A pattern wafer defect detection method based on adaptive local frequency domain filtering

Abstract

The application discloses a kind of based on adaptive local frequency domain filtering graphic wafer defect detection method, by composite second difference removes background component, enhances defect feature from two dimensions.For frequency filtering method based on normalized frequency filter factor.Due to noise component is located in high frequency, only pay attention to low frequency area reflecting defect component, according to the frequency characteristics of different regions, the image is adaptively filtered and adjusted, compared with traditional mean, median filtering and other spatial domain filtering method and low-pass, high-pass filtering and other frequency filtering method, this method effectively suppresses background noise without losing defect details, improves defect contrast and peak signal-to-noise ratio;Using the improved peak-centroid method based on selected abnormal area, the defect signal of different characteristics is more applicable and accurate, and different types of graphic wafer surface defects can be accurately detected and positioned.

Description

Technical Field

[0001] This invention belongs to the field of defect detection technology, specifically relating to a method for detecting defects in patterned wafers based on adaptive local frequency domain filtering. Background Technology

[0002] Semiconductor testing equipment is used throughout the hundreds of processes in semiconductor chip production, serving as an effective means to ensure IC manufacturing quality and yield. Wafer defects include random defects, primarily surface contaminants, and system defects generated by the manufacturing process. The introduction of system defects significantly reduces yield and increases manufacturing costs.

[0003] Current semiconductor defect detection methods primarily include scanning electron microscopy (SEM), transmission electron microscopy (TEM), atomic force microscopy (AFM), and optical inspection. For non-destructive, high-precision, full-area wafer defect detection, optical inspection is the most important method. Optical inspection, based on optical principles, often requires a series of image post-processing steps to accurately locate defects. Defect detection on patterned wafers is even more complex and challenging, and post-processing algorithms play a crucial role in improving the signal-to-noise ratio and contrast of defects.

[0004] In 2021, Chen Shiwei of Zhejiang University used homomorphic filtering to improve the effect of uneven illumination on images, obtained standard comparison images by threshold iteration method based on statistics, and obtained defects by difference (Chen Shiwei. Research on multi-scan patterned wafer inspection technology based on bright and dark field imaging [D]. Master's thesis of Zhejiang University, 2021).

[0005] Chinese patent CN201210585636.8 proposes a Fourier filter with multiple switchable blocking units. By switching different filters, the direction of blocking filtering can be selected to complete the detection of wafer defects on patterns in different directions.

[0006] Chinese patent CN202210988249.2 proposes a method for extracting the period of a periodic structure and a method for detecting wafer defects. The method involves performing peak searching twice on the sampled spectrum to obtain the accurate value of the period of the periodic structure. By obtaining the difference image of the median image of the target image and the adjacent images, defect detection and defect classification are performed on the image to be detected.

[0007] Chinese patent CN202310527789.5 inputs the image to be inspected into a preset defect detection model trained based on the unsupervised algorithm PaDim to obtain the detection result characterizing whether there is a defect in the wafer, and segments the image characterizing the defect region from the camera image through a preset image segmentation strategy.

[0008] Existing post-processing algorithms are mostly simple spatial domain processing (such as Laplacian image enhancement, mean, Gaussian, bilateral filtering, etc.) and common frequency domain processing (low-pass, high-pass filters, etc.), which are not applicable to the detection of complex patterned wafer defects and cannot be completed by a single method. Artificial intelligence algorithms, on the other hand, rely on a large amount of labeled data and parameter tuning, and at the same time place high demands on computing resources. Summary of the Invention

[0009] To overcome the shortcomings of the prior art, this invention proposes a patterned wafer defect detection method based on adaptive local frequency domain filtering, which effectively improves the defect contrast and peak signal-to-noise ratio of the defect image. Combined with the improved peak-centroid method for defect extraction proposed in this invention, different types of patterned wafer surface defects can be accurately detected and located.

[0010] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0011] A method for detecting defects in patterned wafers based on adaptive local frequency domain filtering includes the following steps:

[0012] Step 1: Preprocess the acquired wafer image to obtain image I;

[0013] Step 2: Perform second-order column-direction difference on image I to obtain image I. dx ;

[0014] Step 3: For image I dx Perform second-order difference along the row direction to obtain image I. dxy The region that was not fully subjected to second-order difference of rows and columns was removed to obtain image I. C ;

[0015] Step 4: For image I C Local frequency domain analysis was performed to obtain the normalized frequency filter factor S. i，j ;

[0016] Step 5: Apply the normalized frequency filter factor to perform frequency domain filtering to obtain image I. FC ;

[0017] Step Six: Analyze Image I based on statistics FC An abnormal region is selected, and the abnormal region is determined to have a defect.

[0018] Step 7: Defect localization.

[0019] Compared with traditional spatial and frequency filtering methods, this invention proposes an adaptive local frequency domain filtering defect detection method for patterned wafers. It performs composite second-order difference based on the periodicity of the patterned wafer, focuses on the low-frequency part where the defect is located, and performs frequency domain filtering according to the corresponding frequency filtering factor. This effectively suppresses noise and improves defect contrast without losing defect features.

[0020] Furthermore, based on selected abnormal regions, this invention proposes an improved peak-centroid method to achieve accurate defect localization.

[0021] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0022] Figure 1 This is a flowchart of the patterned wafer defect detection method based on adaptive local frequency domain filtering according to the present invention.

[0023] Figure 2 This is the original image acquired using a dark-field microscope according to an embodiment of the present invention.

[0024] Figure 3 This is an image of the original image rotated according to an embodiment of the present invention.

[0025] Figure 4 This is a region of interest image provided according to an embodiment of the present invention.

[0026] Figure 5 This is an image obtained after second-order difference of columns according to an embodiment of the present invention.

[0027] Figure 6 This is a composite second-order difference image provided according to an embodiment of the present invention.

[0028] Figure 7 This is an image obtained by removing the boundary from a composite second-order difference image according to an embodiment of the present invention.

[0029] Figure 8 This is a normalized frequency filter factor diagram provided according to an embodiment of the present invention.

[0030] Figure 9 This invention provides a frequency domain filtered image based on a normalized frequency filter factor according to an embodiment.

[0031] Figure 10 This is a Gaussian filtered image provided according to an embodiment of the present invention.

[0032] Figure 11 This is a frequency domain low-pass filtered image provided according to an embodiment of the present invention.

[0033] Figure 12 This is an abnormal region determination diagram provided by the present invention according to an embodiment.

[0034] Figure 13 This is an abnormal region map in a filtered image provided according to an embodiment of the present invention.

[0035] Figure 14 This is a defect location map of a filtered image provided according to an embodiment of the present invention.

[0036] Figure 15 This is a defect location map of the original image provided according to an embodiment of the present invention. Detailed Implementation

[0037] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings and examples of defect detection.

[0038] refer to Figure 1 As shown, the present invention provides a patterned wafer defect detection method based on adaptive local frequency domain filtering, which mainly includes the following steps:

[0039] Step 1: Preprocess the acquired wafer image to obtain image I.

[0040] This invention is applicable to wafer images acquired using bright-field and dark-field optical microscopes.

[0041] In this embodiment of the invention, images are acquired using a dark-field microscope, such as... Figure 2 As shown, the test sample is a 55nm linewidth periodic pattern. The periodic sample contains three defects: a 55×300nm horizontal bridge, a 55nm diameter cylinder, and a 55×320nm vertical bridge. The simulated pattern wafer has one full-cycle pattern simulating a cell.

[0042] The preprocessing methods in this step mainly include wafer pose adjustment and region selection. Pose adjustment involves rotating the image by an angle α, where α = atan[(y2-y1) / (x2-x1)], and (x1, y1) and (x2, y2) are any two feature points in the image. Region selection involves selecting the wafer region from the image and excluding regions outside the wafer. The resulting image I has a size of M×N, and a single unit region on the wafer is called a Cell, with each Cell having a size of M. c ×N c Then image I contains A complete unit region.

[0043] for Figure 2The image shown is rotated to the design orientation. Two points (600, 1430) and (1419, 1410) are selected. The rotation angle is calculated as α = atan[(1410-1430) / (1419-600)] = -1.398°. Please refer to the image after rotation. Figure 3 The region of interest (477:1764, 1407:2694) was cropped to obtain image I, see [link / reference]. Figure 4 .

[0044] Step 2: Perform second-order column-direction difference on image I to obtain image I. dx .

[0045] In this step, the size of the cell in the column direction, i.e., M, is used. c To obtain image I, perform second-order difference in the column direction as the spacing. dx The implementation method is as follows:

[0046] I dx (x, y) = I(xM) c ,y)-2I(x,y)+I(x+M c ,y)

[0047] Among them, I dx (x, y) represents the image I dx The gray value at pixel coordinates (x, y), M c Let I(x, y) and I(x+M) be the dimensions of a single cell region on the wafer along the column direction. c ,y),I(xM c (x, y) and (x+M) represent the pixel coordinates (x, y) and (x+M) in image I, respectively. c ,y),(xM c The gray value at (y).

[0048] For the embodiments of the present invention Figure 4 The image shown has a cell size of M. c ×N c 56×28 pixels, M=1288 pixels, N=1288 pixels, M / M c =23, N / N c =46, then image I has 23×46 cells. Second-order differencing is performed with a column dimension of 56 as the interval. At this point, second-order differencing cannot be performed between the left and right columns of cells in the image, resulting in image I. dx Please see Figure 5 .

[0049] Step 3: For image I dx Perform second-order difference along the row direction to obtain image I. dxy .

[0050] In this step, the size of the Cell in the row direction, i.e., N, is used. c For image I, the spacing is... dx Perform second-order difference along the row direction to obtain image I. dxy The implementation method is as follows:

[0051] I dxy (x, y) = I dx (x, yN) c )-2I dx (x, y) + I dx (x, y+N) c )

[0052] Among them, I dxy (x, y) represents the image I dxy The gray value at pixel coordinates (x, y), N c I is the size of the cell in the row direction. dx (x, yN) c ), I dx (x, y+N) c ) represent images I respectively dx Pixel coordinates (x, yN) c (x, y+N) c The grayscale value at ().

[0053] For the embodiments of the present invention Figure 5 The image shown, in image I dx Based on this, a second-order difference is performed with a cell row dimension of 28 as the spacing. At this point, the upper and lower rows of cells in the image cannot be second-order differencing, resulting in image I. dxy Please see Figure 6 .

[0054] This invention removes background components through composite second-order difference. This composite second-order difference, first along the column direction and then along the row direction, amplifies changes in the image, enhancing defect features in both dimensions and making defects in the image more obvious. Furthermore, it can also remove I... dxy The boundary regions in the image where the column composite second-order difference is not completed, specifically for image I. dxy By removing the regions on the wafer where the complete composite second-order difference process could not be performed due to the cells being at the boundaries, image I is obtained. C At this point, image I C It was divided into (M / Mc-2)×(N / N) c -2) regions, Embodiment I of the present invention C There are 21 × 44 complete cycles. Please refer to [link / reference]. Figure 7 .

[0055] In this invention, the order of the difference operations in steps two and three can be interchanged. That is, row-direction second-order difference can be performed first, followed by column-direction second-order difference on the image after row-direction second-order difference. The final result is the same or similar, and the processing effect is not changed. However, it should be noted that the difference operation in step three must be based on the difference image completed in the previous step, and it is a combined row and column second-order difference operation.

[0056] Step 4: For image I C Local frequency domain analysis was performed to obtain the normalized frequency filter factor S. i，j .

[0057] In this step, the image I to be analyzed in the local frequency domain is... C Divided into Each unit area, with This means that the image after combining the second-order differences of each cell is described as dCell. i，j Frequency domain analysis is performed on each cell region, including:

[0058] a. For the cell region dCell in column i and row j. i，j Performing a two-dimensional discrete Fourier transform (DFT) converts the signal from the spatial domain to the frequency domain, i.e.:

[0059]

[0060] Among them, F i，j (u, v) represents the spectrum at the frequency index (u, v), dCell i，j (m, n) represents dCell i，j The value at index (m, n) in the spatial field;

[0061] b. Calculate the amplitude spectrum, amplitude |FFTshift i (u, v)| is obtained by taking the absolute value of the complex number, as follows:

[0062] Mag i，j (u, v) = |F i，j (u, v)|

[0063] Among them, Mag i，j (u, v) represents the amplitude spectrum at the frequency index (u, v), i.e., the intensity at that frequency. Phase information is not considered. It is used to analyze the energy distribution of the spectrum.

[0064] c. Regarding Mag i，j Extract its center The sum of values ​​within the frequency range SMag i，j ,Right now:

[0065]

[0066] Where, round means rounding the number to the nearest integer;

[0067] d. Regarding SMag i，j Normalization is performed to obtain the normalized frequency filter factor S. i，j ,Right now:

[0068]

[0069] Where SMag is SMag i，j The set is defined as follows: min(SMag) and max(SMag) are the minimum and maximum values ​​in the set SMag, respectively.

[0070] In this embodiment of the invention, for Figure 7 The image shown has 1≤i≤21 and 1≤j≤44. Mag is extracted in step c. i，j The total value of all frequency components in the central 14×7 region is SMag i，j :

[0071]

[0072] The final normalization process yields the frequency filter factor S. i，j Represented as:

[0073]

[0074] See normalized frequency filter factor. Figure 8 .

[0075] It is worth noting that the present invention can also perform defect detection on patternless wafers. In this case, there is no need to perform differential analysis. Instead, the process proceeds directly to step four, where a reasonable period is selected for frequency domain analysis and corresponding frequency domain filtering is performed, and then defect extraction is performed on the filtered image.

[0076] Step 5: Apply the normalized frequency filter factor to perform frequency domain filtering to obtain image I. FC .

[0077] In this step, image I C Each cell region dCell i，j Frequency domain filtering is performed by applying the corresponding normalized frequency filter factor in the frequency domain, specifically including:

[0078] a. Regarding dCell i，j Apply the corresponding normalized frequency filter factor S in its frequency domain i，j :

[0079] G i，j (u, v) = F i，j (u, v)·S i，j

[0080] Among them, G i，j (u, v) is the spectrum at the frequency index (u, v) after filtering;

[0081] b. Perform a two-dimensional discrete Fourier inverse transform on G. i，j Transforming (u, v) back into the spatial domain yields the result from dCell. i，j The result after frequency domain filtering: fdCell i，j :

[0082]

[0083] fdCell i，j (m, n) represents the gray value at pixel index (m, n) in the cell region image of the i-th column and j-th row in the spatial domain after frequency domain filtering, determined by fdCell. i，j Forming a filtered image I FC .

[0084] In this embodiment of the invention, fdCell i，j (1≤i≤M / Mc-2,1≤j≤N / N c -2) Forming the filtered image I FC See Figure 9 .

[0085] Furthermore, in order to compare the method proposed in this invention with commonly used spatial and frequency domain filtering methods, Figure 10 and Figure 11 Images after Gaussian filtering (filter size 3×3, standard deviation 2) and frequency domain low-pass filtering (cutoff frequency 300 pixels) are presented respectively. It can be seen that compared with other filtering algorithms, the filtering method based on frequency filtering factor proposed in this invention can effectively eliminate image background and improve defect contrast.

[0086] This invention employs a frequency domain filtering method based on a normalized frequency filtering factor. Since noise components are located at high frequencies, only the low-frequency regions reflecting defect components are considered. Adaptive local filtering adjustments are made to the image based on the frequency domain characteristics of different regions. Compared with traditional spatial domain filtering methods such as mean and median filtering, and frequency domain filtering methods such as low-pass and high-pass filtering, this method effectively suppresses background noise and improves defect contrast without losing defect details.

[0087] Step Six: Analyze Image I based on statistics FC An abnormal region is selected, and the abnormal region is determined to have a defect.

[0088] After obtaining the filtered image I FC Then, first examine the abnormal area. kGiven the selected normalized frequency filter factor array S, calculate the normalized frequency filter factor S. i，j mean

[0089]

[0090] Determine S i，j Is it greater than or equal to? Form logical array L i，j This defines the abnormal region, namely:

[0091]

[0092] The value of 1 in the logical array corresponds to the image I. FC In the process, abnormal regions are formed, which reflect the presence of defective unit regions.

[0093] In this embodiment of the invention, the following is obtained: Then determine S i，j Is it greater than or equal to? That is, 0.5092, forming the logical array L. i，j See Figure 12 There are three abnormal areas here, see [link / reference] Figure 13 .

[0094] Step 7: Defect localization.

[0095] Defect localization in this step is done in image I. FC For each anomalous region in image I, an improved peak-centroid method is used. FC For each abnormal region in the data, perform the following steps to identify and locate the defect:

[0096] a. The kth abnormal region Area k Take the absolute values ​​of the grayscale values ​​of all pixels within the set, sort these absolute values ​​in descending order, and select the pixel coordinates corresponding to the top n values ​​to form a coordinate set TP. k :

[0097] Sort k =sort(|Area) k (:)|,′descend′)

[0098] TP k ={(x, y)||Area k (x, y)|∈Sort k (1:n)}

[0099] Where 'descend' represents the array sorted in descending order, Sort kTP is a sorted array of the absolute values ​​of pixel grayscale values ​​within the k-th abnormal region. k Area is the set of pixel coordinates corresponding to the first n values ​​in the array. k (x, y) represents the gray value at pixel coordinates (x, y) in the k-th abnormal region;

[0100] b. Calculate the coordinates TP in each anomaly region. k The total value I of the corresponding numerical values k :

[0101]

[0102] c. Calculate TP k The weighted average of the elements in the middle, i.e., the defect location in each abnormal region:

[0103]

[0104] in These are the location coordinates of each defect in the x and t directions, respectively, for all Centroids. k This forms the defect coordinate set Centroid, and then, based on the location of the region of interest, the defect coordinates are regressed and marked onto the rotated image.

[0105] This step employs an improved peak-centroid method based on selected abnormal regions, which is more applicable and accurate to defect signals with different characteristics, enabling precise location of defect regions and positions.

[0106] In this embodiment of the invention, n = 10 is chosen, that is, the pixel coordinates corresponding to the top 10 sorted values ​​are selected to form a coordinate set TP. k The weighted average result is rounded to obtain image I. FC The three defect locations are (430, 151), (586, 520), and (742, 994). See [link / reference]. Figure 14 By restoring the coordinates to the rotated image, the defect locations were found to be (962, 1585), (1118, 1954), and (1274, 2428), respectively. These locations are marked on the rotated image. See [link to relevant documentation]. Figure 15 .

[0107] The present invention provides a patterned wafer defect detection method that applies a frequency filtering factor to the defect image for frequency domain filtering, which effectively improves the defect contrast and peak signal-to-noise ratio of the defect image. Combined with the improved peak-centroid method for defect extraction, it can accurately detect and locate different types of patterned wafer surface defects.

[0108] To evaluate the positioning error of the peak-centimeter method for defect location, the distance between the defect and the boundary point is calculated, and the known design value is used as the true value to calculate the relative error.

[0109] The designed distances between the three defects and the boundary points are (40μm, 19.8125μm), (50.3μm, 45μm), and (60.1775μm, 80μm), respectively. The distances calculated based on defect location are (38.7780μm, 19.6650μm), (49.5420μm, 45.1260μm), and (60.3060μm, 77.8320μm), respectively. The location errors are (3.06%, 0.74%), (1.51%, 0.28%), and (0.21%, 2.71%), respectively. The location error is less than 5%, which fully demonstrates the effectiveness of the method of the present invention.

[0110] In summary, the present invention provides a patterned wafer defect detection method with the following characteristics: a. Composite second-order difference is used to remove background components, enhancing defect features from two dimensions; b. Adaptive local frequency domain filtering is used to suppress background noise and improve defect contrast; c. An improved peak-centroid method based on selected abnormal regions is used for defect localization, which is more robust and has greater applicability and accuracy in locating defects of different types and features.

[0111] This detection method is applicable to images acquired by optical inspection systems, such as bright-field and dark-field optical microscopes. Furthermore, it is also suitable for defect detection on patternless wafers. This invention significantly improves defect image contrast and peak signal-to-noise ratio. Combined with an improved peak-centroid method for defect extraction, it effectively detects and locates different types of wafer surface defects. It is an adaptive and robust defect detection method that plays a crucial role in micro / nanoscale wafer defect detection.

[0112] The specific embodiments of the present invention have been described above with reference to the accompanying drawings. However, these descriptions should not be construed as limiting the scope of the present invention. The scope of protection of the present invention is defined by the appended claims. Any modifications based on the claims of the present invention are within the scope of protection of the present invention.

Claims

1. A method for detecting defects in patterned wafers based on adaptive local frequency domain filtering, characterized in that, Includes the following steps: Step 1: Preprocess the acquired wafer image to obtain image I; Step Two, with To obtain the image, perform a second-order difference along the column direction on image I, with the spacing as the second difference. ; Step 3, with For the spacing, for the image Perform second-order difference along the row direction to obtain the image. The image is obtained by truncating regions that have not undergone complete second-order row and column differences. ; Step 4: Process the image Local frequency domain analysis was performed to obtain the normalized frequency filter factor. ; Step 5: Apply the normalized frequency filter factor to perform frequency domain filtering to obtain the image. ; Step Six: Analyze the image based on statistics An abnormal region is selected, and the abnormal region is determined to have a defect. Step 7: Defect localization; In step six, the normalized frequency filter factor is calculated. mean : in The size of image I, , Row and column numbers; judge Is it greater than or equal to? To form a logical array This defines the abnormal region, namely: The value of 1 in the logical array corresponds to the image In the process, abnormal regions are formed, which reflect the presence of defective unit regions.

2. The patterned wafer defect detection method based on adaptive local frequency domain filtering according to claim 1, characterized in that, In step one, the preprocessing includes wafer pose adjustment and region selection; the pose adjustment involves adjusting the image according to the rotation angle. α Rotate, where α =atan[( y 2 y 1) / ( x 2 x 1)],( x 1, y 1) and ( x 2, y 2) Any two feature points in the image; the region selection is to select the wafer region in the image and exclude the region outside the wafer.

3. The method for detecting patterned wafer defects based on adaptive local frequency domain filtering according to claim 1, characterized in that, In step two, image I is subjected to a second-order column-direction difference to obtain image I. The implementation method is as follows: in, Representing an image Mid-pixel coordinates grayscale value at that location The dimension of a cell, or unit region, on a wafer is defined in the column direction. These represent the pixel coordinates in image I. , , The grayscale value at that location.

4. The patterned wafer defect detection method based on adaptive local frequency domain filtering according to claim 3, characterized in that, In step three, the image Perform second-order difference along the row direction to obtain the image. The implementation method is as follows: in, Representing an image Mid-pixel coordinates grayscale value at that location This represents the size of the cell in the row direction. , Representing images respectively Mid-pixel coordinates , The grayscale value at that location.

5. The patterned wafer defect detection method based on adaptive local frequency domain filtering according to claim 4, characterized in that, In step four, the image to be analyzed in the local frequency domain is... Divided into Each cell region, in dCell i,j (1≤ i ≤ -2, 1≤ j ≤ 2) Representation; frequency domain analysis is performed on each unit region, including: a. Regarding the first List row cell area Performing a two-dimensional discrete Fourier transform converts the signal in the spatial domain into the frequency domain, i.e.: in, Frequency index The spectrum at that location, express Medium spatial domain index The value at; b. Calculate the amplitude spectrum as follows: in, For frequency index The amplitude spectrum at that frequency, i.e. the intensity at that frequency, without considering phase information; c. To Extract its center round( )×round( The sum of values ​​within the frequency range ,Right now: in, This means rounding the number to the nearest integer. d. Regarding Normalization is performed to obtain the normalized frequency filter factor. ,Right now: in for The set, min(SMag), Sets The minimum and maximum values ​​in the range.

6. The patterned wafer defect detection method based on adaptive local frequency domain filtering according to claim 5, characterized in that, In step five, the image Each unit area Frequency domain filtering is performed by applying the corresponding normalized frequency filter factor in the frequency domain, specifically including: a. To Apply the corresponding normalized frequency filter factor in its frequency domain : in, Frequency index after filtering The spectrum at that location; b. Perform a two-dimensional discrete Fourier inverse transform, and... Transforming back to the spatial domain, we obtain the result from... Result after frequency domain filtering : in For the frequency domain filtered, the first in the spatial domain i List j Pixel index in the cell region image of a row The gray value at that location is determined by Forming a filtered image .

7. The patterned wafer defect detection method based on adaptive local frequency domain filtering according to claim 1, characterized in that, In step seven, defect localization is achieved using an improved peak-centroid method in the image. Perform the following steps for each abnormal region in the data: a. The first k Areas with abnormalities k Take the absolute values ​​of the grayscale values ​​of all pixels within the array, sort these absolute values ​​in descending order, and select the pixels before sorting. The pixel coordinates corresponding to the numerical values ​​form a coordinate set. : in This indicates that the array is sorted in descending order. For the first k A sorted array of absolute values ​​of pixel grayscale values ​​within an abnormal region. For the first part of the array A set of pixel coordinates corresponding to each value; Indicates the first k Pixel coordinates in the abnormal region The grayscale value at that location; b. Calculate the coordinates of each anomaly region. The total value of the corresponding numerical values : c. Calculation The weighted average of the elements in the middle, i.e., the defect location in each abnormal region: in , For each defect in x , y Directional positioning coordinates, all That is, to form a set of defect coordinates Then, based on the location of the region of interest, the defect coordinates are regressed and marked onto the rotated image.

8. The patterned wafer defect detection method based on adaptive local frequency domain filtering according to claim 1, characterized in that, The difference order in steps two and three is reversed, that is: first perform second-order difference in the row direction, and then perform second-order difference in the column direction on the image after second-order difference in the row direction.

9. The patterned wafer defect detection method based on adaptive local frequency domain filtering according to claim 1, characterized in that, The wafer image is an image acquired by a bright-field or dark-field optical microscope; or, defect detection is performed on a wafer without a pattern. In this case, the process proceeds directly to step four, where a period is selected for frequency domain analysis and corresponding frequency domain filtering is performed, and then defect extraction is performed on the filtered image.

Images