Wafer detection pre-alignment system and method based on multi-level redundant architecture

Abstract

The application discloses a wafer detection pre-alignment system and method based on a multi-level redundancy architecture. The system comprises: an automatic candidate point discovery module for extracting candidate alignment points in a reference wafer image; a multi-dimensional quality scoring module for multi-dimensionally quantifying and scoring the candidate points; a multi-level redundancy point grouping module for grouping the candidate points into a primary site point group, a backup point group and an emergency point group in a three-level redundancy architecture; an image registration engine for calculating image offset of a template and a search area and outputting a matching confidence; and a runtime adaptive point selection module for automatically and seamlessly switching between the three-level point groups according to the confidence. The application realizes automatic and objective selection of pre-alignment points and multi-level redundancy fault tolerance, can adapt to process changes, and significantly improves the robustness of pre-alignment and the continuity of the detection process.

Description

Technical Field

[0001] This invention relates to the field of wafer defect detection technology in semiconductor manufacturing, and specifically to an intelligent pre-alignment system and method for wafer inspection. Background Technology

[0002] During semiconductor manufacturing, wafer surfaces are prone to various minute defects (such as particles, scratches, and missing patterns), necessitating rapid and accurate defect detection during production. The current mainstream defect detection method involves die-to-die differential detection between adjacent dies: using the first die as a reference and the second die as the target die, the two are aligned, and the difference is calculated; areas of difference exceeding a threshold are identified as defects.

[0003] This inspection method requires extremely high positioning accuracy of the die grid. If there is a deviation in the die grid positioning, pixel-level misalignment will occur between the reference die and the die to be inspected, resulting in a large amount of residual after differentiation, which will lead to false alarms or missed detections. To ensure the positioning accuracy of the die grid during operation, existing wafer inspection equipment generally adopts a pre-alignment mechanism: before the inspection operation, the operator manually selects one or two die inner corner points with unique geometric features on the wafer as reference points. During operation, the system searches for these points through a template matching algorithm, thereby determining the precise positions of the first two dies and the die spacing, and then calculating the coordinates of the entire die grid.

[0004] The basic process of existing pre-alignment methods is as follows: The operator first manually selects a representative die as a reference, then observes the die's geometric pattern in the real-time video image using a joystick, manually positioning the crosshairs on the unique geometric shapes at the upper left and upper right corners of the die, marking them as the first and second points respectively (for array mode, only one point is marked). The system stores the coordinate data of these points, and during runtime, uses a template matching algorithm (such as grayscale correlation matching or geometric pattern matching) to search for the position in the acquired image that best matches the template point. The die grid position is calculated based on the matching result, thus completing the pre-alignment.

[0005] In addition to manually marking the points themselves, operators also need to set multiple matching parameters based on experience, including: focus offset, imaging mode selection, edge filter type, kernel size and threshold, input filter type, etc.

[0006] Existing pre-alignment methods rely entirely on operators manually selecting points. The quality of these points depends on personal experience and lacks objective evaluation standards. Furthermore, the fixed one or two points lack redundancy and fault tolerance; failure to match any point leads to detection abort. Simultaneously, process variations between batches cause the original geometric pattern appearance to gradually evolve, but current technology lacks the ability to assess and warn of point stability across batches. Point failures are only detected when detection fails during runtime.

[0007] At the matching algorithm level, traditional grayscale correlation matching or geometric pattern matching is highly sensitive to fluctuations in imaging conditions such as changes in illumination, focus drift, and local blurring. All parameters need to be set manually and do not output matching confidence, making it impossible for operators to predict the reliability of the points before operation. In addition, there are many and interdependent parameters associated with the points (focus offset, imaging mode, filter, light source training, etc.), and improper configuration directly affects matching performance; the additional scanning band for pre-alignment increases the detection time overhead and lacks optimization methods. Summary of the Invention

[0008] The technical problems to be solved by this invention are: how to achieve automated and intelligent selection of pre-alignment points, eliminating reliance on the operator's personal experience; how to construct a multi-layered redundant point architecture, seamlessly switching when point matching fails, avoiding interruption of the detection process; how to make point selection and matching strategies adaptable to process changes between batches, ensuring long-term stability; how to improve the robustness of the matching algorithm to complex imaging conditions; and how to achieve automatic optimization of supporting parameters.

[0009] To address the aforementioned technical problems, this invention provides a wafer inspection pre-alignment system based on a multi-level redundancy architecture, comprising:

[0010] The automatic candidate point discovery module is used to perform full-image analysis on the reference grain image, extract multiple candidate alignment points, and form a candidate point set;

[0011] The multi-dimensional quality scoring module is used to quantitatively score each candidate point on multiple preset quality dimensions, obtain a comprehensive quality score, and output a sorted list.

[0012] The multi-layer redundant site grouping module is used to group candidate sites into a three-layer redundant site group based on the comprehensive quality score, which includes at least a main site site group, a backup site site group, and an emergency site site group, and save it to the test formula;

[0013] An image registration engine is used to receive template image patches and search region image patches during detection runtime, calculate image offsets, and output matching confidence scores.

[0014] The runtime adaptive point selection module is used to automatically switch to the next level of points for pre-alignment when the confidence of the current level point is lower than a preset threshold during the detection process, based on the matching confidence.

[0015] This invention also provides a wafer inspection pre-alignment method based on a multi-level redundancy architecture, comprising: an automatic candidate site discovery step, which performs full-image analysis on a reference die image to extract candidate sites; a multi-dimensional quality scoring step, which scores and sorts the candidate sites in multiple dimensions; a multi-level redundant site grouping step, which groups the candidate sites into three-level redundant site groups; an image registration step, which calculates the offset between the template image block and the search area image block and outputs the confidence score during inspection; and a runtime adaptive site selection step, which automatically switches between the three-level site groups based on the confidence score.

[0016] Compared with the prior art, the technical solution of the present invention has the following advantages:

[0017] (1) High degree of automation and strong objectivity and consistency: By automatically discovering candidate sites and combining uniqueness, sharpness, feature content and orientation integrity into a four-dimensional quantitative scoring, site selection is transformed from subjective human experience judgment to objective and repeatable quantitative evaluation, eliminating the impact of individual operator differences on site quality. Different operators and different times can obtain consistent site selection results for the same wafer.

[0018] (2) Strong fault tolerance and uninterrupted detection process: The three-level redundant point grouping architecture, combined with the runtime matching confidence-driven automatic switching strategy, realizes seamless switching when point matching fails (switching time is in milliseconds), avoiding the problem of "single point failure causing detection to stop" in the existing technology. The entire switching process is transparent to the operator and can be automatically completed without manual intervention.

[0019] (3) Adaptive to process changes and continuous optimization: Based on the online learning mechanism of process stability based on exponential moving average, the system can continuously track and update the stability status of each point across batches, and automatically perform degradation, elimination and replacement operations when the performance of the point continues to decline. This solves the core problem in the prior art that "the point gradually fails as the process drifts and no one knows it until it is passively discovered when the detection fails during operation".

[0020] (4) Strong matching robustness and high sub-pixel accuracy: The image registration engine first performs coarse positioning through phase correlation, and then calculates precise image offset based on gradient descent pixel alignment, which solves the problems of insufficient phase correlation accuracy and gradient descent pixel alignment being sensitive to initial values.

[0021] (5) Automated parameter configuration and reduced usage threshold: The filter recommendation driven by focus curve fitting and noise estimation automates a large number of matching parameters that need to be manually configured one by one, reducing the decline in matching performance caused by improper parameter configuration and also reducing the requirements for the operator's professional skills.

[0022] (6) The system architecture is modular and highly scalable: the functional modules are connected through standardized interfaces and can be flexibly combined according to actual needs; the weight coefficients of the multidimensional quality score can be customized according to the pattern characteristics of different process layers; the parameters of process stability learning can be flexibly configured according to the stability of the production line process. Attached Figure Description

[0023] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0024] Figure 1 This is a system architecture diagram provided in an embodiment of the present invention.

[0025] Figure 2 yes Figure 1 A structural diagram of the image registration engine.

[0026] Figure 3 This is a diagram comparing the effects of an image registration engine and traditional grayscale matching.

[0027] Figure 4 This is a flowchart of the automatic candidate site discovery module. Detailed Implementation

[0028] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0029] 1. System Overall Structure (Divided into Modules)

[0030] The intelligent pre-alignment system, as shown in Figure 1, includes the following eight modules:

[0031] 1) Automatic candidate point discovery module: Performs full-image analysis on the reference die image, integrates multi-scale Harris corner detector and scale-invariant feature transform (SIFT) or ORB feature point detector, and automatically outputs a set of all candidate alignment points that meet the basic geometric conditions through a four-fold screening mechanism of edge exclusion, duplicate region exclusion, low contrast exclusion and non-maximum suppression.

[0032] 2) Multidimensional quality scoring module: Receives a set of candidate points, quantifies and scores each candidate point in four dimensions: uniqueness, sharpness, feature content, and orientation completeness, and outputs a ranked list of candidate points with a comprehensive quality score.

[0033] 3) Multi-level redundant point grouping module: Based on the scoring results, candidate points are grouped into three levels: primary point (G_primary), backup point (G_backup), and emergency point (G_fallback). Combined with spatial distribution uniformity constraints, it ensures that the points at each level are uniformly covered within the die. The grouped three-level redundant point group is output and saved to the detection formula.

[0034] 4) Image Registration Engine: Receives a point template patch and a runtime search region patch. It first estimates the overall translation using phase correlation, and then calculates the precise sub-pixel image offset (u,v) using a gradient descent-based pixel alignment method. Let the template... The peak intensity is Search area peak intensity The confidence level of phase correlation is Confidence of pixel alignment ,in x and y are the coordinates on the image. The confidence level of image registration. .

[0035] Ideally, the peak value is 1 when the two images are completely identical; noise, occlusion, or pattern repetition can cause the peak value to drop, and the offset between the two images is not eliminated, thus leading to a decrease in confidence.

[0036] 5) Adaptive parameter optimization module: During the setup phase, the module performs light source brightness scanning (using brightness histogram entropy, local contrast, and overexposure / underexposure ratio as quality indicators), focus scanning (Z-axis stepping, Laplacian variance fitting of focus curve), and noise estimation (wavelet domain noise level calculation) on candidate points. It automatically recommends the optimal light source brightness, focus offset, and filter type and parameters, which are then provided to the automatic candidate point discovery module and the image registration engine, respectively.

[0037] 6) Process stability learning module: During the operation phase, the matching confidence C_match and matching residual R_match of each test are collected. The process stability score S_stability of each point is continuously updated using exponential moving average (EMA). The module automatically performs downgrading, elimination and replacement operations based on the score and feeds the updated stability score back to the runtime adaptive point selection module.

[0038] 7) Runtime adaptive point selection module: Before each batch of detection begins, the global brightness, global focus and local anomalies of the real-time acquired image are quickly measured. Based on the measurement results, the optimal point combination is dynamically selected from the three-level redundant point group, and the template patch and search area patch of the selected point are sent to the image registration engine.

[0039] The eight modules described above can be divided into three main units based on their operational phases: the "Setup Phase Core Unit," consisting of the automatic candidate site discovery module, the multi-dimensional quality scoring module, and the multi-layer redundant site grouping module, is responsible for automatically discovering, evaluating, and grouping sites during the test formulation setup process; the "Running Phase Core Unit," consisting of the image registration engine, the process stability learning module, and the runtime adaptive site selection module, is responsible for performing site matching, continuous learning, and dynamic optimization during test operation; and the "Auxiliary Support Unit," based on adaptive parameter optimization, provides automatic parameter recommendation capabilities during the setup phase. All modules are connected via standardized interfaces and can be flexibly combined or upgraded independently according to actual needs.

[0040] 2. Example

[0041] 1) Example 1

[0042] Example 1 Module Composition:

[0043] The automatic candidate point discovery module, the multi-dimensional quality scoring module, and the multi-level redundant point grouping module are sequentially connected to form an automated point selection pipeline in the setup phase. The automatic candidate point discovery module performs a full-image analysis on the reference die image and outputs a set of candidate points; the multi-dimensional quality scoring module scores each candidate point and outputs a list of candidate points with scores; the multi-level redundant point grouping module groups the candidate points into different levels based on the scoring results and outputs a three-level redundant point group.

[0044] Example 1 Working Method:

[0045] Step 1—Automatic Candidate Point Discovery: The high-resolution image of the reference die is first preprocessed with Gaussian filtering for noise reduction and edge-preserving filtering. Then, a multi-scale Harris corner detector is used to extract a set of candidate corner points, while a scale-invariant feature point set is extracted using either Scale Invariant Feature Transform (SIFT) or ORB feature point detector. The union of these two sets is taken as the initial candidate set. The initial candidate set undergoes a four-stage filtering process: points located in the edge regions of the die are excluded (to avoid edge effects); points located in repetitive pattern regions are excluded (the autocorrelation function is calculated for the neighborhood window of each candidate point; points with autocorrelation peaks exceeding a threshold are considered repetitive regions and excluded); points located in low-contrast regions are excluded; and non-maximum suppression is applied to excessively dense points. This filtering process yields a set of valid candidate points.

[0046] Step 2—Multidimensional Quality Scoring: Calculate quality scores in four dimensions for each candidate point:

[0047] The uniqueness score S_unique = 1 − maxⱼ(NCC(patch(p_i), patch(p_j))), where NCC is the normalized cross-correlation. This score measures the uniqueness of the local pattern of a point within the entire die. The more unique the point, the less likely it is to be mismatched to the wrong location during runtime.

[0048] The sharpness score S_sharpness = Var(Laplacian(patch(p_i))) uses Laplacian variance to measure the local image sharpness of a point. Points with higher sharpness contain more edge and texture information.

[0049] The feature content score S_feature = (∑|∇I(x,y)|) / (W×W), which is the mean of the gradient magnitude, measures the feature density of the points.

[0050] The orientation integrity score S_orientation = min(H_grad_x, H_grad_y) / max(H_grad_x, H_grad_y), where H_grad_x and H_grad_y are the entropies of the gradient histograms in the horizontal and vertical directions within the window, respectively. It measures whether the point contains significant horizontal and vertical features at the same time. The closer the score is to 1, the better.

[0051] The overall quality score S_total = w1·S_unique + w2·S_sharpness + w3·S_feature + w4·S_orientation, with the weight coefficients all set to 0.25 by default. These can be customized and adjusted according to the pattern characteristics of different process layers.

[0052] Step 3—Multi-layer Redundancy Site Grouping: Based on the comprehensive quality score, candidate sites are grouped into a three-layer redundancy architecture. The first layer is the primary site group G_primary (default 4 sites), which selects several sites with the highest comprehensive scores, divides the die into a corresponding number of sub-regions, and selects the best site within each sub-region to ensure uniform spatial distribution. The second layer is the backup site group G_backup (default 4 sites), which selects the second-best sites from the remaining candidates, and maintains a certain spatial distance from the primary site sites to avoid common-cause failures. The third layer is the emergency site group G_fallback (default 2 sites), which selects the next best sites as a last resort, using a relatively lenient matching standard. The three-layer site groups and their feature descriptor caches are saved together in the detection recipe.

[0053] Example 1: Runtime Control Logic

[0054] Step 1: The system prioritizes using the primary site location in G_primary for alignment matching.

[0055] Step 2: Detect the matching confidence C_match for each main site location:

[0056] If the C_match of all master sites is greater than or equal to the confidence threshold τ_conf, then the matching result is used directly to complete the pre-alignment;

[0057] If the C_match value of any point is lower than τ_conf, the system will automatically switch to the corresponding backup point in G_backup and use the matching result of the backup point.

[0058] Step 3: If the corresponding point in G_backup also fails to match (C_match<τ_conf), then the emergency point in G_fallback is further enabled.

[0059] Step 4: The entire switchover process is completed automatically in the background without interrupting the testing process. Operators can view the switchover record and reason annotations in the operation report. The system also sends the matching data (C_match and matching residual R_match) to the process stability learning module for subsequent analysis.

[0060] Example 1 Implementation Results:

[0061] It achieves fully automated discovery, multi-dimensional objective scoring, and redundant grouping of pre-aligned points, transforming point selection from subjective judgment relying on operator experience to objective evaluation based on quantitative indicators. The three-layer redundancy architecture provides two-level failover capability, seamlessly switching to backup or emergency points when the primary site fails due to process changes, localized contamination, or other reasons, avoiding interruptions to the testing process. Operators can view the scoring details and spatial distribution of each point on the interface, confirming the system's recommendations or manually fine-tuning them, balancing automation efficiency with the flexibility of manual oversight.

[0062] 2) Example 2

[0063] Example 2 Module Composition:

[0064] Based on the system architecture shown in Figure 1, this embodiment adds an image registration engine. This engine consists of a phase correlation module and a gradient descent-based pixel alignment module.

[0065] Example 2 Connection method:

[0066] The image registration engine's inputs are connected to an automatic candidate point discovery module (to acquire image patches of template points) and a runtime image acquisition system (to acquire image patches of the search area). First, the overall translation is estimated using the phase correlation method. Then, a gradient descent-based pixel alignment method is used to calculate the precise sub-pixel image offset, outputting the final matching coordinates and confidence score.

[0067] Example 2 Working Method:

[0068] During the setup phase, the system uses the automatic candidate point discovery module to identify candidate points and then extracts all point template patches at once.

[0069] During the detection process, for each search region (default search window 128×128 pixels), the overall translation is first estimated using the phase correlation method, and then the precise sub-pixel image offset is calculated using the pixel alignment method based on gradient descent. The final matching coordinates and confidence scores are then output.

[0070] Example 2 Implementation Results:

[0071] After optimizing the image registration method, the matching success rate increased from approximately 72% for traditional grayscale correlation matching to approximately 96%, and the sub-pixel matching accuracy improved from ±0.5 pixels to ±0.15 pixels. After INT8 quantization and template feature pre-caching, the single-point inference time is less than 5 milliseconds, meeting the real-time requirements of online high-throughput detection. Furthermore, each matching outputs a confidence score, providing a quantifiable decision-making basis for subsequent process stability learning and runtime adaptive selection.

[0072] 3) Example 3

[0073] Example 3 Module Composition:

[0074] Based on the system architecture shown in Figure 1, a process stability learning module and a runtime adaptive site selection module are added.

[0075] Example 3 Working Method:

[0076] After each test run, the process stability learning module collects the actual matching performance data for each used point, including the matching confidence C_match and the matching residual R_match. The process stability score S_stability for each point is continuously updated using the exponential moving average (EMA) algorithm.

[0077] S_stability(p_i)^(t)=α·S_stability(p_i)^(t−1)+(1−α)·f(C_match^(t),R_match^(t)).

[0078] Where α is the exponential moving average coefficient (default α = 0.85), t represents the t-th batch, and f() is a function that maps the matching confidence and residuals to stability increments.

[0079] Example 3: Dynamic Location Management Logic

[0080] Step 1: If the matching confidence C_match of a certain point is lower than the downgrade threshold τ_degrade for K consecutive batches, the system will automatically downgrade the point (the main station will be downgraded to the backup station, and the backup station will be downgraded to the emergency station). After downgrading, the unused candidate point with the highest current comprehensive score will be selected from the candidate pool to supplement the corresponding level.

[0081] Step 2: If the matching confidence C_match of a certain point in M ​​consecutive batches is lower than the elimination threshold τ_remove (M > K, τ_remove < τ_degrade), the system will automatically remove the point permanently from the candidate set and select a replacement point from the candidate pool.

[0082] Step 3: Before each batch of detection begins, the runtime adaptive point selection module performs rapid measurements on the real-time acquired images (global brightness anomaly detection, global focus quality assessment, and local occlusion detection), and dynamically adjusts the point usage strategy based on the measurement results: when the overall brightness is normal, the default main station location is used; when local occlusion is detected and the occluded area overlaps with a main station location, the module automatically switches to the corresponding backup station location; when global blur is detected, the station location with the highest sharpness score is prioritized; when excessively high or low brightness is detected, the station location with the highest feature content score is prioritized.

[0083] Example 3 Implementation Results:

[0084] The system possesses cross-batch self-learning and adaptive optimization capabilities. When certain locations gradually fail due to process drift, the system can identify risky locations in advance by observing the continuous decline in matching confidence before operational failures occur. It then automatically performs downgrading or elimination operations and adds new locations. The runtime adaptive selection strategy further enhances the system's ability to cope with fluctuations in real-time imaging conditions.

[0085] 4) Example 4

[0086] Example 4 Module Composition:

[0087] Based on the system architecture shown in Figure 1, an adaptive parameter optimization module is added.

[0088] Example 4 Working Method—Adaptive Parameter Optimization:

[0089] Light source brightness optimization: The image region containing candidate points is traversed through light source brightness levels. At each level, images are acquired and the brightness histogram entropy, local contrast, and the proportion of overexposed and underexposed pixels are calculated. The brightness level with the best comprehensive index is selected as the recommended value. Focus offset optimization: The region containing candidate points is scanned in Z-axis steps. At each step, images are acquired and the Laplacian variance is calculated as a focus quality index. After fitting the focus curve, the Z-axis offset corresponding to the peak value is used as the recommended focus offset. Filter parameter recommendation: Based on the noise level of the region containing candidate points (obtained through wavelet domain noise estimation), the filter type is automatically recommended—pass-through or smoothing filter is recommended when noise is low, median filter is recommended when noise is high, and edge filter (default) is recommended when noise is moderate.

[0090] Example 4 Implementation Results:

[0091] Adaptive parameter optimization automates the process of manually setting multiple matching parameters (light source brightness, focus offset, filter type and parameters, etc.) that operators previously had to set manually, reducing the workload of parameter configuration and the risk of human error.

[0092] The four embodiments above achieve progressive technical effects through different module combinations. Embodiment 1 is the basic embodiment, realizing automated review and redundant grouping of point selection; Embodiment 2, based on Embodiment 1, replaces traditional grayscale matching with deep feature matching, improving matching robustness and accuracy; Embodiment 3, based on Embodiment 2, adds online learning and adaptive selection capabilities, realizing continuous optimization across batches; Embodiment 4, based on Embodiment 3, adds parameter self-optimization capabilities, realizing a complete setup-run closed loop.

[0093] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A wafer inspection pre-alignment system based on a multi-level redundant architecture, comprising: include: The automatic candidate point discovery module is used to perform full-image analysis on the reference grain image, extract multiple candidate alignment points, and form a candidate point set; The multi-dimensional quality scoring module is used to quantify and score each candidate point in the candidate point set on multiple preset quality dimensions, obtain the comprehensive quality score of each candidate point, and output a list of candidate points sorted by comprehensive quality score. The multi-layer redundant site grouping module is used to group candidate sites into three-layer redundant site groups, including at least a main site group, a backup site group, and an emergency site group, based on the comprehensive quality score of the candidate site list, and save the information of each layer of site groups to the test formula. The image registration engine is used to receive the template image block and the search area image block corresponding to the target point during the detection process, calculate the image offset between the template image block and the search area image block, and output the matching confidence score. The runtime adaptive point selection module is used to obtain the matching confidence of each point output by the image registration engine during the detection runtime. When the matching confidence of the point at the current level is lower than the preset threshold, it switches to the point at the next level for pre-alignment. The switching is performed layer by layer between the main point group, the backup point group and the emergency point group. The multidimensional quality scoring module scores each candidate point in four dimensions: uniqueness, sharpness, feature content, and directional integrity. The uniqueness score is measured by calculating the maximum normalized cross-correlation between the local image patch of the candidate point and other regions within the grain; the sharpness score is measured by calculating the Laplacian variance of the local image patch of the candidate point; the feature content score is measured by calculating the mean gradient magnitude of the local image patch; and the orientation integrity score is measured by calculating the uniformity of the gradient orientation distribution of the local image patch in the horizontal and vertical directions. The overall quality score is a weighted sum of the scores from the four dimensions mentioned above; When the multi-layer redundant point grouping module groups candidate points into each level, it divides the grain into sub-regions corresponding to the number of points in each level based on the overall quality score and applies spatial distribution uniformity constraints. In each sub-region, the candidate point with the best overall quality score is selected to ensure that the points of each level are uniformly covered in the grain and that the points of different levels maintain a predetermined spatial distance. The process stability learning module is used to collect the matching confidence and matching residual of the used points after each test run, continuously update the process stability score of each point using the exponential moving average algorithm, and perform point downgrading, elimination and replacement operations based on the process stability score. The runtime adaptive point selection module is also used to perform global brightness anomaly detection, global focus quality assessment, and local occlusion detection on the real-time acquired images before the start of each batch of detection, and dynamically adjust the point usage strategy based on the detection results: When the overall system is functioning normally, use the primary site location group; When partial occlusion is detected and the occluded area overlaps with the current main site location, the system switches to the corresponding backup site location. When global blur is detected, the point with the highest sharpness score is used; When excessively high or low brightness is detected, the point with the highest feature content score is used.

2. The system of claim 1, wherein, The automatic candidate point discovery module extracts an initial candidate set by combining a multi-scale corner detector and a scale-invariant feature point detector, and then performs a four-fold screening on the initial candidate set, namely edge exclusion, duplicate region exclusion, low contrast exclusion and non-maximum suppression, to obtain the candidate point set.

3. The system of claim 1, wherein, The image registration engine includes a phase correlation module and a gradient descent-based pixel alignment module; The phase correlation module is used to estimate the overall translation of the template image block and the search region image block; The gradient descent-based pixel alignment module is used to calculate the image offset using the overall translation amount as the initial value, and outputs the final matching coordinates and matching confidence.

4. The system of claim 1, wherein, The site downgrading, elimination, and replacement operations include: When the matching confidence of a certain point is lower than the downgrade threshold for K consecutive batches, the point is downgraded to the next level, and the unused candidate point with the highest comprehensive quality score is selected from the candidate pool to supplement the original level. When the matching confidence of a certain point is lower than the elimination threshold for M consecutive batches, the point is removed from all levels, and a replacement point is selected from the candidate pool to supplement the corresponding level, where M>K and the elimination threshold is lower than the downgrade threshold.

5. The system of any one of claims 1 to 4, wherein, It also includes an adaptive parameter optimization module, which performs light source brightness scanning, focus scanning and noise estimation on the image region where the candidate point is located, recommends the optimal light source brightness, focus offset and filter type and parameters, and provides the recommended parameters to the automatic candidate point discovery module and the image registration engine respectively.

6. A wafer detection pre-alignment method based on a multi-level redundant architecture, using the system of any one of claims 1-5, characterized in that, Includes the following steps: Automatic candidate site discovery steps: Perform full-image analysis on the reference grain image to extract multiple candidate alignment sites and form a candidate site set; Multidimensional quality scoring steps: Quantify and score each candidate point in the candidate point set on multiple preset quality dimensions to obtain the comprehensive quality score of each candidate point, and output a list of candidate points sorted by comprehensive quality score; Multi-layer redundant site grouping steps: Based on the comprehensive quality score of the candidate site list, the candidate sites are grouped into a three-layer redundant site group that includes at least a main site site group, a backup site site group, and an emergency site site group, and the information of each layer of site group is saved to the test formula; Image registration steps: During the detection process, the template image block and the search region image block corresponding to the target point are obtained. The image offset is calculated by combining the phase correlation method and the pixel alignment method based on gradient descent, and the matching confidence is output. Runtime adaptive point selection steps: During the detection runtime, the matching confidence of each point is obtained. When the matching confidence of the point at the current level is lower than the preset threshold, the point is switched to the next level for pre-alignment. The switching is carried out layer by layer between the main point group, the backup point group and the emergency point group.

Images