A method for defect detection of a reticle

Abstract

The application discloses a kind of defect detection methods for reticle mask, it is related to mask defect detection technical field, according to instruction magnification and mask name automatic call and solidification detection parameter;With light calibration point gray quadratic polynomial fitting normalization correction reference map;After dividing detection area, extract ORB feature, get subpixel homonyme point pair by hamming distance and RANSAC;Difference detection defect is sought after affine matrix fine alignment Die image.The application can replace artificial rejudgment, and detection time is reduced to second level, defect identification accuracy is ≥99%, satisfy 28 nm node high-efficiency accurate demand.

Description

Technical Field

[0001] This invention relates to the field of mask defect detection technology, and more particularly to a defect detection method for reticle masks. Background Technology

[0002] In semiconductor photolithography, the pattern integrity of the reticle directly determines the wafer yield. The industry commonly uses automated mask inspection systems (such as the KLA-Tencor series) that integrate high-resolution optical imaging, image processing, and defect classification systems to perform 100% mask scanning. These devices typically offer two algorithm modes:

[0003] Die-to-Database (D2DB): Compares the chip cell under test (Die) with the layout database pixel by pixel;

[0004] Die-to-Die (D2D): Performs a layout-free comparison between adjacent dies, suitable for scenarios with repetitive graphics.

[0005] While existing technologies have achieved high-throughput image acquisition, the "secondary review" and "fine classification" of defects still heavily rely on manual labor.

[0006] The detection system only outputs the coordinates of suspected defects and grayscale difference images. Operators need to confirm the authenticity of each defect under a high-powered microscope and manually mark the type (opaque spot, pindot, pin hole, black convex / white convex, open / short circuit, thinline, particle, contact hole abnormality, etc.).

[0007] Human judgment is susceptible to subjective experience, visual fatigue, and differences in ambient lighting, resulting in a misjudgment rate of >15%, and the re-judgment time for a single mask is >30 min, which cannot meet the dual requirements of "high precision + high throughput" for advanced nodes of 28 nm and below.

[0008] In recent years, deep learning-based defect classification algorithms have been introduced, but existing solutions still remain at the single-point application of "offline training - online inference":

[0009] Training data requires extensive manual annotation, which is costly.

[0010] The reasoning model is disconnected from the D2D / D2DB comparison algorithm, which fails to utilize the difference features generated during the comparison stage, resulting in information redundancy.

[0011] It is sensitive to subpixel-level positioning errors and lighting differences, and is prone to producing false defects. Summary of the Invention

[0012] The technical problem to be solved by this invention is to address the shortcomings of existing technologies, specifically by providing a defect detection method for reticle masks, as detailed below:

[0013] 1) In a first aspect, the present invention provides a defect detection method for reticle masks, the specific technical solution of which is as follows:

[0014] Based on the detection magnification information and the name of the mask template to be tested in the detection command, the detection parameters of the equipment used to perform defect detection on the mask template to be tested are adjusted, and the detection parameters are associated and stored with the name of the mask template to be tested;

[0015] Based on the gray values ​​of the light calibration points in the reference image corresponding to the mask to be tested, the light source brightness of the reference image is corrected, and the corrected reference image is generated.

[0016] The detection region is divided into the corrected reference image. The ORB feature points corresponding to the test Die image and the reference ORB feature points corresponding to the reference Die image are extracted from the division results. Hamming distance matching and RANSAC are used to remove mismatches and establish sub-pixel level corresponding point pairs.

[0017] By combining subpixel-level corresponding point pairs, a transformation matrix is ​​determined to convert the die image under test into a reference die image. Based on the transformation matrix, the die image under test and the reference die image are aligned, and defect detection is performed based on the alignment result.

[0018] The beneficial effects of the defect detection method for reticle masks provided by this invention are as follows:

[0019] By associating and storing detection parameters with mask names, one-click switching of detection configurations for different process nodes is achieved, improving equipment versatility and line changeover efficiency. Light source brightness correction is performed using the grayscale values ​​of optical calibration points, effectively eliminating grayscale deviations caused by uneven illumination and improving subsequent comparison accuracy. ORB+RANSAC is used to construct sub-pixel-level corresponding point pairs and generate a transformation matrix, achieving high-precision geometric alignment between the test die and the reference die, significantly reducing the false defect rate caused by positional errors. The entire process of parameter adjustment, brightness correction, region division, feature matching, and defect detection is automated, completely replacing manual review and reducing detection time from minutes to seconds. Defect identification accuracy is increased to over 99%, meeting the accuracy and capacity requirements for mask defect detection at advanced nodes of 28 nm and below.

[0020] Based on the above solution, the present invention can be further improved as follows.

[0021] Furthermore, the process of determining the light source brightness correction for the reference image is as follows:

[0022] Multiple light calibration points are randomly selected at the boundary between the transparent and non-transparent areas in the reference image. The gray values ​​corresponding to the multiple light calibration points are obtained. A light distribution model is constructed by fitting a quadratic polynomial. The reference image is then normalized using the light distribution model to complete the light source brightness correction.

[0023] The beneficial effects of the above-mentioned further solutions are as follows:

[0024] By randomly sampling multiple light calibration points at the boundary between transparent and non-transparent areas and constructing an illumination distribution model using quadratic polynomial fitting, pixel-level normalization of the entire reference image is achieved. This effectively eliminates gray-level gradient errors caused by uneven light sources in different areas of the mask, thereby reducing illumination difference noise in the subsequent image comparison stage by more than two orders of magnitude. This significantly reduces the number of false defects and improves the gray-level contrast of real defects, thus simultaneously improving the reliability and detection rate of sub-pixel level defect detection.

[0025] Furthermore, the specific process for establishing sub-pixel level corresponding point pairs is as follows:

[0026] The feature point pairs consisting of the ORB feature points to be tested and the reference ORB feature points are coarsely matched by Hamming distance matching, and the feature point pairs that meet the preset requirements are finely matched by RANSAC to remove mismatches, so as to obtain sub-pixel level corresponding point pairs.

[0027] The beneficial effects of the above-mentioned further solutions are as follows:

[0028] After coarse matching is performed using Hamming distance, RANSAC iterative elimination of mismatches can be achieved in one go, obtaining high-precision, uniformly distributed sub-pixel level pairs of the same name. This reduces the error of solving the subsequent affine matrix to within 0.05 pixels, thereby reducing the false defect rate caused by feature point drift by more than 90%, while ensuring that real defects can still be detected stably at the most stringent 28 nm node.

[0029] Furthermore, the specific process for determining defects based on the alignment results is as follows:

[0030] The aligned die image to be tested is compared with the reference die image, and the suspected defect area is extracted by binarization segmentation.

[0031] After preprocessing the suspected defective areas, feature analysis is performed using a preset analysis algorithm to determine the final defect result.

[0032] The beneficial effects of the above-mentioned further solutions are as follows:

[0033] By rapidly locating suspected defect areas through differential and binarized segmentation, and then removing background noise through morphological preprocessing, subsequent feature analysis only needs to be performed within 2%–5% of the pixels, reducing the overall detection time by 70%. At the same time, by using preset algorithms to control multi-dimensional features such as area, width, height, and grayscale mean, the false defect rate is reduced to below 0.3%, ensuring that 100% of real defects with ≥3 pixels on the 28 nm node mask are detected and automatically classified.

[0034] Furthermore, it also includes:

[0035] Based on the feature representation of the mask template under test, the offset between the mask template under test and the defect detection equipment is calculated, and the position of the mask template under test is corrected based on the offset.

[0036] The beneficial effects of the above-mentioned further solutions are as follows:

[0037] By calculating and compensating for translational / rotational offsets between the device and the mask in real time using mask feature representations, the initial error of subsequent ORB matching and affine alignment is reduced by an order of magnitude, thereby improving the overall positioning accuracy to the sub-pixel level. This significantly reduces the number of false defects caused by platform drift and directly shortens the entire inspection time by about 20%.

[0038] 2) In a second aspect, the present invention also provides a defect detection system for reticle masks, the specific technical solution of which is as follows:

[0039] The initialization module is used to: adjust the detection parameters of the device used to perform defect detection on the mask template under test based on the detection magnification information and the name of the mask template under test in the detection instruction, and associate and store the detection parameters with the name of the mask template under test;

[0040] The light source brightness correction module is used to: correct the light source brightness of the reference image based on the gray value of the light calibration point in the reference image corresponding to the mask to be tested, and generate the corrected reference image;

[0041] The detection region determination module is used to: divide the detection region into the corrected reference image, extract the ORB feature points corresponding to the test Die image and the reference ORB feature points corresponding to the reference Die image in the division result, use Hamming distance matching and RANSAC to remove mismatches, and establish sub-pixel level corresponding point pairs.

[0042] The detection module is used to: combine subpixel-level corresponding point pairs to determine the transformation matrix that converts the die image to be tested into a reference die image; based on the transformation matrix, align the die image to be tested with the reference die image; and perform defect detection based on the alignment result.

[0043] Based on the above solution, the present invention can be further improved as follows.

[0044] Furthermore, the process of determining the light source brightness correction for the reference image is as follows:

[0045] Multiple light calibration points are randomly selected at the boundary between the transparent and non-transparent areas in the reference image. The gray values ​​corresponding to the multiple light calibration points are obtained. A light distribution model is constructed by fitting a quadratic polynomial. The reference image is then normalized using the light distribution model to complete the light source brightness correction.

[0046] Furthermore, the specific process for establishing sub-pixel level corresponding point pairs is as follows:

[0047] The feature point pairs consisting of the ORB feature points to be tested and the reference ORB feature points are coarsely matched by Hamming distance matching, and the feature point pairs that meet the preset requirements are finely matched by RANSAC to remove mismatches, so as to obtain sub-pixel level corresponding point pairs.

[0048] Furthermore, the specific process for determining defects based on the alignment results is as follows:

[0049] The aligned die image to be tested is compared with the reference die image, and the suspected defect area is extracted by binarization segmentation.

[0050] After preprocessing the suspected defective areas, feature analysis is performed using a preset analysis algorithm to determine the final defect result.

[0051] Furthermore, it also includes:

[0052] The correction module is used to: calculate the offset between the mask under test and the defect detection device based on the feature identifier of the mask under test, and perform position correction on the mask under test based on the offset.

[0053] 3) In a third aspect, the present invention also provides an electronic device, the electronic device including a processor coupled to a memory, the memory storing at least one computer program, the at least one computer program being loaded and executed by the processor to enable the electronic device to perform any of the above methods.

[0054] 4) In a fourth aspect, the present invention also provides a computer-readable storage medium storing at least one computer program, which is loaded and executed by a processor to enable a computer to implement any of the above methods.

[0055] It should be noted that the beneficial effects of the technical solutions of the second to fourth aspects of the present invention and their corresponding possible implementations can be found in the above description of the technical effects of the first aspect and its corresponding possible implementations, and will not be repeated here. Attached Figure Description

[0056] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0057] Figure 1 This is one of the flowcharts illustrating a defect detection method for a reticle mask according to an embodiment of the present invention;

[0058] Figure 2 This is a second schematic flowchart of a defect detection method for a reticle mask according to an embodiment of the present invention;

[0059] Figure 3 This is a schematic diagram of the light source brightness correction process in a defect detection method for a reticle mask according to an embodiment of the present invention;

[0060] Figure 4 This is a schematic diagram of the detection area of ​​a defect detection method for a reticle mask according to an embodiment of the present invention;

[0061] Figure 5 This is a schematic diagram of the defect detection results of a defect detection method for a reticle mask according to an embodiment of the present invention;

[0062] Figure 6 This is a structural framework diagram of an electronic device according to the present invention. Detailed Implementation

[0063] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0064] like Figure 1 As shown, an embodiment of the present invention provides a defect detection method for a reticle mask, comprising the following steps:

[0065] S1, based on the detection magnification information and the name of the mask template to be tested in the detection command, adjust the detection parameters of the equipment used to perform defect detection on the mask template to be tested, and store the detection parameters in association with the name of the mask template to be tested;

[0066] S2, Based on the gray values ​​of the light calibration points in the reference image corresponding to the mask to be tested, perform light source brightness correction on the reference image and generate the corrected reference image;

[0067] S3, perform detection region division on the corrected reference image, extract the ORB feature points corresponding to the test Die image and the reference ORB feature points corresponding to the reference Die image from the division results, use Hamming distance matching and RANSAC to remove mismatches, and establish sub-pixel level corresponding point pairs.

[0068] S4. Combining subpixel-level corresponding point pairs, determine the transformation matrix to convert the test die image into a reference die image. Based on the transformation matrix, align the test die image with the reference die image and perform defect detection based on the alignment result.

[0069] The beneficial effects of the defect detection method for reticle masks provided by this invention are as follows:

[0070] By associating and storing detection parameters with mask names, one-click switching of detection configurations for different process nodes is achieved, improving equipment versatility and line changeover efficiency. Light source brightness correction is performed using the grayscale values ​​of optical calibration points, effectively eliminating grayscale deviations caused by uneven illumination and improving subsequent comparison accuracy. ORB+RANSAC is used to construct sub-pixel-level corresponding point pairs and generate a transformation matrix, achieving high-precision geometric alignment between the test die and the reference die, significantly reducing the false defect rate caused by positional errors. The entire process of parameter adjustment, brightness correction, region division, feature matching, and defect detection is automated, completely replacing manual review and reducing detection time from minutes to seconds. Defect identification accuracy is increased to over 99%, meeting the accuracy and capacity requirements for mask defect detection at advanced nodes of 28 nm and below.

[0071] It should be noted that the mask under test refers to the high-precision quartz or glass template used in the photolithography process to transfer chip patterns onto a silicon wafer. In this solution, it specifically refers to the mask currently being prepared for defect inspection, which is covered with repeatedly arranged chip dies. The purpose of the inspection is to identify any pattern defects or contamination defects in these dies.

[0072] In another embodiment of this solution, the process of adjusting the detection parameters of the device includes:

[0073] Step 1: Issuance of Detection Command

[0074] The software layer sends the detection magnification (e.g., ×150, ×200) and the name of the mask to be tested (e.g., “N28-LV-001”) as keywords to the machine control unit.

[0075] Step 2: Recipe library matching

[0076] The machine searches for an entry with the same name in the local Recipe database based on the mask template name; if it does not exist, it selects the default template based on the detection magnification.

[0077] The Recipe entry contains all the parameters required for the corresponding scaling factor:

[0078] Optical magnification and numerical aperture (NA); light source power and exposure time; image acquisition resolution (pixels / μm); defect size conversion factor; algorithm parameters such as grayscale threshold, control index, and ORB feature extraction threshold.

[0079] Step 3: Closed-loop calibration of machine hardware

[0080] Magnification switching: The drive motor adjusts the objective lens group so that the actual magnification differs from the Recipe setting by less than 0.1%.

[0081] Light source power adjustment: The LED current is adjusted in a closed loop through real-time feedback from the optical power meter to ensure that the gray value at the calibration point falls within the dynamic range of 45%–55% specified in the Recipe.

[0082] Pixel size calibration: Using a standard grid, the pixel size is calibrated at the current magnification. The actual size relationship is calibrated in real time, and the defect size conversion factor is updated.

[0083] Platform movement resolution reset: Recalculate the stepper motor pulse equivalent based on the magnification to ensure that the scanning step distance matches the Recipe.

[0084] Step 4: Parameter solidification and associated storage

[0085] All the above real-time calibration results are written back to the Recipe entry and stored in the machine log along with the mask template name and detection timestamp, enabling one-click reproduction of "name-parameter".

[0086] In another embodiment of this solution, the light source brightness correction is preceded by a correction process for the mask to be tested, specifically:

[0087] After successful initialization, the algorithm immediately loads the detection magnification, product name, and ROI parameters issued by the software into memory. It reads the baseline template image corresponding to the product name from the template library and extracts ORB feature points from its four corners and center. At the same time, it transfers the newly acquired original test image into the memory buffer. The algorithm performs ORB descriptor matching on the two images. First, it uses Hamming distance for brute-force matching to obtain coarse matching point pairs. Then, it uses RANSAC iteratively to eliminate mismatches, retaining ≥8 pairs of corresponding points with the highest confidence. Based on these corresponding points, it calculates the translation vector (dx, dy) and rotation angle θ of the test image relative to the baseline template, generating a 2×3 affine matrix. The control system drives the high-precision XYθ platform for real-time compensation based on this matrix. After the platform moves into position, it acquires the image again and repeats the above matching-calculation-compensation closed loop until (dx, dy) < 0.05 pixels and θ < 0.002 rad, confirming that the product has completed sub-pixel-level precise positioning.

[0088] In another embodiment of this solution, the reference image corresponding to the mask under test refers to the image of the mask under test identified by the device. Based on the above, the process of light source brightness correction is as follows:

[0089] Read the entire mask image that has been corrected and whose coordinates are aligned. First, traverse all horizontal transition edges between the transparent area (white high-grayscale area) and the non-transparent area (black low-grayscale area) of the image, and randomly select no less than 16 edge points as candidate light calibration points, recording their (x,y) coordinates and corresponding grayscale g. Perform quadratic polynomial fitting on these (g,x,y) triples to obtain the two-dimensional illumination distribution model I(x,y)=ax²+by²+cxy+dx+ey+f, where the coefficients are solved by the least squares method. Use this model to calculate the theoretical grayscale value of each pixel in the entire image, and compare it with the measured grayscale value to obtain the normalized coefficient matrix. Finally, multiply the original image pixel by pixel by this coefficient matrix to complete the brightness normalization, output the corrected reference image, and attach an illumination distribution map of the same size for subsequent threshold segmentation. The light source brightness correction is now complete.

[0090] In another embodiment of this solution, the process of dividing the corrected reference image into regions is specifically as follows:

[0091] The normalized reference image is read into memory. First, the geometric coordinates of the pattern layer, cutting path, alignment mark, and blank area are parsed using a preset mask layout rule file. The image is then binarized into patterned and non-patterned areas through threshold segmentation. Then, morphological opening and closing operations are used to remove isolated noise and close small gaps. The entire image is divided into three types of regions according to the rules: the pattern defect detection area is formed by expanding the boundaries of all effective chip patterns by 2μm; the contamination defect detection area covers the cutting path and blank area and expands by 1μm; the shielding area is directly taken from the alignment mark and known non-functional areas. Then, a corresponding binary mask is generated for each region, and the mask and the corrected image are sent to the next step of ORB feature extraction.

[0092] In another embodiment of this scheme, the process of establishing sub-pixel level corresponding point pairs is as follows:

[0093] ORB feature processing is performed simultaneously in the test die image and the corresponding reference die image: First, candidate key points are extracted in the mask intersection of the two types of regions using FAST corner detection. For each point, the Harris response is calculated and the top N strongest points are retained. The gray-scale centroid method is used to assign the main direction and generate a 256-bit rBRIEF descriptor. After completion, the descriptor sets of the two images are sent to Hamming distance brute-force matching. The nearest neighbor and second nearest neighbor distance ratio <0.8 are retained as coarse matching. RANSAC is iterated 1000 times, and 4 pairs of points are randomly selected each time to estimate the homography matrix. Point pairs with reprojection error <1 pixel are judged as interior points. Finally, the largest interior point set is retained as sub-pixel level homonymous point pairs.

[0094] In another embodiment of this scheme, the process of determining the transformation matrix includes:

[0095] Extract all subpixel-level corresponding point pairs. First, solve the 2×3 affine matrix A=[a11 a12 tx;a21 a22 ty] using least squares. Then, use bilinear interpolation to map the image to be tested pixel by pixel to the reference Die coordinate system according to A. During interpolation, for each target integer coordinate (x′,y′), reverse the calculation to obtain the floating-point coordinate (x,y) in the image to be tested and take the 2×2 neighborhood gray-level weighted average. After mapping, crop or pad with zeros according to the size of the reference image to ensure that the width and height of the two images are completely consistent. Finally, calculate the mutual information or mean square error of the entire overlapping area as the alignment quality index. If the index is better than the set threshold, output the aligned image to be tested. Otherwise, backtrack to the feature matching step to re-select point pairs and recalculate the matrix until the accuracy requirements are met.

[0096] In another embodiment of this solution, the process of aligning the die image to be tested with the reference die image includes:

[0097] The grayscale difference image is obtained by calculating the absolute difference pixel by pixel between the aligned test die image I′test and the reference die image Iref. Based on the pattern area and the contaminated area, the corresponding Otsu adaptive threshold T is called to differentiate the data. Figure 2 The initial defect mask B is obtained by quantization. An opening operation (erosion followed by dilation) is performed on B using a 3×3 structuring element to remove isolated noise points with an area <3 pixels. A closing operation is then used to close minor breaks. Afterward, 8-connected component labeling is performed on the cleaned mask. For each connected component, the area A, the bounding rectangle width and height (w, h), the average grayscale difference μ_D, the aspect ratio r, and the roundness c are extracted. Finally, these features are input into the multi-dimensional control thresholds for each defect type in the Recipe for judgment. If a connected component satisfies any real defect condition, its centroid coordinates, defect type, size, and corresponding thumbnail are recorded, generating a defect list Dlist, and the process directly proceeds to the result merging and output step.

[0098] In another embodiment of this solution, the process of defect detection based on the alignment processing result includes:

[0099] The current die's defect list (Dlist) is transformed to the global coordinates of the entire Reticle. The Dlists of all detected dies on the same Reticle are traversed. Defects in overlapping areas are merged and deduplicated based on IoU > 0.5. Defects are then sorted by type, area, and confidence level. Global defect information is packaged into a unified data structure: defect index, global coordinates (x, y), type label, size, grayscale difference, thumbnail path, and corresponding die number. An XML / CSV template is used to generate a machine-readable report, which is then synchronously uploaded to the MES database. Simultaneously, a PDF report containing a defect distribution map and a ZIP package containing small defect images are output according to the customer's format. Finally, a "detection complete" signal is returned to the software layer, and the memory cache is released, awaiting the next mask template task.

[0100] Furthermore, the process of determining the light source brightness correction for the reference image is as follows:

[0101] Multiple light calibration points are randomly selected at the boundary between the transparent and non-transparent areas in the reference image. The gray values ​​corresponding to the multiple light calibration points are obtained. A light distribution model is constructed by fitting a quadratic polynomial. The reference image is then normalized using the light distribution model to complete the light source brightness correction.

[0102] Furthermore, the specific process for establishing sub-pixel level corresponding point pairs is as follows:

[0103] The feature point pairs consisting of the ORB feature points to be tested and the reference ORB feature points are coarsely matched by Hamming distance matching, and the feature point pairs that meet the preset requirements are finely matched by RANSAC to remove mismatches, so as to obtain sub-pixel level corresponding point pairs.

[0104] Furthermore, the specific process for determining defects based on the alignment results is as follows:

[0105] The aligned die image to be tested is compared with the reference die image, and the suspected defect area is extracted by binarization segmentation.

[0106] After preprocessing the suspected defective areas, feature analysis is performed using a preset analysis algorithm to determine the final defect result.

[0107] Furthermore, it also includes:

[0108] Based on the feature representation of the mask template under test, the offset between the mask template under test and the defect detection equipment is calculated, and the position of the mask template under test is corrected based on the offset.

[0109] Example 1, such as Figure 2 As shown, the algorithm receives the detection task parameters from the software and completes the initialization configuration of the algorithm module. Based on the detection recipe name passed in by the software, it loads the corresponding template file; the detection magnification, representing image resolution information, is used for subsequent defect size calculation; the ROI region configuration parameters of the image to be tested are used to extract useful detection information from the image. The algorithm outputs a signal indicating whether initialization was successful, loads relevant image data into memory, and configures algorithm running parameters, etc.

[0110] The acquired original image is rotated and translated to match the reference image. The algorithm achieves this by extracting features from the localization markers; it primarily uses the ORB feature extraction algorithm to automatically extract the localization markers. ORB (Oriented Fast and Rotated BRIEF) is a computer vision algorithm for feature point detection and description, designed to provide a fast and efficient alternative to the computationally expensive SIFT and SURF methods. ORB determines the principal orientation of each feature point by calculating the gray-level centroid within its neighborhood; this method is called the gray-level centroid method. The gray-level centroid is a point obtained by calculating the weighted average position of the pixels surrounding the feature point, with the weights being its gray-level values. BRIEF (Binary Robust Independent Elementary Features) is a binary string-based descriptor that constructs a descriptor by comparing pairs of pixels (e.g., recording 1 if the intensity of the first pixel is greater than the second, otherwise 0). When generating the BRIEF descriptor, ORB does not directly use the original pixel pairs, but rotates and adjusts these pixel pairs according to the calculated principal direction, thus ensuring that feature points can be correctly matched even when the image is rotated.

[0111] like Figure 3As shown, the image brightness difference caused by uneven light source is corrected to improve the comparison accuracy of subsequent images. Multiple "grayscale transition points" are selected as calibration points at the boundary between transparent and opaque areas. The gray values ​​at these locations are sampled as illumination intensity samples, and a quadratic polynomial fitting is used to construct an illumination distribution model. The entire image is then normalized.

[0112] like Figure 4 As shown, the detection area includes pattern structure areas (such as lines and circuit areas), contamination defect detection areas, and non-functional areas (shielded areas); the types of areas to be detected are divided, and different detection parameters are set according to different areas, excluding shielded areas to avoid false alarms.

[0113] This invention achieves feature matching by extracting ORB features, which extracts feature points from the test die and the reference die and matches them to establish a correspondence.

[0114] The image of the die to be tested is transformed into the coordinate space of the reference die, ensuring a one-to-one correspondence between pixels. Based on feature points, the correspondence between the images of the die to be tested and the reference die is found, and the affine transformation matrix is ​​calculated according to the correspondence. Based on bilinear interpolation, the image of the die to be tested is transformed into the coordinate system of the reference die through the affine transformation matrix. To ensure that the image sizes of the die to be tested and the reference die are consistent, the image size is usually adjusted after the affine transformation to maintain consistency between the image sizes of the die to be tested and the reference die.

[0115] Affine transformation is a geometric transformation widely used in image processing and computer vision. It can perform operations such as translation, rotation, scaling, and cropping of images. The affine transformation matrix is ​​a 2×3 matrix, with the following form:

[0116] T=[ ],

[0117] For a two-dimensional point (x, y), after an affine transformation, a new point (x', y') is obtained. The transformation formula is:

[0118] [ ]=[ ]* [ ]

[0119] Where: tx and ty represent translation amounts, and a11, a12, a21, and a22 represent rotation, scaling, and shearing.

[0120] In image alignment tasks, there are typically at least 3 pairs of non-collinear corresponding points (source_points). The target_points are estimated using the least squares algorithm to estimate the affine transformation matrix.

[0121] The aligned test die and reference die images are subtracted to extract abnormal regions and determine whether the extracted suspected defects are real defects. First, based on the formula diff = abs(img1 - img2), the aligned test die and reference die are subtracted, and the suspected defect regions are extracted by binarization segmentation. Then, morphological operations are performed to remove background noise; feature analysis (area, width, height, grayscale mean, etc.) is performed; finally, control indicators are used to screen and obtain the final real defects.

[0122] For each pair of corresponding pixels (x, y), calculate the grayscale difference between the test die and the reference die: D(x,y) = |Itest(x,y) - Imodel(x,y)|, where D(x,y) represents the pixel value of the difference image at position (x, y); Itest(x,y) and Imodel(x,y) represent the pixel values ​​of the test image and the template image at position (x, y), respectively. Apply a threshold T to the difference image to convert the difference result into a binary image.

[0123] B(x,y)=

[0124] This highlights areas of difference and removes minor differences caused by noise or other factors.

[0125] Connectivity analysis is performed on the final binary image to extract and label each individual defect region. For each connected region, its area, size, and other features can be calculated for further defect analysis.

[0126] Detection results merging and output: Summarize the detection results of all dies and output defect information (coordinates, defect type, size information, defect thumbnail, etc.).

[0127] It should be noted that, as Figure 5 As shown, various defects include: opaque spots, pindots, pinholes, black bumps, white bumps, open circuits, short circuits, thin lines, particles, contact holes, etc.

[0128] Description of various defects:

[0129] Opaque Spot: White circular spots appear in areas that were originally black (the light-transmitting areas are obscured).

[0130] Pinhole: Black circular spots appear in the original white area (light shines through the shaded area);

[0131] White Bump: Localized areas are darker or brighter (by definition);

[0132] Black Bump may appear brighter or whiter in certain areas (depending on how it is used to indicate light transmission);

[0133] Broken Line: A portion of a line segment is missing.

[0134] Short: An additional connecting area is added between the two lines;

[0135] Thin Line: A line that is thinner than the others.

[0136] Particle occlusion point.

[0137] In the above embodiments, although the steps are numbered S1, S2, etc., they are only specific embodiments given by the present invention. Those skilled in the art can adjust the execution order of S1, S2, etc. according to the actual situation, which is also within the protection scope of the present invention. It can be understood that in some embodiments, some or all of the above embodiments may be included.

[0138] This invention also provides a defect detection system for reticle masks, the specific technical solution of which is as follows:

[0139] The initialization module is used to: adjust the detection parameters of the device used to perform defect detection on the mask template under test based on the detection magnification information and the name of the mask template under test in the detection instruction, and associate and store the detection parameters with the name of the mask template under test;

[0140] The light source brightness correction module is used to: correct the light source brightness of the reference image based on the gray value of the light calibration point in the reference image corresponding to the mask to be tested, and generate the corrected reference image;

[0141] The detection region determination module is used to: divide the detection region into the corrected reference image, extract the ORB feature points corresponding to the test Die image and the reference ORB feature points corresponding to the reference Die image in the division result, use Hamming distance matching and RANSAC to remove mismatches, and establish sub-pixel level corresponding point pairs.

[0142] The detection module is used to: combine subpixel-level corresponding point pairs to determine the transformation matrix that converts the die image to be tested into a reference die image; based on the transformation matrix, align the die image to be tested with the reference die image; and perform defect detection based on the alignment result.

[0143] It should be noted that the beneficial effects of the defect detection system for reticle masks provided in the above embodiments are the same as those of the defect detection method for reticle masks described above, and will not be repeated here. Furthermore, the system provided in the above embodiments is only illustrated by the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the system can be divided into different functional modules according to the actual situation to complete all or part of the functions described above. In addition, the system and method embodiments provided in the above embodiments belong to the same concept, and their specific implementation process is detailed in the method embodiments, and will not be repeated here.

[0144] like Figure 6 As shown, an electronic device 300 according to an embodiment of the present invention includes a processor 320 coupled to a memory 310. The memory 310 stores at least one computer program 330, which is loaded and executed by the processor 320 to enable the electronic device 300 to implement any of the above-mentioned methods. Specifically:

[0145] The electronic device 300 can vary considerably due to differences in configuration or performance. It may include one or more processors 320 (Central Processing Units, CPUs) and one or more memories 310. The one or more memories 310 store at least one computer program 330, which is loaded and executed by the one or more processors 320 to enable the electronic device 300 to implement the defect detection method for reticle masks provided in the above embodiments. Of course, the electronic device 300 may also have wired or wireless network interfaces, a keyboard, and input / output interfaces for input and output. The electronic device 300 may also include other components for implementing device functions, which will not be elaborated upon here.

[0146] An embodiment of the present invention provides a computer-readable storage medium storing at least one computer program, which is loaded and executed by a processor to enable a computer to implement any of the above-described methods.

[0147] Alternatively, the computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a compact disc read-only memory (CD-ROM), magnetic tape, a floppy disk, and an optical data storage device, etc.

[0148] In an exemplary embodiment, a computer program mask or computer program is also provided, which includes computer instructions stored in a computer-readable storage medium. A processor of an electronic device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the electronic device to perform any of the methods described above.

[0149] It should be noted that the terms "first" and "second" in the specification and claims of this application are used to distinguish similar objects and represent a limitation on a specific order or sequence. Where appropriate, the order of use for similar objects can be interchanged so that the embodiments of this application described herein can be implemented in an order other than that shown or described.

[0150] Those skilled in the art will recognize that this invention can be implemented as a system, method, or computer program mask. Therefore, this disclosure can be specifically implemented in the following forms: it can be entirely hardware, entirely software (including firmware, resident software, microcode, etc.), or a combination of hardware and software, generally referred to herein as a "circuit," "module," or "system." Furthermore, in some embodiments, the invention can also be implemented as a computer program mask in one or more computer-readable media containing computer-readable program code.

[0151] Any combination of one or more computer-readable media may be used. A computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. A computer-readable storage medium can be, but is not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples (a non-exhaustive list) of computer-readable storage media include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this document, a computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in connection with an instruction execution system, apparatus, or device.

[0152] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A defect detection method for reticle masks, characterized in that, include: Based on the detection magnification information and the name of the mask template to be tested in the detection command, the detection parameters of the equipment used to perform defect detection on the mask template to be tested are adjusted, and the detection parameters are associated and stored with the name of the mask template to be tested; Based on the grayscale values ​​of the light calibration points in the reference image corresponding to the mask to be tested, the light source brightness of the reference image is corrected, and a corrected reference image is generated. The corrected reference image is divided into detection regions. The ORB feature points corresponding to the test Die image and the reference ORB feature points corresponding to the reference Die image are extracted from the division results. Hamming distance matching and RANSAC are used to remove mismatches and establish sub-pixel level corresponding point pairs. By combining the subpixel-level corresponding point pairs, a transformation matrix is ​​determined to convert the test die image into the reference die image. Based on the transformation matrix, the test die image and the reference die image are aligned, and defect detection is performed based on the alignment result. Before the light source brightness correction, the mask to be tested also needs to be corrected, specifically: After successful initialization, the detection magnification, product name, and ROI parameters issued by the software are loaded into memory. The reference template image corresponding to the product name is read from the template library, and ORB feature points are extracted from the four corners and the center of the reference template image. At the same time, the original test image is transferred to the memory buffer. ORB descriptor matching is performed on the two images. First, Hamming distance is used for brute-force matching to obtain coarse matching point pairs. Then, RANSAC is used to iteratively eliminate mismatches, retaining ≥8 pairs of corresponding points with the highest confidence. Based on the highest ≥8 pairs of corresponding points, the translation vector (dx, dy) and rotation angle θ of the test image relative to the reference template are calculated, and a 2×3 affine matrix is ​​generated. The control system drives the high-precision XTθ platform for real-time compensation according to the affine matrix. After the platform moves into place, the image is acquired again and the above matching-calculation-compensation closed loop is repeated until the translation vector (dx, dy) < 0.05 pixels and θ < 0.002 rad, confirming that the product has completed sub-pixel-level precise positioning.

2. The defect detection method for a reticle mask according to claim 1, characterized in that, The process of determining the light source brightness correction for the reference image is as follows: Multiple light calibration points are randomly selected at the boundary between the transparent and non-transparent areas in the reference image. The gray values ​​corresponding to the multiple light calibration points are obtained. A light distribution model is constructed by performing quadratic polynomial fitting. The reference image is then normalized using the light distribution model to complete the light source brightness correction.

3. The defect detection method for a reticle mask according to claim 1, characterized in that, The specific process for establishing sub-pixel level corresponding point pairs is as follows: The feature point pairs consisting of the ORB feature points to be tested and the reference ORB feature points are coarsely matched by Hamming distance matching, and the feature point pairs that meet the preset requirements are finely matched by RANSAC to remove mismatches, so as to obtain the sub-pixel level corresponding point pairs.

4. The defect detection method for a reticle mask according to claim 1, characterized in that, The specific process for determining defects based on the alignment results is as follows: The aligned Die image to be tested is compared with the reference Die image, and the suspected defect region is extracted by binarization segmentation. After preprocessing the suspected defective areas, feature analysis is performed using a preset analysis algorithm to determine the final defect result.

5. A defect detection method for a reticle mask according to claim 1, characterized in that, Also includes: Based on the feature identifier of the mask to be tested, the offset between the mask to be tested and the defect detection device is calculated, and the position of the mask to be tested is corrected based on the offset.

6. A defect detection system for reticle masks, employing the defect detection method for reticle masks as described in claim 1, characterized in that, The system includes: The initialization module is used to: adjust the detection parameters of the device used to perform defect detection on the mask template under test based on the detection magnification information and the name of the mask template under test in the detection instruction, and associate and store the detection parameters with the name of the mask template under test; The light source brightness correction module is used to: correct the light source brightness of the reference image based on the gray value of the light calibration point in the reference image corresponding to the mask to be tested, and generate the corrected reference image; The detection region determination module is used to: divide the detection region into the corrected reference image, extract the ORB feature points corresponding to the test Die image and the reference ORB feature points corresponding to the reference Die image in the division result, use Hamming distance matching and RANSAC to remove mismatches, and establish sub-pixel level corresponding point pairs. The detection module is used to: combine the subpixel-level corresponding point pairs to determine the transformation matrix for converting the die image to be tested into the reference die image; based on the transformation matrix, align the die image to be tested with the reference die image; and perform defect detection based on the alignment result.

7. A defect detection system for a reticle mask according to claim 6, characterized in that, The process of determining the light source brightness correction for the reference image is as follows: Multiple light calibration points are randomly selected at the boundary between the transparent and non-transparent areas in the reference image. The gray values ​​corresponding to the multiple light calibration points are obtained. A light distribution model is constructed by performing quadratic polynomial fitting. The reference image is then normalized using the light distribution model to complete the light source brightness correction.

8. A defect detection system for a reticle mask according to claim 6, characterized in that, The specific process for establishing sub-pixel level corresponding point pairs is as follows: The feature point pairs consisting of the ORB feature points to be tested and the reference ORB feature points are coarsely matched by Hamming distance matching, and the feature point pairs that meet the preset requirements are finely matched by RANSAC to remove mismatches, so as to obtain the sub-pixel level corresponding point pairs.

9. An electronic device, characterized in that, The electronic device includes a processor coupled to a memory storing at least one computer program, which is loaded and executed by the processor to enable the electronic device to perform the method as described in any one of claims 1 to 5.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores at least one computer program, which is loaded and executed by a processor to enable the computer to perform the method as described in any one of claims 1 to 5.

Images