliminating the small area, and eliminating the small hole and filling the small hole in the contour by the closing operation.

Another embodiment of the present invention is to provide a wafer image detection system for performing the method of the present invention. The wafer image detection system comprises a carrier platform, a camera mechanism and a computer device. The carrier platform is used to place a wafer. The camera mechanism is used to capture a wafer image. The computer device comprises an image display module, a plurality of computing modules, a database and an output device. The computing module is used to analyze and determine the wafer image. The database is used to store the results of previous analyses and to build machine learning materials to speed up the comparison. The output device is used to display the analysis result.

Yet another embodiment of the present invention is to provide a computer program product for performing the method of the present invention. The computer program product stores a plurality of computing modules, including a Kenny edge detector module, a Hough transform module, a binary module, a type module, and a machine learning module. After the computer loads these modules and executes them, the detection of the wafer image defects can be completed.

Thereby, the invention can simultaneously detect high-strength 瑕疵 and low-intensity 晶片 of the wafer, and the invention can detect low-intensity 瑕疵 by detecting high-strength 先 first, and can reduce the calculation amount of each stage after screening by level. Fast and accurate wafer defect detection is achieved. To improve the five most commonly used tests at present. The problems encountered with the wafer defect method.

The Summary of the Invention is intended to provide a simplified summary of the present disclosure in order to provide a basic understanding of the disclosure. This Summary is not an extensive overview of the disclosure, and is not intended to be an

110‧‧‧First steps

120‧‧‧ second step

130‧‧‧ third step

140‧‧‧ fourth step

150‧‧‧ fifth step

160‧‧‧ sixth step

170‧‧‧ seventh step

180‧‧‧ eighth step

200‧‧‧ camera organization

300‧‧‧bearing platform

400‧‧‧ computer equipment

410‧‧‧Image display module

420‧‧‧ Computing Module

421‧‧‧Kenny Edge Detection Module

422‧‧‧Hove Conversion Module

423‧‧‧ Binarization Module

424‧‧‧Type Module

425‧‧‧ machine learning module

430‧‧‧Database

440‧‧‧output device

500‧‧‧ wafer imagery

510‧‧‧ Candidates

511‧‧‧小细

512‧‧‧High-intensity

513‧‧‧Low strength test

520‧‧‧瑕疵 characteristics

521‧‧‧瑕疵 intensity smoothness

522‧‧‧瑕疵Texture complexity

523‧‧‧瑕疵 texture structure

524‧‧‧First edge pixel

525‧‧‧second edge pixel

530‧‧‧ light spots

540‧‧‧ wafer part

550‧‧‧Background section

560‧‧‧Complete wafer boundary contour

570‧‧‧ Straight line

FIG. 1 is a flow chart showing a method for detecting a wafer according to an embodiment of the present invention.

FIG. 2 is a schematic diagram of a process for detecting a wafer cassette according to another embodiment of the present invention.

Part (a) of Figure 3 shows the image of the wafer without tilt; (b) part of the wafer image which is tilted due to the rotation error; and (c) part of the edge of the wafer using the Kenny edge detection module. a schematic diagram represented by a straight line; (d) a schematic diagram of a portion of the candidate region of high intensity; (e) a schematic diagram of an example of a histogram of the image; (f) a partial image of the wafer of binarization; (g) an image There is a schematic diagram of the tiny flaws; (h) is the image of the wafer after removing the fine flaws; (i) is a schematic diagram of the candidate smoothness of the flaw; (j) is partially detected by the Kenny edge detection module. A candidate wafer image map; (k) is a schematic diagram of the candidate tantalum texture structure; (1) is a wafer image map for high-intensity flaw detection.

Part (a) of Figure 4 is a schematic diagram of the xy plane; (b) is a schematic diagram of a point on the xy plane mapped to the ab plane; (c) a schematic diagram of a straight line normal representation.

Figure 5 is a schematic diagram of repairing a wafer image by a type operation, (a) partially binarized wafer image; (b) part of the wafer image after disconnection; and (c) part for closing The image of the wafer after the operation; (d) is a schematic diagram of the set A; (e) is a schematic diagram of the structural element B; (f) is a schematic diagram of the erosion of the set A by the structural element B; A schematic diagram of the result of the expansion operation of the result of the erosion of the set A by the structural element B; (h) part of the original wafer image; (i) part of the schematic of the wafer image after the disconnection operation; (j) part of the closure Schematic diagram of the wafer image after the operation.

Part (a) of Figure 6 is a schematic diagram of the texture structure of the candidate ;; (b) is a texture structure diagram of the 瑕疵; (c) is a texture structure diagram of the weak light spot; (d) ~ (g) part is the edge The pixel ( x, y ) is the four cases of the oblique direction; the (h) part is a schematic diagram showing the method of the mask w ; (i) the part is a schematic diagram of the relative value of the mask w .

Part (a) of Figure 7 is the original wafer image; (b) part of the wafer image for removing the brighter part of the original wafer image; (c) part of the energy map; (d) part of the remaining area contour (f) is a wafer image map that uses a pattern to repair a wafer image to obtain a complete wafer boundary profile; (f) a wafer image portion that is a difference between a predicted wafer boundary and a complete wafer boundary contour.

Please refer to Figure 1 to Figure 7 together. Figure 1 shows the detection chip. A flow chart of the method of detecting ;; Figures 3 through 7 show experimental imagery of the detection process. The method for detecting a wafer cassette of the present embodiment sequentially includes the following steps.

In a first step 110, the wafer is placed on the carrying platform 300, and a wafer image 500 is captured by the camera mechanism 200.

In a second step 120, the wafer image 500 is subjected to horizontal image correction to reduce the analysis result due to the rotation error. As shown in FIG. 3(b), when the wafer image 500 is captured, a rotation error may occur due to the position at which the wafer is placed. First, as shown in FIG. 3(c), a Kenny edge detector module 421 detects the edge of the wafer image 500, and uses a Hough conversion module 422 to find the longest straight line in the wafer image 500 and obtain the angle thereof. Then, the wafer image 500 is rotated and corrected by a two-dimensional rotation formula. The Hough conversion module 422 is a common way to detect straight lines in an image. Considering any point ( x _i , y _i ) on the image, the straight line equation passing through this point can be expressed as a, where a and b are the slopes of the straight line, respectively. Intercept: y _i = ax _i + b .

Can be rewritten as: b =- ax _i + y _i .

For a point (x _i, y _i) and the other point (x _j, y _j) on a straight line, two points in the same straight line on the ab plane intersect at a point (a ', b'), thus the ab plane The point with the highest frequency of occurrence can be regarded as the a and b parameters of the most frequent line segment in the space coordinates. For example, Figure 4 (b) maps the points on the xy plane to the ab plane, where the two lines intersect at (a', b') Point. However, since the slope of the vertical line segment is infinite, it is impossible to use the oblique section to represent the vertical line segment, so the implementation is based on the normal expression, such as the following equation, where ρ is the normal distance from the origin, and θ is the horizontal line. Angle from the vertical line: x cos θ + y sin θ = ρ .

As shown in Fig. 4(c), the method is the same as the oblique truncation described above, and the wafer image 500 is mapped from the xy plane to the ρ θ plane, and the highest frequency (ρ, θ) is obtained in the ρ θ plane. The longest straight line in the wafer image 500, that is, the edge of the wafer, is found, and θ is the desired wafer tilt angle. When the angle θ of the wafer tilt is determined, the wafer image 500 can be corrected by the angle θ so that the edge of the wafer is horizontal. The following equation is used to calculate the position (x', y') of any point (x, y) in the wafer image 500 after being angle corrected:

In a third step 130, a candidate module 510 of high intensity in the wafer image 500 is extracted by a binarization module 423. The binarization module 423 selects a threshold value, and is set to 1 when the grayscale value of the wafer image 500 is greater than the threshold value, and is set to 0 when the grayscale value of the wafer image 500 is less than the threshold value. In order to effectively separate the candidate 瑕疵 510, it is necessary to understand the overall intensity distribution of the wafer image 500. Therefore, the distribution of the intensity in the wafer image 500 is observed by the histogram. For example, the abscissa of the histogram of Fig. 3 (e) represents the intensity scale, and the ordinate indicates the number of pixels appearing on the intensity scale. Since the portion of the wafer image 500 that has high intensity values and significant intensity variations is to be detected, the average of the wafer image 500 can only be used to separate the background portion 550 from the wafer portion 540, for which purpose the intensity on the wafer is calculated. The average value and the intensity standard deviation, the effect of removing the background will be detected, and the threshold value T is calculated by the equation listed below, where I(x, y) is the intensity value of the wafer image 500, and m _c is the wafer image 500. The average value of the intensity, N _c is the number of pixels whose intensity value is greater than the average value of the wafer image 500:

The wafer image 500 calculated by the binarization module 423 is as shown in Fig. 3(f).

In a fourth step 140, a small defect 511 is removed using the two-phase heteromorphism module 424. Because the threshold value of the binarization module 423 is difficult to adjust to a precise data, as shown in FIG. 3(g), the wafer image 500 passing through the binarization module 423 may have some small flaws 511, which are too small or Too slender is considered to be impossible to exist, so the use of the two-phase heteromorphism module 424 removes these fine candidate 瑕疵 510 and eliminates the notch on the contour, making the outline of the candidate 瑕疵 510 more complete. Erosion and expansion are the basic operations of two types of morphology. A is defined by B erosion as:

A is defined by B expansion as:

The set B is called a structural element in morphology. Erosion will cause the object to shrink, and expansion will enlarge the object, as shown in Figure 5 (d) ~ (f), Figure 5 (d) is the set A, Figure 5 (e) is the structural element B, Figure 5 ( f) The action of eroding the set A by the structural elements, and the gray part is the result of the erosion. Figure 4 (g) shows the expansion of the dotted line and the expansion of the gray part. The disconnection and closure operations are two operations that combine expansion and erosion, and are often used in image processing. The disconnection operation is defined by first eroding the object and then expanding:

Conversely, the closing operation is an action of swelling and then eroding the object, defined as:

Figure 5 (h) is a schematic diagram of the original wafer image 500, Figure 5 (i) is a schematic diagram of the original wafer image 500 after the disconnection operation, and Figure 5 (j) is a schematic diagram of the original wafer image 500 after the closing operation. . The break operation smoothes the outline of the image and eliminates small areas; the closed operation effectively eliminates small holes and fills small holes in the outline. Therefore, the fine outline in the wafer image 500 is eliminated by the breaking operation, and the result is as shown in Fig. 5(b). However, there are some small holes in the outline, so the closed hole is used to fill the small hole in the contour to extract the complete contour, and the effect is shown in Fig. 5(c). The wafer image 500 calculated by the two-phased type module 424 is as shown in FIG. 3(h).

In a fifth step 150, at least one spot 530 and at least one high intensity 512 in the wafer image 500 are classified by a feature 520 using a machine learning module 425 in a support vector machine. Both the light spot 530 and the high-strength 瑕疵 512 have high contrast characteristics, but the intensity variation in the light spot 530 is relatively smooth, and since the surface of the wafer is destroyed by the flaw, the surface of the uneven surface is reflected unevenly by the light, so the reflection is very uneven. The standard deviation is used to express the intensity smoothness 521. The standard deviation is mainly used to measure the degree of variation of the grayscale image data. The larger value indicates that the intensity distribution of the wafer image 500 is more dispersed. Using this characteristic, each candidate 瑕疵510 C _i is calculated separately. The standard deviation is one of the 瑕疵 features 520. The standard deviation σ _i formula is as follows, where m _I is the intensity average of the candidate 瑕疵 510 C _i , I(x, y) is the pixel (x, y) intensity value within the candidate 瑕疵 510, and C _i is a set. Containing all pixels (x, y) in the i- th candidate 瑕疵 510, N _i is the number of pixels in C _i ,

As shown in Fig. 3(i), gray indicates a candidate 瑕疵 510 whose standard deviation is larger than the threshold value, and white indicates a candidate 瑕疵 510 whose standard deviation is smaller than the threshold value. In addition, 瑕疵 also usually has more complex textures, such as Figure 3 (j) uses Kenny edge detection module 421 to detect edge and texture complexity 522. The Kenny edge detection module 421 is an optimized edge detection method. Compared with other detection methods, the edge detected by the Kenny edge detection module 421 is the most complete, and the detected edge continuity is good and the positioning is clear. using the formula listed under represented texture complexity, wherein R _ei representative of the i-th candidate flaw texture complexity of 510, the number of edge pixels in the representative N _ei i-th candidate defects 510, N _i is the i th The total number of pixels in the candidate 瑕疵 510: The larger the R _ei , the more complex the texture within the candidate 瑕疵 510 is, and the more likely it is 瑕疵. In addition, as shown in FIG. 6(b), the 瑕疵 texture 523 includes a plurality of edges in an oblique direction; and as shown in FIG. 6(c), the texture existing in the spot is the texture of the original wafer, and most of them are horizontal. And vertical direction. Therefore, as shown in FIG. 3(k), the Kenny edge detection module 421 determines whether the edge direction is horizontal, vertical or oblique. In order to judge whether the edge direction is horizontal, vertical or oblique, the present invention defines four diagonal coordinates ( x -1, y -1), ( x +1, y -1 ) when the Kenny edge pixel ( x , y ) When there are edge pixels in ( x +1, y +1) and ( x -1, y +1), the edge pixels ( x, y ) are oblique directions, as shown in Fig. 6(d)~( g) In either case, define the edge pixel ( x, y ) to be oblique. The edge direction within the candidate 瑕疵 510 is calculated by a spatial filter in practice. As shown in Fig. 6(h), a 3×3 filter mask w is used as an example to determine whether pixels in the 3×3 neighborhood have edge pixels in the diagonal direction, and FIG. 6(h) is a mask w. The relative coefficient value, the coefficient of the four diagonals of the mask is 1, and the remaining five coefficients are zero. The result image f detected by the Kenny edge detection module 421 and the 3×3 w mask are linearly filtered, and the following manners are indicated:

When the response of the spatial filter R(x, y) =1, the direction of the second edge pixel 525 (x, y) is inclined; otherwise, when R (x, y) = 0, the first edge pixel 524 ( x, y) is horizontal and vertical. Figure 6 (a) ~ (c) is a schematic diagram of the texture structure in the candidate 瑕疵 510, wherein the lighter color indicates that the first edge pixel 524 has R(x, y) =0, and the darker color indicates the second. Edge pixel 525 has R(x, y) =1. Since there are more pixels in the oblique direction in the texture, the ratio r _i of the pixels in the oblique direction of the candidate 瑕疵 510 is taken as one of the 瑕疵 features 520, and the 瑕疵 texture 523 features are extracted by the following equation. Where N _ri is the number of pixels in the i- th candidate 瑕疵 510, and N _gi is the number of non-tilt directions of the pixel:

After the 瑕疵 intensity smoothness 521, the 瑕疵 texture complexity 522, and the 瑕疵 texture structure 523 are defined as the 瑕疵 feature 520. Then, the machine learning module 425 is used to learn the training data in the training phase, and then the 瑕疵 and the light spot in the wafer image 500 are distinguished by the 瑕疵 feature through the support vector machine in the test phase, so as to detect the high intensity 瑕疵 512. The purpose of this is shown in Fig. 3 (1). The support vector machine is trying to find the best spatial separation hyperplane for the data to distinguish different samples, maximizing the interval between the hyperplane and the sample, thus minimizing the upper limit of the error. The training samples are ( xi, yi ), i =1, 2, ..., n, xi Rn , yi {+1, -1}, where Rn is the original space of the training sample and the test sample, and xi is the sample vector. Yi is to mark which type of information xi belongs to. To ensure uniqueness, it is assumed that the distance supporting the hyperplane and the best hyperplane is reduced to 1/∥ w ∥, so the training samples have the following restrictions: The above two restrictions can be reduced to a limit, as follows: To get the best separation of the hyperplane, minimize ∥ w ∥ ² , so the target function is:

Use Lagrange Multiplier Method to solve the above problem, find w, bi and αi which can make L the minimum value, αi is Lagrange multiplier, and n is the number of support vectors, where 0≦ αi ≦C, C is used to measure the number of errors. And the parameters of the classification complexity. : get The following decision functions can then be used to classify the data:

However, in practice, the data is usually not divided by using a linear hyperplane, so when a nonlinear problem is encountered, the core data is used to map the original data to the high-dimensional feature space, and then linearly split. The present invention uses a Radial Basis Function (RBF) whose function is: Change the decision function to the radio core function by:

The radiological core function RBF is used to map the data from the original space to the high-dimensional space, so that the data is more easily separated. Then, in the training phase, the three characteristics 520 described above are used as input values of the support vector machine to update the weight to the minimum. As a result of the error, in the test phase, the determined decision function can be used to distinguish between the 瑕疵 and the light spot 530, and it is determined whether or not 瑕疵 is present in the wafer image 500 to achieve the purpose of 瑕疵 detection.

In a sixth step 160, the wafer boundary is found by using the Hough conversion module 422, and the wafer image 500 is divided into the wafer portion 540 and the background portion 550 by the average image intensity of the wafer to calculate an energy map. Low-intensity germanium typically occurs at the edge portion of the wafer, and the wafer boundary is found using the Hough conversion module 422. Since the intensity of the portion of the wafer image 500 will increase the average intensity of the entire wafer image 500, it is difficult to separate the wafer portion 540 and the background portion 550 of the wafer image 500, as shown in FIG. 7(b). A pixel having a pixel intensity value greater than the intensity average of the wafer image 500. Next, the wafer image 500 intensity average is divided into two parts, larger and smaller, the intensity is larger, the wafer portion 540 is represented by the first color; and the intensity is smaller, the background portion 550 is represented by the second color. In order to make the wafer edge obtained by the wafer image 500 smoother, a mask is used to calculate the energy map. When the pixel ( x, y ) has a pixel larger than the average value in the mask, the pixel ( x, y ) is regarded as In the wafer portion, E( x , y ) is set to 255, and vice versa, E( x , y ) is set to 0 for the background portion. The energy map of the wafer image 500 is calculated for any pixel ( x, y ) listed below, and W is a mask having a size of N x N:

The energy map is shown in Figure 7(c). There may be noise or spots 530 in the wafer image 500 such that there are multiple contours in the energy map. To remove the noise and spots 530, the largest area of the area is preserved as in Figure 7(d).

In a seventh step 170, the outline of the wafer image 500 is filled in with the pattern module 424, and the fine flaw 511 in the wafer image 500 is removed to obtain a complete wafer boundary contour 560. As shown in FIG. 7(d), the outline of the wafer is incomplete, there is a hole in the wafer, and there is a noise in the wafer image 500. Therefore, the hole in the wafer area of the wafer image 500 is filled by the closing operation, and then the disconnection operation is performed. The noise in the wafer image 500 is removed. As shown in Fig. 7(e), it can be seen that the outline of the wafer is relatively complete and the edge of the wafer is smoother.

In an eighth step 180, the Hough conversion module 422 is used to find a line 570 representing the boundary of the wafer, and the difference between the line 570 and the full wafer boundary profile 560 is calculated to detect the low intensity 瑕疵 513. The flawless wafer image 500 should have a straight line boundary; however, when there is a low intensity 瑕疵 513 in the wafer image 500, the wafer image 500 has a gap at the wafer boundary, so the Kenny edge detector module 421 is first detected. The complete wafer boundary of Figure 7(e), complete wafer boundary profile 560 as shown in Figure 7(f). The Hough conversion module 422 then finds a line 570 that best represents the wafer boundary as the predicted wafer boundary. Calculating the difference between this line 570 and the full wafer boundary profile 560, when there is no flaw in the wafer boundary in the wafer image 500, the detected full wafer boundary profile 560 should be not much different from the predicted wafer edge. However, when there is a defect in the boundary of the wafer, there is a gap in the boundary of the wafer, and there is a significant distance difference between the boundary of the wafer and the predicted wafer boundary. When the number of pixels with a large gap is larger than the threshold value, the wafer boundary is considered to have a gap, and the wafer image will be tested. 500 is judged as a 瑕疵 image, and the detection of low intensity 瑕疵 513 is achieved.

Another embodiment of the present invention is to provide a wafer image detection system for performing the method of the present invention. The wafer image detection system includes a carrier platform 300, a camera mechanism 200, and a computer device 400. A wafer is placed on the carrying platform 300, a wafer image 500 is captured by the camera mechanism 200, and the wafer image 500 is electrically connected to the computer device 400. The computer device 400 includes an image display module 410, a plurality of computing modules 420, a database 430, and an output device 440. After the wafer image 500 is analyzed and judged by the operation module 420, the previous analysis result is stored in the database 430, and the machine learning data is established to speed up the backward comparison speed, and the final detection result is displayed on the output device 440. The computing module 420 includes a Kenny edge detector module 421, a Hough transform module 422, a binarization module 423, a type module 424, and a machine learning module 425.

Yet another embodiment of the present invention is to provide a computer program product for performing the method of the present invention. The computer program product stores a plurality of computing modules 420, including a Kenny edge detector module 421 and a Hough transform module. The group 422, a binarization module 423, a type module 424, and a machine learning module 425 can perform detection of the wafer image 500 当 when the computer loads the modules and executes them.

Although the present invention has been disclosed in the above embodiments, it is not intended to limit the present invention, and the present invention can be modified and modified without departing from the spirit and scope of the present invention. The scope is subject to the definition of the scope of the patent application attached.

110‧‧‧First steps

120‧‧‧ second step

130‧‧‧ third step

140‧‧‧ fourth step

150‧‧‧ fifth step

160‧‧‧ sixth step

170‧‧‧ seventh step

180‧‧‧ eighth step

Claims (10)

A method for detecting a wafer image includes the following steps: a first step of capturing a wafer image; a second step of correcting the wafer image in a horizontal direction; and a third step of extracting the wafer image by using a binarization module a candidate for medium-high intensity; in the fourth step, the two-phase heteromorphic module is used to remove a small flaw; in the fifth step, a machine learning module is used to classify the support vector machine by a feature. At least one spot and at least one high-intensity flaw in the wafer image; in the sixth step, the wafer boundary is converted by using a Hough transform module, and the wafer image is divided into a wafer portion and a background portion by using the average image intensity of the wafer To calculate an energy map; in a seventh step, the contour of the wafer image is filled incompletely by the patterning method, and the fine flaw in the wafer image is removed to obtain a complete wafer boundary contour; and an eighth step is Using the Hough transform module to find a straight line representing the boundary of the wafer, and calculating the difference between the straight line and the boundary contour of the complete wafer for detecting Low intensity flaws. The method for detecting a wafer image according to claim 1, wherein the second step detects the edge of the image by a Kenny edge detector module, and uses the Hough transform module to find the longest image in the wafer. a straight line and find its angle, and then use a two-dimensional rotation formula to rotate the wafer image positive. The method for detecting a wafer image according to claim 1, wherein the pattern modules comprise a disconnect operation and a close operation. The method for detecting a wafer image according to claim 1, wherein the 瑕疵 feature comprises 瑕疵 intensity smoothness, 瑕疵 texture complexity, and 瑕疵 texture structure. The method for detecting a wafer image according to claim 1, further comprising an adjusting step performed between the fifth step and the sixth step, wherein removing the pixel intensity value in the wafer image is greater than the image intensity of the wafer A pixel of the average for the energy map to effectively separate the wafer portion from the background portion. The method for detecting a wafer image according to claim 1, wherein the energy map is divided into two parts of a larger and a smaller part by an average value of the image intensity of the wafer, and the portion with a higher intensity is represented by a first color, and the intensity is smaller. The part is represented by the second color. The method for detecting a wafer image according to claim 1, wherein in the eighth step, the number of image pixels having a large gap between the line and the boundary contour of the wafer is less than a threshold value, and the line is determined to be absent. When the number of the image pixels having a large difference from the boundary contour of the wafer is greater than the threshold value, it is determined as a single image. A method for detecting a wafer image, the method of claim 1, wherein the wafer image detection system comprises: a carrier platform for placing a wafer; a camera mechanism for capturing a wafer image; and a The computer device comprises: an image display module; a plurality of computing modules for analyzing and determining the image of the wafer; a database for storing an analysis result and establishing machine learning data to speed up the backward comparison speed; and an output The device is used to display the analysis result. The method for detecting a wafer image defect according to claim 8, wherein the computing module comprises a Kenny edge detector module, a Hough transform module, a binarization module, and a type Module and a machine learning module. a computer program product for detecting wafer images, when the computer is loaded with a Kenny edge detector module, a Hough conversion module, a binary module, a type module and After a machine learning module is executed, the method described in claim 1 can be completed.