Method and system for detecting defects on semiconductor wafers based on machine vision

Abstract

The application discloses a kind of semiconductor wafer defect detection method and system based on machine vision, it is related to wafer defect detection technical field, its method includes, obtaining the semiconductor wafer image to be detected;Wafer image is extracted based on wafer image;Based on wafer image and preset image, the specific position of existing defect is obtained, and the preset image is the wafer image without defect;According to the specific position of existing defect, the morphological characteristics of defect are determined;According to the morphological characteristics of defect, the type of defect is determined, the present application detects the defect of semiconductor wafer based on machine vision, uses machine to replace human eye to do measurement and judgment, solve the problem that misjudgment or detection speed is slow when detecting by artificial through microscopic system, simultaneously based on wafer image carries out image processing, avoid the interference of irrelevant area, improve detection efficiency and accuracy.

Description

Technical Field

[0001] This invention relates to the field of wafer defect detection technology, and more specifically to a semiconductor wafer defect detection method and system based on machine vision. Background Technology

[0002] With the development of the semiconductor industry, the demand for semiconductor wafer production and manufacturing is increasing day by day. However, in the process of semiconductor wafer production and preparation, quality defects are inevitable. Therefore, factories have to use tools such as microscopes to manually observe the products.

[0003] For example, Chinese invention patent CN114160450A provides a semiconductor wafer inspection system and its inspection method. This invention inspects wafers through a microscopic system. However, when inspecting wafers manually through a microscopic system, problems such as misjudgment or slow inspection speed are likely to occur. Therefore, the existing technology has shortcomings. Summary of the Invention

[0004] To address the shortcomings of existing technologies, the present invention aims to provide a semiconductor wafer defect detection method and system based on machine vision. The method converts semiconductor wafers into image signals using machine vision products, transmits them to dedicated image processing equipment, and further processes the images to achieve the detection of semiconductor wafer defects.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] This invention provides a semiconductor wafer defect detection method based on machine vision, comprising:

[0007] Acquire an image of the semiconductor wafer to be inspected;

[0008] Extract grain images based on the wafer images;

[0009] Based on the grain image and the preset image, the specific location of the defect is obtained, wherein the preset image is a grain image without defects;

[0010] Based on the specific location of the defect, determine the morphological characteristics of the defect;

[0011] The type of defect is determined based on its morphological characteristics.

[0012] As a further improvement of the present invention, the extraction of the grain image based on the wafer image includes:

[0013] The wafer image is convolved with a Gaussian kernel to obtain a Gaussian pyramid, and the adjacent layers of the Gaussian pyramid are subtracted to obtain a Gaussian difference pyramid.

[0014] Each pixel within the Gaussian difference pyramid is compared with all its neighboring pixels to determine the extreme points;

[0015] The Gaussian difference pyramid is expanded at the extreme points using the Taylor formula, and the offset of the extreme points is calculated. Extreme points with offsets less than a preset value are used as feature points.

[0016] A feature vector is constructed based on the feature points, and the feature vector is matched with a preset feature vector to obtain the grain image. The preset feature vector is calculated based on the preset image.

[0017] As a further improvement of the present invention, constructing a feature vector based on the feature points includes:

[0018] Centered on the feature point, a circular region within a preset radius is selected, and the circular region is located inside the region where the grain is located;

[0019] The circular region is divided into m annular regions, where m is a positive integer greater than 1;

[0020] The circular region is divided into n intervals, each interval having the same angle, resulting in m×n sub-regions, where n is a positive integer greater than 1;

[0021] Within each sub-region, the magnitude of each pixel is calculated, and the magnitudes of each pixel are weighted to obtain the feature value of the sub-region;

[0022] The feature values ​​of each sub-region are combined to obtain the feature vector of the feature point.

[0023] As a further improvement of the present invention, the formula for calculating the amplitude is as follows: Where (x, y) are the coordinates of the pixel, and σ is the scale space factor. and They are respectively Difference in the horizontal and vertical directions, in and These are the first-order differences of the Gaussian pyramid along the horizontal and vertical directions, respectively.

[0024] As a further improvement of the present invention, the step of obtaining the specific location of the defect based on the grain image and the preset image includes:

[0025] The grain image is decomposed into an illumination response model, and the illumination response model is then subjected to a logarithmic transformation.

[0026] The result of the logarithmic transform is then subjected to a Fourier transform and frequency domain filtering.

[0027] The result of the frequency domain filtering is converted into the spatial domain to obtain the target image;

[0028] Based on the target image and the preset image, the specific location of the defect is obtained.

[0029] As a further improvement of the present invention, the specific location of the defect is obtained based on the target image and the preset image, including:

[0030] The target image and the preset image are converted into grayscale images to obtain a grayscale target image and a grayscale preset image.

[0031] The difference between the grayscale target image and the grayscale preset image is obtained by subtracting the grayscale target image from the grayscale preset image.

[0032] The specific location of the defect is determined based on the difference.

[0033] As a further improvement of the present invention, determining the specific location of the defect based on the difference includes:

[0034] If the difference is negative, the negative number is compared with a preset threshold.

[0035] If the negative number is greater than or equal to a preset threshold, the absolute value of the negative number is taken; if the negative number is less than the preset threshold, the negative number is set to zero, and the updated difference is obtained.

[0036] The specific location of the defect is determined based on the updated difference.

[0037] As a further improvement of the present invention, determining the type of the defect based on its morphological characteristics includes:

[0038] The complexity and texture information of the defect are determined based on its grayscale entropy.

[0039] The class measure of the defect is calculated based on the contour dimensions of the defect;

[0040] The integrity of the defect is calculated based on its area and outline length.

[0041] The type of defect is determined based on the complexity and texture information, the class moment, and the completeness.

[0042] As a further improvement of the present invention, the formula for calculating the gray entropy is: gary_entropy=-∑(P(i)×log2(P(i))), where i is the gray value of the defect and P(i) is the normalized frequency of the gray value i.

[0043] This invention provides a semiconductor wafer defect detection system based on machine vision, comprising:

[0044] Acquisition module: Used to acquire images of the semiconductor wafer to be inspected;

[0045] Matching module: used to extract grain images from the wafer image;

[0046] The positioning module is used to determine the specific location of the defect based on the grain image and a preset image, wherein the preset image is a grain image without defects.

[0047] Classification module: used to determine the morphological characteristics of the defect based on its specific location, and to determine the type of the defect based on its morphological characteristics.

[0048] This invention acquires images of semiconductor wafers using machine vision products, extracts grain images from the wafer images, performs image processing on the grain images, and detects wafer defects based on the results of the image processing. In other words, it uses machines to replace human eyes for measurement and judgment, which improves detection efficiency and makes up for the shortcomings of manual inspection methods. Attached Figure Description

[0049] Figure 1 This is a flowchart of the method steps of the present invention;

[0050] Figure 2 This is a flowchart illustrating the process of extracting grain images.

[0051] Figure 3 This is a schematic diagram of a circular area;

[0052] Figure 4 This is a schematic diagram of the angle interval;

[0053] Figure 5 This is a schematic diagram of the system structure of the present invention. Detailed Implementation

[0054] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof.

[0055] The term "and / or" in the following text is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. Additionally, the character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0056] like Figure 1 As shown in the figure, this application provides a semiconductor wafer defect detection method based on machine vision, including:

[0057] Acquire an image of the semiconductor wafer to be inspected;

[0058] Extracting grain images from wafer images;

[0059] Based on the grain image and the preset image, the specific location of the defect is obtained. The preset image is a grain image without defects.

[0060] Determine the morphological characteristics of the defect based on its specific location;

[0061] The type of defect is determined based on its morphological characteristics.

[0062] The semiconductor wafer defect detection method based on machine vision provided in this embodiment acquires images of the semiconductor wafer to be inspected using machine vision products, such as CMOS and CCD. The images of the semiconductor wafer to be inspected are converted into image signals and transmitted to a dedicated image processing device. The wafer defect morphology information is obtained through image processing, and the wafer defect is finally detected. Furthermore, the machine vision-based detection method can improve the productivity of repetitive tasks, solve the problems of misjudgment or slow detection speed caused by visual fatigue in manual inspection, and make up for the shortcomings of the prior art.

[0063] A wafer is a crucial component of semiconductors, containing tens of thousands of dies. Industrial wafer defect detection essentially involves detecting die defects. Each die is surrounded by a street-like background area, which is considered irrelevant. In image processing, including operations on these irrelevant areas leads to high computational load and long processing times. Therefore, it's necessary to extract the complete die image while removing the street background. Further, such as... Figure 2 As shown, this embodiment provides a method for extracting a die image from a wafer image, which improves computational efficiency while avoiding additional interference from irrelevant regions. Specifically, it includes:

[0064] The wafer image is convolved with a Gaussian kernel to obtain a Gaussian pyramid, and the adjacent layers of the Gaussian pyramid are subtracted to obtain a Gaussian difference pyramid.

[0065] Each pixel within the difference of Gaussian pyramid is compared with all its neighboring pixels to determine the extreme points;

[0066] The Gaussian difference pyramid is expanded at the extreme points using the Taylor formula, and the offset of the extreme points is calculated. Extreme points with offsets less than a preset value are used as feature points.

[0067] A feature vector is constructed based on the feature points, and the feature vector is matched with a preset feature vector to obtain a grain image. The preset feature vector is calculated based on the preset image.

[0068] Specifically, the Gaussian convolution kernel is Where (x, y) are the coordinates of the pixel in the wafer image, and σ is the scale space factor, also known as the blur coefficient. The larger the value of σ, the more blurred the image. Convolving the wafer image with a Gaussian kernel yields a Gaussian pyramid. Where I(x,y) is the wafer image, and the difference-of-Gaussian pyramid is obtained by subtracting adjacent layers (Gaussian images of adjacent scales) of the Gaussian pyramid. in, denoted as convolution, and k represents the scale.

[0069] All adjacent pixels include 8 neighboring pixels at the same scale and 9×2 pixels at the adjacent upper and lower scales, totaling 26 points. Extrema can be determined using the Difference of Gaussian (GOG) method, which involves subtracting the Gaussian filtering results under different parameters using the difference of a Gaussian function to find the extrema. Then, the Gaussian pyramid is expanded at the extrema using the Taylor series. Taking the derivative of the Taylor expansion and setting it to zero yields the offset of that extrema. When the offset of the extrema in any dimension (x, y, or σ) is less than a preset value, the extrema is designated as a feature point. Multiple extrema and feature points are possible; preferably, the preset value can be set to 0.5.

[0070] Furthermore, this embodiment provides a method for constructing feature vectors, including:

[0071] A circular region with a preset radius is selected centered on the feature point. The circular region is located inside the region where the grain is located.

[0072] Divide the circular region into m annular regions, where m is a positive integer greater than 1;

[0073] Divide the circular region into n intervals, each interval having the same angle, to obtain m×n sub-regions, where n is a positive integer greater than 1;

[0074] Within each sub-region, the magnitude of each pixel is calculated, and the magnitudes of each pixel are weighted to obtain the feature value of the sub-region;

[0075] The feature values ​​of each sub-region are combined to obtain the feature vector of the feature point.

[0076] The preset radius needs to be adjusted according to the size of the grain. To avoid interference from irrelevant areas, the circular area should be located inside the area where the grain is located.

[0077] For example, such as Figure 3 As shown, within a 21×21 neighborhood around the feature point, a circular region with a radius of 10 is taken, centered on the feature point. This circular region is then divided into four annular regions from the inside out. Since the closer the pixels around the feature point are to the feature point, the greater their influence on the calculated feature value, the step size of each annular sub-region is set to 4, 3, 2, and 1, respectively. Figure 4 As shown, the circular region is further divided into 8 equal intervals, each with an angle of 45 degrees, resulting in 32 sub-regions. Within each sub-region, the amplitude of each pixel is calculated.

[0078] The specific formula for calculating the amplitude is as follows: Where (x, y) are the coordinates of the pixel, and σ is the scale space factor. and They are respectively Difference in the horizontal and vertical directions, in These are the first-order differences of the Gaussian pyramid along the horizontal and vertical directions, respectively, where S is the SOBEL operator. x and S y Let S be the partial derivatives of S in the horizontal and vertical directions, respectively.

[0079] Then, the magnitude of each pixel within each sub-region is weighted, with the weight parameter being... Where (m0,n0) represents feature points, (m,n) represents pixel points, and W m,n Let s be the weight coefficient of the pixel, a be the step size of the annular region containing the pixel, a be the radius of the annular region containing the pixel, and k be the scale of the pixel, i.e., the level of the feature point in the Gaussian pyramid. The ratio of the step size s to the radius a of the annular region containing the pixel describes the position of the annular region relative to the central feature point region, further refining the influence weight of pixels at different positions on the central feature point. Then, the feature values ​​of each sub-region are combined to obtain a 4×8=32-dimensional feature vector for the feature point, and this feature vector is normalized to obtain the final feature vector.

[0080] The final feature vector is matched with the preset feature vector to obtain the grain image. Since there are multiple feature points, there are also multiple feature vectors. The matching method is to calculate the Euclidean distance between each feature vector and the preset feature vector. If the distance is less than the preset value, the match is successful. The circular area corresponding to the successfully matched feature vector is the extracted grain image.

[0081] In existing technologies, feature vectors are typically generated by using a 16×16 square region within the neighborhood of a feature point to generate a 128-dimensional feature vector. However, high-dimensional feature vectors require a large amount of computation for matching, which can severely affect the matching speed and cause significant matching errors. Therefore, this embodiment uses a circular region to generate feature vectors. Circularity is rotationally invariant, which not only fully preserves feature information but also eliminates the need to determine the principal direction of the feature points. Furthermore, the dimension of the feature vector can be effectively adjusted as needed, meeting the extraction requirements of different wafer images. This allows for accurate extraction of the grain image based on the original wafer image, and further image analysis can be performed on the grain image. Compared to analysis performed directly on the original image, this avoids interference from street background areas and can more accurately identify wafer defect areas.

[0082] In the field of machine vision, when acquiring images using CMOS or CCD, the images are inevitably affected by lighting, environmental object collisions, and image transmission, resulting in noise in most acquired images. Therefore, image denoising and enhancement are necessary steps in image processing. Existing image denoising and enhancement methods typically employ spatial domain-based processing techniques such as Gaussian filtering and median filtering. However, spatial domain-based processing methods are prone to robustness issues. Therefore, this embodiment further provides a method for denoising and enhancing grain images, employing a frequency domain-based processing method that focuses more on the overall picture rather than local details. Specifically, it includes:

[0083] The grain image is decomposed into an illumination response model, and the illumination response model is then subjected to a logarithmic transformation.

[0084] The result of the logarithmic transform is then subjected to a Fourier transform and frequency domain filtering.

[0085] The result of frequency domain filtering is converted into the spatial domain to obtain the target image.

[0086] Specifically, for a grain image f(x, y), it can be decomposed into an illumination response model:

[0087] f(x, y) = i(x, y)r(x, y)

[0088] Where i(x, y) represents the illumination component at position (x, y) in f(x, y), and r(x, y) represents the reflection component at position (x, y) in f(x, y). A logarithmic transformation of the above illumination response model yields:

[0089] z(x,y)=lnf(x,y)=lni(x,y)+lnr(x,y)

[0090] Where z(x, y) is the illumination response model after logarithmic transformation, and then the above results are subjected to Fourier transform to obtain:

[0091] F{z(x, y)}=F{lni(x, y)}+F{lnr(x, y)}

[0092] Where F{z(x,y)} represents z(x,y) after Fourier transform, the above equation can be simplified to: Z(u,v)=F i (u, v) + F r (u, v), where F i (u, v) is the result of the Fourier transform of lni(x, y), F r (u, v) is the result of the Fourier transform of lnr(x, y), and Z(u, v) is the result of the Fourier transform of z(x, y). Frequency domain filtering of Z(u, v) yields:

[0093] B(u,v)=H(u,v)Z(u,v)=H(u,v)F i (u, v) + H(u, v)F r (u, v)

[0094] Where B(u, v) represents the result of frequency domain filtering of Z(u, v), and H(u, v) represents the frequency domain filter, which transforms the result after the above filter H(u, v) filtering operation from the frequency domain to the spatial domain:

[0095] b(x, y) = F -1 B(u, v) = F -1 (H(u,v)F i (u, v))+F -1 (H(u,v)F r (u, v))

[0096]

[0097] The resulting g(x, y) is the target image.

[0098] The method provided in this embodiment removes noise and enhances image details by performing frequency domain processing on the image. At the same time, the impact of uneven illumination is reflected in the low-frequency information of the spectrum, which is not considered during calculation. Therefore, it is not affected by uneven illumination and has higher robustness. The obtained target image will serve as the basis for subsequent wafer defect detection.

[0099] Furthermore, this embodiment provides a method for comparing a target image with a preset image to obtain the specific location of the defect, including:

[0100] Convert the target image and the preset image into grayscale images to obtain a grayscale target image and a grayscale preset image;

[0101] The difference between the grayscale target image and the grayscale preset image is obtained by subtracting the grayscale target image from the grayscale preset image.

[0102] Based on the difference, the specific location of the defect is determined.

[0103] Specifically, in order to simplify image storage while preserving image details, grayscale processing can be performed on color images. The processing method can be the averaging method, which uses the average value of the three channels in the RGB space at each pixel as the grayscale value of that pixel.

[0104] The preset image is a template grain image without any defects. If the target image does not contain defects, the grayscale information at corresponding positions in the two images is the same. If the target image contains defects, then the grayscale information at corresponding positions is the same except for the grayscale information at the defect positions.

[0105] Next, the grayscale target image and the grayscale preset image are subtracted, that is, the difference between corresponding pixels in spatial location is calculated. The grayscale preset image is used as the subtracted image, and the grayscale target image is used as the subtrahend image. The difference in grayscale values ​​is calculated. Since the grayscale value of each pixel is in the range of [0, 255], the difference in grayscale values ​​of each corresponding pixel is between [-255, 255]. If the difference is between [0, 255], it is called a positive difference; if the difference is between [-255, 0), it is called a negative difference. That is, when the grayscale value of a pixel in the subtracted image is less than the grayscale value of the subtrahend image, or in other words, the former is darker than the latter, a negative difference will occur. The negative difference exceeds the protection range of grayscale values, which will cause the difference information between the two images to be lost. Therefore, a certain method can be used to convert negative values ​​into positive values, thereby preserving the difference information.

[0106] Common conversion methods include the zeroing method and the absolute value method. The zeroing method sets the negative difference to zero, while the absolute value method takes the absolute value of the negative difference to obtain the positive difference. However, in actual manufacturing processes, due to manufacturing processes and other reasons, the pixels at corresponding positions in two images are not completely identical, which may cause interference when performing the difference calculation. Therefore, this embodiment further provides a method for updating negative difference values, including:

[0107] If the difference is negative, compare the negative number with a preset threshold.

[0108] If the negative number is greater than or equal to the preset threshold, the absolute value of the negative number is taken; if the negative number is less than the preset threshold, the negative number is set to zero, and the updated difference is obtained.

[0109] Based on the updated difference, the specific location of the defect is determined.

[0110] The preset threshold is determined according to the detection requirements.

[0111] The specific formula is as follows:

[0112]

[0113] Where res1(x,y) is the difference obtained by the zeroing method, res2(x,y) is the difference obtained by the absolute value method, res(x,y) is the updated difference, c(x,y) represents the grayscale preset image, g1(x,y) represents the grayscale target image, and K is the preset threshold.

[0114] The method provided in this embodiment determines the specific location of the defect in the grain image by calculating the difference between the corresponding positions of the two images, and provides a method for updating negative values, which can retain complete difference information and make the detection results more accurate.

[0115] Furthermore, after determining the specific location of the defect, adjacent defect pixels can be connected to obtain the region shape of the defect. Based on the morphological characteristics of the defect, the defect type can be determined, specifically including:

[0116] The complexity and texture information of a defect are determined based on its grayscale entropy.

[0117] Calculate the class degree of the defect based on its outline dimensions;

[0118] The integrity of a defect is calculated based on its area and outline length.

[0119] The type of defect is determined based on complexity, texture information, class moment, and completeness.

[0120] Currently, the types of defects in crystals can be broadly categorized into wafer fragments, arc discharge ablation, external contamination, corrosion pits, and redundant structures. Wafer fragments are mainly caused by wafer breakage and scratches, and their appearance is often angular. Arc discharge ablation is mainly caused by high concentrations of charged ions accumulating charge in non-conductive areas of the crystal surface through an electric arc in a plasma environment, and its appearance is often a charred, blurred-edge appearance. External contamination is generally in the form of fine particles, some irregular in shape, and some in the form of strips or near-circular shapes. Corrosion pit defects generally appear as cavities on the metal substrate or circuit, caused by etching by the developer or ACT tank. Redundant structure defects generally appear as abnormal bridging between adjacent structures or circuit structures appearing in incorrect locations, and the morphology of this type of defect is similar to that of a normal circuit structure.

[0121] Therefore, it can be concluded that the type of grain defect is strongly correlated with the morphological characteristics of grain defect. The type of grain defect can be determined based on the morphological characteristics of grain defect, and then the type of wafer defect can be obtained, thus realizing the detection of wafer defects. Furthermore, since a wafer contains multiple grains, a wafer may contain multiple defects.

[0122] Specifically, the formula for calculating grayscale entropy is as follows:

[0123] gary_entropy=-∑(P(i)×log2(P(i)))

[0124] Where i represents the gray value at the location of the grain defect, P(i) is the normalized frequency of the gray value i, and gray entropy is a parameter that measures the gray information of the grain defect region, representing the complexity and texture information of the grain defect. A higher gray entropy value indicates that the grain defect has more information and texture variation, while a lower gray entropy value indicates that the grain defect is relatively smooth with less texture variation. For types such as arc discharge ablation, external contamination, and corrosion pits, the gray entropy is generally higher.

[0125] The specific formula for calculating the class moment is as follows:

[0126] rectangularity=area / rectangularity_area

[0127] Where rectangularity is the similarity of the grain shape and area is the area of ​​the defect. Rectangularity_area represents the minimum rectangular outline size that encloses the grain defect region. Its value range is generally 0-1. The similarity parameter can effectively represent the similarity between the defect outline and the rectangle. For some defect types, such as redundant structural defects, the similarity is generally higher.

[0128] The formula for calculating completeness is as follows:

[0129] roundness=4π×area / perimeter 2

[0130] In this context, roundness represents integrity, and perimeter represents the length of the defect profile. The value of roundness is usually between 0 and 1. The closer the value of roundness is to 1, the closer the outline shape of the grain defect is to a circle. When the edge of the defect is obviously serrated, its roundness is low, indicating that the grain may have a fragmented wafer type defect.

[0131] Several judgment methods can also be combined to determine the defect type. For example, if the defect has high integrity and high gray entropy, the defect is more likely to be a redundant structure. If the defect has low integrity and low gray entropy, the defect is more likely to be a broken wafer.

[0132] The method provided in this implementation can preliminarily determine the type of defect based on its morphological characteristics, but it cannot accurately determine the specific type of defect. If a more accurate type needs to be defined, a more accurate computational model, such as a neural network, is required.

[0133] Furthermore, such as Figure 5 As shown in the figure, this application provides a semiconductor wafer defect detection system based on machine vision, including:

[0134] Acquisition module: Used to acquire images of the semiconductor wafer to be inspected;

[0135] Matching module: used to extract grain images based on wafer images;

[0136] The positioning module is used to determine the specific location of defects based on the grain image and a preset image, where the preset image is a grain image without defects.

[0137] Classification module: Used to determine the morphological characteristics of defects based on their specific locations, and to determine the type of defects based on their morphological characteristics.

[0138] The acquisition module is specifically a machine vision device, such as CMOS and CCD, while the matching module, comparison module, positioning module and classification module are located in the image processing device. After the acquisition module acquires the image, it transmits the image to the image processing device for image analysis.

[0139] The semiconductor wafer defect detection method and system based on machine vision provided in this application obtains a target image by extracting a grain image and performing image denoising and enhancement processing. The location of the defect in the target image is determined based on a preset image, and the type of defect is preliminarily determined based on the morphological characteristics of the defect. This improves the detection speed and makes up for the shortcomings of manual inspection.

[0140] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0141] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1A device that provides the functions specified in one or more boxes.

[0142] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0143] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0144] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A semiconductor wafer defect detection method based on machine vision, characterized in that, include: Acquire an image of the semiconductor wafer to be inspected; Extract grain images based on the wafer images; Based on the grain image and the preset image, the specific location of the defect is obtained, wherein the preset image is a grain image without defects; Based on the specific location of the defect, determine the morphological characteristics of the defect; Based on the morphological characteristics of the defect, determine the type of the defect; The step of extracting the grain image based on the wafer image includes: The wafer image is convolved with a Gaussian kernel to obtain a Gaussian pyramid, and the adjacent layers of the Gaussian pyramid are subtracted to obtain a Gaussian difference pyramid. Each pixel within the Gaussian difference pyramid is compared with all its neighboring pixels to determine the extreme points; The Gaussian difference pyramid is expanded at the extreme points using the Taylor formula, and the offset of the extreme points is calculated. Extreme points with offsets less than a preset value are used as feature points. A feature vector is constructed based on the feature points, and the feature vector is matched with a preset feature vector to obtain the grain image. The preset feature vector is calculated based on the preset image. The construction of feature vectors based on the feature points includes: Centered on the feature point, a circular region within a preset radius is selected, and the circular region is located inside the region where the grain is located; Divide the circular region into A circular region, It is a positive integer greater than 1; Divide the circular region into Given several intervals, each with the same angle, we obtain... Sub-regions It is a positive integer greater than 1; Within each sub-region, the amplitude of each pixel is calculated, and the amplitudes of each pixel are weighted to obtain the feature value of the sub-region. The weighting parameter is... ,in Representing feature points, Represents pixels, This represents the weight coefficient for that pixel. The step size is the annular region containing the pixel. The radius of the annular region containing the pixel is... The scale of the pixel; The feature values ​​of each sub-region are combined to obtain the feature vector of the feature point; The step of determining the specific location of the defect based on the grain image and the preset image includes: The grain image is decomposed into an illumination response model, and the illumination response model is then subjected to a logarithmic transformation. The result of the logarithmic transform is then subjected to a Fourier transform and frequency domain filtering. The result of the frequency domain filtering is converted into the spatial domain to obtain the target image; Based on the target image and the preset image, the specific location of the defect is obtained.

2. The semiconductor wafer defect detection method based on machine vision according to claim 1, characterized in that, The specific formula for calculating the amplitude is as follows: ,in The coordinates of the pixel are... For scale space factor, and They are respectively Difference in the horizontal and vertical directions, ,in and These are the first-order differences of the Gaussian pyramid along the horizontal and vertical directions, respectively.

3. The semiconductor wafer defect detection method based on machine vision according to claim 1, characterized in that, Based on the target image and the preset image, the specific location of the defect is obtained, including: The target image and the preset image are converted into grayscale images to obtain a grayscale target image and a grayscale preset image. The difference between the grayscale target image and the grayscale preset image is obtained by subtracting the grayscale target image from the grayscale preset image. The specific location of the defect is determined based on the difference.

4. The semiconductor wafer defect detection method based on machine vision according to claim 3, characterized in that, Determining the specific location of the defect based on the difference includes: If the difference is negative, the negative number is compared with a preset threshold. If the negative number is greater than or equal to a preset threshold, the absolute value of the negative number is taken; if the negative number is less than the preset threshold, the negative number is set to zero, and the updated difference is obtained. The specific location of the defect is determined based on the updated difference.

5. The semiconductor wafer defect detection method based on machine vision according to claim 1, characterized in that, Based on the morphological characteristics of the defect, the type of the defect is determined, including: The complexity and texture information of the defect are determined based on its grayscale entropy. The class measure of the defect is calculated based on the contour dimensions of the defect; The integrity of the defect is calculated based on its area and outline length. The type of defect is determined based on the complexity and texture information, the class moment, and the completeness.

6. The semiconductor wafer defect detection method based on machine vision according to claim 5, characterized in that, The specific formula for calculating the grayscale entropy is as follows: ,in The grayscale value of the defect is... grayscale value The normalized frequency.

7. A semiconductor wafer defect detection system based on machine vision, characterized in that, include: Acquisition module: Used to acquire images of the semiconductor wafer to be inspected; Matching module: used to extract grain images from the wafer image; The positioning module is used to determine the specific location of the defect based on the grain image and a preset image, wherein the preset image is a grain image without defects. Classification module: used to determine the morphological characteristics of the defect based on its specific location, and to determine the type of the defect based on its morphological characteristics; The step of extracting the grain image based on the wafer image includes: The wafer image is convolved with a Gaussian kernel to obtain a Gaussian pyramid, and the adjacent layers of the Gaussian pyramid are subtracted to obtain a Gaussian difference pyramid. Each pixel within the Gaussian difference pyramid is compared with all its neighboring pixels to determine the extreme points; The Gaussian difference pyramid is expanded at the extreme points using the Taylor formula, and the offset of the extreme points is calculated. Extreme points with offsets less than a preset value are used as feature points. A feature vector is constructed based on the feature points, and the feature vector is matched with a preset feature vector to obtain the grain image. The preset feature vector is calculated based on the preset image. The construction of feature vectors based on the feature points includes: Centered on the feature point, a circular region within a preset radius is selected, and the circular region is located inside the region where the grain is located; Divide the circular region into A circular region, It is a positive integer greater than 1; Divide the circular region into Given several intervals, each with the same angle, we obtain... Sub-regions It is a positive integer greater than 1; Within each sub-region, the amplitude of each pixel is calculated, and the amplitudes of each pixel are weighted to obtain the feature value of the sub-region. The weighting parameter is... ,in Representing feature points, Represents pixels, This represents the weight coefficient for that pixel. The step size is the annular region containing the pixel. The radius of the annular region containing the pixel is... The scale of the pixel; The feature values ​​of each sub-region are combined to obtain the feature vector of the feature point; The step of determining the specific location of the defect based on the grain image and the preset image includes: The grain image is decomposed into an illumination response model, and the illumination response model is then subjected to a logarithmic transformation. The result of the logarithmic transform is then subjected to a Fourier transform and frequency domain filtering. The result of the frequency domain filtering is converted into the spatial domain to obtain the target image; Based on the target image and the preset image, the specific location of the defect is obtained.

Images