Integrated circuit mask plate defect multispectral cooperative detection method and system

Abstract

The invention relates to the technical field of defect detection, and discloses an integrated circuit mask plate defect multispectral cooperative detection method, which comprises the following steps: carrying out preliminary scanning on a mask plate by using a multispectral camera, selecting an optimal wave band combination, switching to a target wave band through a liquid crystal adjustable optical filter, and obtaining a high-confidence multispectral image; the optical path offset is calculated through the linear relation, the deflection angle of the micro-mirror is dynamically adjusted, and images are collected for the second time after imaging offset is compensated; image filtering and registration are completed in the FPGA, and a high-contrast fusion image is generated; if the confidence coefficient is insufficient, triggering a wave band reselection mechanism to recheck; dynamically optimizing a defect type-spectrum mapping table based on real-time detection data; when the environment fluctuation exceeds the limit, an anti-interference mode is automatically switched, and micro-mirror calibration is triggered; and monitoring performance degradation of the optical assembly in real time, switching redundant optical paths, cleaning damaged parts, triggering manual reinspection when conflicts occur, and finally generating a visual report. According to the invention, the detection accuracy of the integrated circuit mask plate can be improved.

Description

Technical Field

[0001] This application relates to the technical field of defect detection, and in particular to a multi - spectral collaborative detection method and system for integrated circuit mask defects. Background Art

[0002] In integrated circuit manufacturing, mask defect detection is a core link to ensure the chip yield. However, existing technologies have significant limitations in complex scenarios. Dynamic changes in environmental parameters (such as temperature and vibration) can cause optical path offsets. Traditional compensation methods (such as fixed filters or mechanical calibration) have slow response speeds and insufficient accuracy, making it difficult to cope with real - time disturbances such as high - frequency vibration or sudden changes in temperature and humidity. For example, the optical path redundancy technology proposed in Document 7 only focuses on fault switching and does not achieve dynamic closed - loop compensation for small offsets.

[0003] Secondly, the selection of detection bands depends on preset rules and fails to dynamically optimize in combination with real - time defect features, resulting in limited detection sensitivity. As can be seen from the above, how to improve the detection accuracy of integrated circuit masks remains to be solved. Summary of the Invention

[0004] In order to improve the detection accuracy of integrated circuit masks, this application provides a multi - spectral collaborative detection method and system for integrated circuit mask defects.

[0005] In a first aspect, this application provides a multi - spectral collaborative detection method for integrated circuit mask defects, adopting the following technical solutions:

[0006] A multi - spectral collaborative detection method for integrated circuit mask defects includes:

[0007] Preliminarily scan the mask through a multi - spectral camera to obtain the corresponding initial multi - band image, identify the corresponding candidate defect types based on the initial multi - band image. The candidate defect types include metal residue, transparent film defects, and phase distortion. Dynamically select the optimal band combination from a pre - trained defect type - spectrum mapping table according to the candidate defect type. If the candidate defect type is a metal residue defect, activate the near - infrared band and the deep ultraviolet band. If the candidate defect type is a transparent film defect, activate the extreme ultraviolet band and the visible light band. If the candidate defect type is a phase distortion defect, activate the ultraviolet band and the mid - infrared band. Switch to the target band combination within 10 microseconds through a liquid crystal tunable filter to obtain a multi - spectral image corresponding to high confidence.

[0008] Based on the temperature and humidity sensors and vibration sensors, the corresponding temperature change amount, humidity change amount and vibration acceleration component are determined in real time. Based on the temperature change amount and vibration acceleration component, the offsets of the optical path in the X-axis and Y-axis are calculated. The offsets have a linear relationship with the temperature change amount and vibration acceleration component. The deflection angle of the micromirror is dynamically adjusted according to the offsets and the focal length of the optical system to compensate for the imaging offset in real time. After compensation, the multi-spectral camera acquires images twice to obtain the corresponding secondary multi-spectral images.

[0009] The field-programmable gate array is used to filter the secondary multi-spectral images and dynamically adjust the filtering neighborhood radius. The multi-band image registration is completed in the FPGA to generate a fused high-contrast image, and the coordinates of the candidate defect regions are extracted. The target detection model is loaded in the graphics processing unit. The target detection model integrates the spectral attention mechanism to locate the coordinates of the candidate defect regions, and outputs the coordinates of the defect bounding box and the defect confidence. If the detected defect confidence is lower than the preset threshold, the band re-selection mechanism is triggered, and the inspection is re-performed by switching to the supplementary band combination.

[0010] The defect type - spectrum mapping table is updated based on the real-time detection results. If there is a band combination with a continuous detection rate exceeding 90% among the same type of defects, the corresponding priority is increased. If there is a band combination with a false alarm rate exceeding 5%, the corresponding weight is reduced or it is temporarily disabled. When the environmental fluctuation exceeds the limit, it automatically switches to the anti-interference mode, closes the bands sensitive to environmental impacts, and triggers the full-scan calibration process of the MEMS micromirror. The real-time detection results include the defect type, position coordinates, confidence, environmental parameters and offsets.

[0011] The performance degradation indicators corresponding to the optical components are monitored in real time. The performance degradation indicators include the decrease in the transmittance of the filter and the degree of lens surface contamination. When any indicator exceeds the preset threshold, it automatically switches to the redundant optical path and triggers the cleaning procedure. And an energy spectrum analysis module is integrated to verify the elemental composition of the defect region. If there is a conflict between the optical detection result and the energy spectrum data, the manual re-inspection process is triggered. A visual inspection report is generated, marking the defect position, type, confidence and elemental composition.

[0012] Optionally, the method further includes: analyzing the temperature and humidity changes and vibration frequency in real time through a multi-sensor fusion algorithm to construct a dynamic interference model. Among them, the optical path offset caused by temperature change is corrected by a non-linear compensation curve, and the offset caused by vibration is compensated after separating the high-frequency noise and low-frequency trend through an adaptive filter. The deflection angle of the micromirror is dynamically adjusted through a step-by-step iterative algorithm. After each iteration, a verification image is acquired and the deflection parameters are feedback-optimized.

[0013] Optionally, the band reselection mechanism, the method further includes: when the defect confidence is lower than the threshold, calling the historical defect database to match the current defect characteristics, if the match is successful, directly activating the historical optimal band combination; if the match is unsuccessful, selecting the standby band combination that is closest to the current defect spectral response through the spectral-feature correlation map, and the spectral-feature correlation map is generated based on a pre-trained deep learning model and includes a multi-dimensional similarity matrix of defect types and band combinations.

[0014] Optionally, when the energy spectrum analysis module is integrated with the multi-spectral detection system, a timing synchronization trigger mechanism is adopted, and the method further includes: when the optical detection determines a high-confidence defect, the energy spectrum analysis module performs an elemental composition scan on the same area; if the defect type detected by the optical detection conflicts with the elemental distribution detected by the energy spectrum analysis module, three-dimensional topography scanning is started, and the reflectance deviation of the optical image is corrected through the topography data, and finally a fusion verification report is generated.

[0015] Optionally, the method further includes: when it is detected that the performance of the optical component degrades, predicting the detection accuracy loss after switching to the redundant optical path through the fault impact assessment model, if the loss is lower than the preset threshold, automatically switching; if the loss exceeds the threshold, triggering the multi-band collaborative detection mode, maintaining the detection by enhancing the weight of the non-degraded band and introducing an adaptive noise reduction algorithm, and at the same time recording the degradation rate of the degraded component to predict the maintenance period.

[0016] Optionally, during the dynamic update of the defect type - spectrum mapping table, the method further includes: establishing a band combination performance evaluation system, generating a corresponding multi-dimensional performance index vector through the defect detection rate, environmental interference intensity, false alarm / missed alarm times, and system resource occupancy rate collected in real time; adopting an adaptive weight adjustment algorithm to perform dynamic priority sorting on the band combination based on the multi-dimensional performance index vector, where the weight adjustment rule includes: if there is a band combination with a continuous detection confidence exceeding 90% in the same type of defect, then increase the corresponding priority and trigger the performance migration with other low-efficiency bands; if there is a band combination with a false alarm rate exceeding the threshold and an environmental interference sensitivity higher than the average value, then reduce the corresponding weight or mark it as a to-be-verified state, and enable the collaborative detection mode of the alternative band combination; when the band combination marked as the to-be-verified state continues to have false alarms, start the knowledge transfer process across defect types, and inject the data of the corresponding band combination in the detection of other defect types into the training set of the current defect type.

[0017] Optionally, the multi-band image registration method further includes: extracting semantic feature maps of each band image through a deep learning model, and constructing a cross-band feature correlation map; calculating the sub-pixel displacement field of adjacent band images by using the optical flow method, and performing global consistency correction in combination with the feature correlation map; in the fusion stage, dynamically weighting the signal-to-noise ratios of each band through an attention mechanism, where the weight value is negatively correlated with the current environmental interference intensity of the band, and generating a corresponding fused image.

[0018] In a second aspect, the present application provides a multi-spectral collaborative detection system for integrated circuit mask defects, adopting the following technical solution:

[0019] A multi-spectral collaborative detection system for integrated circuit mask defects includes:

[0020] A candidate defect type recognition module, which preliminarily scans the mask through a multi-spectral camera to obtain corresponding initial multi-band images, and is used to identify corresponding candidate defect types based on the initial multi-band images. The candidate defect types include metal residue, transparent thin film defects, and phase distortion; dynamically select the optimal band combination from a pre-trained defect type - spectrum mapping table according to the candidate defect type. If the candidate defect type is a metal residue defect, activate the near-infrared band and the deep ultraviolet band. If the candidate defect type is a transparent thin film defect, activate the extreme ultraviolet band and the visible light band. If the candidate defect type is a phase distortion defect, activate the ultraviolet band and the mid-infrared band; switch to the target band combination within 10 microseconds through a liquid crystal tunable filter to obtain a multi-spectral image corresponding to high confidence;

[0021] An offset calculation module, which determines the corresponding temperature change, humidity change, and vibration acceleration component in real time based on a temperature and humidity sensor and a vibration sensor, and calculates the offsets of the optical path in the X-axis and Y-axis based on the temperature change and the vibration acceleration component. The offset is linearly related to the temperature change and the vibration acceleration component; dynamically adjust the deflection angle of the micromirror in real time according to the offset and the focal length of the optical system to compensate for the imaging offset; after compensation, collect images twice through a multi-spectral camera to obtain corresponding secondary multi-spectral images;

[0022] A high-contrast image generation module, which filters the secondary multi-spectral images and dynamically adjusts the filtering neighborhood radius through a field-programmable gate array, completes multi-band image registration in the FPGA, is used to generate a fused high-contrast image, and extracts the coordinates of candidate defect regions; load a target detection model in the graphics processor. The target detection model integrates a spectral attention mechanism, locates the coordinates of the candidate defect regions, and outputs the coordinates of the defect bounding box and the defect confidence. If the detected defect confidence is lower than a preset threshold, trigger a band re-selection mechanism and switch to a supplementary band combination for re-inspection;

[0023] An update module, which is used to update the defect type - spectrum mapping table based on real-time detection results. If the continuous detection rate of a band combination in the same type of defect exceeds 90%, the corresponding priority is increased; if the false alarm rate of a band combination exceeds 5%, the corresponding weight is reduced or it is temporarily disabled; when the environmental fluctuation exceeds the limit, it automatically switches to the anti-interference mode, closes the bands sensitive to environmental impacts, and triggers the full-scan calibration process of the MEMS mirror. Among them, the real-time detection results include defect type, position coordinates, confidence level, environmental parameters, and offset.

[0024] A performance degradation index monitoring module, which is used to monitor the performance degradation indexes corresponding to optical components in real time. The performance degradation indexes include the decrease in the transmittance of the filter and the degree of lens surface contamination. When any index exceeds the preset threshold, it automatically switches to the redundant optical path and triggers the cleaning procedure; and it integrates an energy spectrum analysis module to verify the elemental composition of the defect area. If there is a conflict between the optical detection result and the energy spectrum data, it triggers the manual re-inspection process; generates a visual detection report, marking the defect position, type, confidence level, and elemental composition.

[0025] In a third aspect, the present application provides an integrated circuit mask defect multi-spectral collaborative detection system, adopting the following technical solution:

[0026] An integrated circuit mask defect multi-spectral collaborative detection system includes a processor, and a program of the integrated circuit mask defect multi-spectral collaborative detection method described in any one of the above is run in the processor.

[0027] In a fourth aspect, the present application provides a storage medium, adopting the following technical solution:

[0028] A storage medium stores a program of the integrated circuit mask defect multi-spectral collaborative detection method described in any one of the above. Description of the Drawings

[0029] Figure 1 is a flowchart of an integrated circuit mask defect multi-spectral collaborative detection method shown according to an exemplary embodiment.

[0030] Figure 2 is a structural block diagram of an integrated circuit mask defect multi-spectral collaborative detection method device shown according to an exemplary embodiment. Detailed Embodiments

[0031] The following details the embodiments of the present application, and the examples of the embodiments are shown in the drawings.

[0032] In the description of this specification, the description referring to the terms "certain embodiments", "one embodiment", "some embodiments", "schematic embodiments", "examples", "specific examples", or "some examples" means that the specific features, structures, materials, or characteristics described in connection with the embodiments or examples are included in at least one embodiment or example of the present application. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiments or examples. Moreover, the specific features, structures, materials, or characteristics described may be combined in any one or more embodiments or examples in a suitable manner.

[0033] An embodiment of the present application discloses a multi-spectral collaborative detection method for integrated circuit mask defects. Referring to Figure 1 , it includes:

[0034] S100: Initially scan the mask plate with a multi-spectral camera to obtain the corresponding initial multi-band image, and identify the corresponding candidate defect types based on the initial multi-band image. The candidate defect types include metal residue, transparent thin film defects, and phase distortion; dynamically select the optimal band combination from the pre-trained defect type - spectrum mapping table according to the candidate defect type. If the candidate defect type is a metal residue defect, activate the near-infrared band and the deep ultraviolet band. If the candidate defect type is a transparent thin film defect, activate the extreme ultraviolet band and the visible light band. If the candidate defect type is a phase distortion defect, activate the ultraviolet band and the mid-infrared band; switch to the target band combination within 10 microseconds through a liquid crystal tunable filter to obtain the multi-spectral image corresponding to high confidence.

[0035] Among them, the configuration of the multi-spectral camera: The first step is hardware setting. Use a multi-spectral camera equipped with a multi-band filter set (such as visible light, near-infrared, ultraviolet, mid-infrared, etc.) to ensure coverage of the band range required for defect detection. For example, the near-infrared band (700 - 1000 nm) is used to detect the thickness of metal residues, the ultraviolet band (200 - 400 nm) is used to identify surface contamination, and the mid-infrared band (3 - 8 μm) is used to analyze phase distortion caused by thermal stress. The second step is scanning parameters. Set the exposure time, gain, and filter switching speed of the camera to ensure that the band switching is completed within 10 microseconds. For example, use a preset general band combination (such as visible light + near-infrared) during the initial scan to quickly cover the whole area.

[0036] Initial multi - band image acquisition: The first step is optical path control. The liquid crystal tunable filter (LCTF) is used to dynamically switch bands (such as sequentially activating the visible light, near - infrared, and ultraviolet bands). After each band switch, the camera acquires the image corresponding to that band. For example, the visible light band is used to observe surface reflection differences, and the ultraviolet band is used to detect material absorption characteristics. The second step is image alignment. The fixed focal length and mechanical structure of the camera are used to ensure the spatial consistency of images in different bands, avoiding offsets caused by mechanical vibrations (which need to be corrected in subsequent environmental compensation steps).

[0037] Candidate defect type identification: The initial multi - band images are input into a pre - trained deep learning model to extract spectral features in each band (such as edges, textures, reflectance curves). For example: Metal residues show high reflectivity in the near - infrared band and significant absorption in the ultraviolet band; Transparent film defects have abnormal transmittance in visible light and interference fringes in the extreme ultraviolet band; Phase distortion shows uneven thermal distribution in the mid - infrared band. The candidate defect type is judged through multi - band feature fusion (such as spectral attention mechanism), and the confidence level is output (such as 85% confidence level for metal residues).

[0038] Through multi - band global scanning, the preliminary defect location of the mask plate is completed within 10 seconds, reducing the computational load of subsequent high - precision detection; In addition, combined with spectral feature differences, single - band misjudgment is avoided (such as a transparent film may be misjudged as a clean area in visible light).

[0039] Then, the specific steps for dynamically selecting the optimal band combination are executed:

[0040] First is the call of the defect type - spectrum mapping table: Retrieve the optimal band combination corresponding to the defect type from the pre - trained "defect type - spectrum mapping table". For example: For metal residues, the near - infrared band (strong penetration, detecting thickness) + deep ultraviolet band (sensitive to metal surface contamination); For transparent film defects, the extreme ultraviolet band (detecting nanoscale thickness changes) + visible light band (observing abnormal transmittance); For phase distortion, the ultraviolet band (surface topography) + mid - infrared band (thermal stress analysis). And based on the candidate defect type (such as metal residues), retrieve the corresponding band combination from the mapping table.

[0041] Then is the band combination optimization: By real - time reading the data of the temperature and humidity sensors, if the environmental temperature and humidity fluctuate beyond the threshold (such as temperature change > 2°C), preferentially select bands with strong anti - interference ability (such as mid - infrared is not sensitive to temperature and humidity); If vibration interference is detected, enable the anti - vibration band combination (such as reducing the dependence on the ultraviolet band sensitive to vibration); And check the current state of the liquid crystal filter to ensure that the target band combination can be quickly switched.

[0042] Finally, quickly switch to the target band combination through a liquid crystal tunable filter (LCTF): By applying a specific voltage, change the arrangement of liquid crystal molecules to form a transmission window for the target band (such as near-infrared + deep ultraviolet); complete the band switching within 10 microseconds (refer to the fast response characteristics of liquid crystal materials) to avoid image misalignment caused by the movement of the mask.

[0043] At the target band, the camera acquires high-resolution images with optimized exposure and gain parameters. For example: for metal residues, in the deep ultraviolet band, the edge contour contrast of metal contaminants is increased by 50%; for transparent film defects, in the extreme ultraviolet band, the interference fringes caused by the film thickness difference are clearly visible.

[0044] Obtain multi-dimensional characteristics such as reflection, absorption, transmission, and thermal radiation of defects through multi-spectral preliminary scanning. Combine the multi-band fusion analysis of the deep learning model to quickly identify candidate defect types (such as metal residues, transparent films, or phase distortions), and based on the pre-trained defect type - spectral mapping table and real-time environmental parameters (such as temperature, humidity, vibration), dynamically select the optimal band combination (such as metal residues activating the near-infrared + deep ultraviolet bands), complete the band switching within 10 microseconds through a liquid crystal tunable filter, and finally generate a multi-spectral image with high confidence. This process realizes fast preliminary screening, accurate classification, and high signal-to-noise ratio imaging, provides a reliable data basis for subsequent steps such as environmental compensation and multi-modal verification, and significantly improves the detection sensitivity of target defects (such as the detection accuracy of metal residues is increased by 30%) and the overall efficiency of the system.

[0045] S200, based on the temperature and humidity sensors and vibration sensors, determine the corresponding temperature change, humidity change, and vibration acceleration component in real time. Calculate the offsets of the optical path in the X-axis and Y-axis based on the temperature change and vibration acceleration component. The offsets are linearly related to the temperature change and vibration acceleration component; dynamically adjust the deflection angle of the micromirror according to the offsets and the focal length of the optical system to compensate for the imaging offset in real time; after compensation, acquire images twice through the multi-spectral camera to obtain the corresponding secondary multi-spectral images.

[0046] In the embodiment of the present application, analyze the temperature and humidity changes and vibration frequency in real time through a multi-sensor fusion algorithm to construct a dynamic interference model. Among them, the optical path offset caused by temperature change is corrected by a non-linear compensation curve, and the offset caused by vibration is compensated after separating the high-frequency noise and low-frequency trend through an adaptive filter; the deflection angle of the micromirror is dynamically adjusted through a step-by-step iterative algorithm, and verification images are acquired and the deflection parameters are optimized and fed back after each iteration.

[0047] Among them, 1. Determine the temperature, humidity and vibration parameters in real time. The specific implementation process includes the following steps: Through the integrated temperature and humidity sensor (such as a high-precision capacitive sensor) and a triaxial accelerometer (such as a MEMS accelerometer), monitor the changes in temperature, humidity and vibration acceleration components in the environment in real time. For example, sample 100 times per second to ensure that minute environmental fluctuations are captured. Then input the collected data into a multi-sensor fusion algorithm for synchronization and preprocessing; for example, calculate the temperature change ΔT (unit: °C), humidity change ΔRH (unit: %RH) and vibration accelerations aX, aY, aZ on the X, Y, and Z axes (unit: g). These data are used to construct a dynamic interference model subsequently.

[0048] Through high-frequency sampling, accurately capture the subtle changes in temperature, humidity and vibration in the environment, providing accurate basic data for subsequent optical path offset calculation; and use a multi-sensor fusion algorithm to ensure that data from different sources are aligned at the same timestamp, reducing errors caused by sensor delay or desynchronization.

[0049] 2. Calculate the optical path offset and establish a dynamic interference model. The specific implementation process includes the following steps: Based on the temperature change ΔT and vibration acceleration components (mainly considering the X and Y axes), calculate the offsets ΔX and ΔY of the optical path on the X and Y axes according to the pre-calibrated linear relationship; for example, for temperature change, assume ΔX = k1 * ΔT, ΔY = k2 * ΔT (k1, k2 are calibration coefficients); for vibration, ΔX = k3 * aX, ΔY = k4 * aY.

[0050] Temperature compensation: The optical path offset caused by temperature change is corrected using a non-linear compensation curve to more accurately reflect the actual offset. For example, a non-linear compensation curve is fitted through experimental data to replace the simple linear model. Vibration compensation: Use an adaptive filter to separate high-frequency noise from low-frequency noise, and only compensate for low-frequency vibrations with greater influence to avoid overreacting to high-frequency noise.

[0051] 3. Dynamically adjust the deflection angles of the micromirrors. The specific implementation process includes the following steps: According to the calculated optical path offsets ΔX and ΔY and the focal length f of the optical system, calculate the deflection angles θX and θY of the micromirrors that need to be adjusted. The formula is θX = arctan(ΔX / f), θY = arctan(ΔY / f). Use a step-by-step iterative algorithm to dynamically adjust the deflection angles of the micromirrors; after each adjustment, collect and verify the images and compare them with the ideal position, and feedback the results to optimize the deflection parameters for the next adjustment. For example, initially set θX = 0.5°. If the detection result shows that there is still a slight offset, then gradually increase or decrease θX until the best alignment state is reached.

[0052] Through precise calculation and iterative optimization, it is ensured that the deflection angle of the micromirror can effectively compensate for the optical path deviation and restore the imaging quality. And by utilizing the verification image feedback mechanism, a closed-loop control system is formed to continuously optimize the deflection parameters and ensure stability during long-term operation.

[0053] 4. Secondary multispectral image acquisition. The specific execution process includes the following steps: After completing the micromirror deflection angle adjustment, use the multispectral camera for secondary image acquisition. At this time, the camera parameters (such as exposure time, gain, etc.) should be appropriately adjusted according to the current environmental conditions and target band characteristics to obtain high-quality multispectral images; after the acquisition is completed, the secondary image is evaluated for quality (such as clarity, contrast, etc.) to ensure that the compensation effect meets expectations. If slight offset or other quality problems are found, the micromirror deflection angle can be adjusted again or the image can be re-acquired.

[0054] By real-time monitoring of temperature, humidity and vibration parameters, the optical path offset on the X-axis and Y-axis caused by these environmental factors is accurately calculated, and nonlinear compensation curves and adaptive filtering techniques are used to correct the offset caused by temperature changes and vibrations, respectively. Then, the micromirror deflection angle is dynamically adjusted according to the calculation results to compensate for the imaging offset. The adjustment process is optimized through a step-by-step iterative algorithm, and verification image feedback optimization is collected after each adjustment to ultimately obtain high-quality secondary multispectral images. This process effectively eliminates the impact of environmental interference on imaging quality, ensures high precision and reliability of detection, and significantly improves the accuracy of integrated circuit mask defect detection.

[0055] S300, through the field programmable gate array, performs filtering on the secondary multispectral image and dynamically adjusts the filtering neighborhood radius, completes multi-band image registration in the FPGA, generates a fused high-contrast image, and extracts the coordinates of the candidate defect area; loads the target detection model in the graphics processor, the target detection model integrates the spectral attention mechanism, locates the coordinates of the candidate defect area, outputs the defect bounding box coordinates and the defect confidence, and if the defect confidence is detected to be lower than the preset threshold, triggers the band reselection mechanism and switches to the supplementary band combination for re-inspection.

[0056] In the embodiment of the present application, the band reselection mechanism and method further include:

[0057] When the defect confidence is lower than the threshold, the historical defect database is called to match the current defect characteristics. If the match is successful, the historical optimal band combination is directly activated;

[0058] If the match is not successful, the backup band combination closest to the current defect spectral response is selected through the spectrum-feature association map. The spectrum-feature association map is generated based on a pre-trained deep learning model and contains a multi-dimensional similarity matrix between defect types and band combinations.

[0059] Among them, 1. Filtering and multi-band image registration in FPGA. The specific execution process includes the following steps: Perform filtering processing on the secondary multi-spectral image through a Field Programmable Gate Array (FPGA), and dynamically adjust the filtering neighborhood radius to adapt to different noise levels. For example, expand the neighborhood radius from 3×3 to 5×5 according to the noise intensity to remove random noise in the image.

[0060] Complete the precise registration of multi-band images in the FPGA to ensure the spatial consistency of each band image; adopt a feature point matching algorithm (such as SIFT or SURF) to extract feature points, and use the optical flow method to calculate the sub-pixel displacement field between adjacent bands for global consistency correction to generate a fused high-contrast image. In addition, for the extraction of candidate defect regions, extract the coordinates of candidate defect regions from the fused image to provide accurate position information for subsequent object detection.

[0061] Improve the clarity and contrast of the image and enhance the visibility of subtle defect features through adaptive filtering and precise registration; and utilize the parallel processing ability of the FPGA to quickly complete the filtering and registration tasks to ensure real-time performance.

[0062] 2. Object detection and localization in GPU. The specific execution process includes the following steps: Load an object detection model (such as YOLOv5) with an integrated spectral attention mechanism on a Graphics Processing Unit (GPU). The object detection model can process multi-band features simultaneously. Input the coordinates of the candidate defect region into the model for precise defect localization; the object detection model outputs the coordinates of the defect bounding box and the defect confidence. If the detected defect confidence is lower than a preset threshold (such as 80%), then trigger the band re-selection mechanism and switch to the supplementary band combination for re-inspection.

[0063] Achieve precise localization of the candidate defect region by leveraging the powerful feature extraction ability of the deep learning model; in addition, by setting the confidence threshold, automatically identify low-confidence defects and initiate the re-inspection process to reduce false alarms and missed detections.

[0064] 3. Band re-selection mechanism. The specific execution process includes the following steps: When the defect confidence is lower than the threshold, first call the historical defect database and try to match the current defect features. If the match is successful, directly activate the band combination that performed best for this type of defect historically. If no match is found in the historical database, select the backup band combination that is closest to the current defect spectral response through the spectral-feature association map generated by a pre-trained deep learning model. This map contains a multi-dimensional similarity matrix between defect types and band combinations to help the system make the best choice.

[0065] By matching through the historical database or selecting from the spectral-feature correlation map, the most suitable band combination for the current defect type is ensured, improving the detection accuracy. In addition, whether it is a known or new type of defect, the best detection solution can be found through the intelligent band selection mechanism, enhancing the flexibility of the system and its ability to handle complex situations.

[0066] By performing filtering and multi-band image registration on the secondary multi-spectral image in the FPGA, a high-contrast image is generated and the coordinates of the candidate defect regions are extracted. Then, the target detection model is loaded in the GPU to perform precise defect localization and confidence evaluation on the candidate regions. For defects with low confidence, a band re-selection mechanism is adopted. Through matching with the historical defect database or selecting from the spectral-feature correlation map, a suitable alternative band combination is selected for re-inspection. This process not only improves the image quality and detection accuracy, but also enhances the flexibility and adaptability of the system through the intelligent band selection mechanism, ensuring the efficiency and accuracy of the integrated circuit mask defect detection.

[0067] S400, update the defect type - spectrum mapping table based on the real-time detection results. If there is a band combination with a continuous detection rate exceeding 90% among the same type of defects, increase its corresponding priority; if there is a band combination with a false alarm rate exceeding 5%, reduce its corresponding weight or temporarily disable it; when the environmental fluctuation exceeds the limit, automatically switch to the anti-interference mode, turn off the bands sensitive to environmental impacts, and trigger the full-scan calibration process of the MEMS mirror.

[0068] Among them, the following steps are included:

[0069] 1. Update the defect type - spectrum mapping table in real time. The specific execution process includes the following steps: After each detection, the system records the real-time detection results, including the defect type, position coordinates, confidence level, environmental parameters (such as temperature, humidity) and offset. Based on these data, the system evaluates the performance of each band combination; if a certain band combination has a continuous detection rate exceeding 90% among the same type of defects, the priority of this band combination will be increased. For example, if the detection rate of the near-infrared and deep-ultraviolet band combination for metal residues reaches 92%, this combination will be preferentially selected in future detections of metal residues. When the false alarm rate of a certain band combination exceeds 5%, the system will automatically reduce its weight or temporarily disable this combination; for example, if the false alarm rate of the visible light plus ultraviolet band combination for transparent film defects is 6%, this combination may be marked as "to be verified status" and its usage frequency will be reduced until further verification.

[0070] By dynamically adjusting the priority and weight of the band combination, it is ensured that the system always uses the most effective band combination for detection, thereby improving the detection efficiency and accuracy; by promptly identifying and processing the band combinations with high false alarm rates, the false alarm probability in future detections is reduced, enhancing the reliability of the system.

[0071] 2. Automatic switching mechanism when environmental fluctuations exceed the limit. The specific execution process includes the following steps: The system continuously monitors environmental parameters (such as data provided by temperature and humidity sensors and vibration sensors). Once it is found that the environmental fluctuations exceed the preset threshold (such as the temperature change is greater than 2°C or the vibration acceleration exceeds 0.5g), the system automatically switches to the anti-interference mode. In the anti-interference mode, the system will automatically turn off those bands that are particularly sensitive to environmental changes (such as the ultraviolet band is vulnerable to vibration), to avoid detection errors caused by environmental factors. And trigger the MEMS mirror to perform a full scan calibration process, recalculate and adjust the mirror deflection angle, compensate for the optical path offset caused by environmental changes, and ensure that the imaging quality is not affected.

[0072] By automatically switching to the anti-interference mode, the influence of environmental fluctuations on the detection results is reduced, ensuring that the system can still operate stably in a complex environment. And the full scan calibration of the MEMS mirror can effectively compensate for the optical path offset caused by environmental changes, guarantee the quality and accuracy of image acquisition, and thus maintain a high level of detection accuracy.

[0073] Based on the real-time detection results, dynamically update the defect type - spectrum mapping table. By increasing the priority of the efficient band combination and reducing the weight of the high false alarm rate band combination, continuously optimize the detection strategy to ensure that the system always uses the optimal configuration for defect detection. In addition, when the environmental fluctuations exceed the threshold, automatically switch to the anti-interference mode, turn off the bands sensitive to the environment, and through the full scan calibration process of the MEMS mirror, effectively compensate for the influence brought by environmental changes. This process not only enhances the system's self-learning and adaptability, but also significantly reduces the risks of false alarms and missed detections, providing a solid guarantee for the high-quality detection of integrated circuit mask plates. In this way, S400 significantly improves the stability and detection accuracy of the system in a complex environment.

[0074] S500, real-time monitor the performance degradation indicators corresponding to the optical components. The performance degradation indicators include the decrease in the transmittance of the filter and the degree of lens surface contamination. When any indicator exceeds the preset threshold, automatically switch to the redundant optical path and trigger the cleaning program; and integrate the energy spectrum analysis module to verify the elemental composition of the defect area. If there is a conflict between the optical detection result and the energy spectrum data, trigger the manual re-inspection process; generate a visual detection report, marking the defect location, type, confidence level and elemental composition.

[0075] Among them, 1. Real-time monitoring of the performance degradation indicators of optical components. The specific implementation process includes the following steps: The system monitors the key performance degradation indicators of optical components in real time, including the decrease in the transmittance of the filter and the degree of lens surface contamination. For example, use built-in sensors or measure whether the transmittance of the filter is lower than the preset threshold (such as 90%) through regular calibration, or detect whether the degree of lens surface contamination exceeds the set standard (such as 20% occlusion). Once any performance degradation indicator exceeds the preset threshold, the system will automatically switch to the standby redundant optical path to ensure the continuity of the detection work without interrupting the production process. At the same time as switching to the redundant optical path, the system will trigger the corresponding cleaning procedure to clean the contaminated optical components and restore their normal working state.

[0076] By real-time monitoring of key performance indicators and taking timely measures, it is possible to prevent detection errors caused by the performance degradation of optical components and ensure the accuracy of detection results. Automatically switching to the redundant optical path can handle component degradation issues without affecting the detection efficiency, while the cleaning procedure helps to extend the service life of optical components.

[0077] 2. Integrate the energy spectrum analysis module for element composition verification. The specific implementation process includes the following steps: After the optical detection identifies potential defects, the integrated energy spectrum analysis module (such as EDS) will scan the element composition of the same area to verify whether specific elements (such as iron, copper, etc. in metal residues) exist in this area; if there is a conflict between the optical detection result and the element distribution obtained from the energy spectrum analysis (for example, the optical detection determines metal residues but EDS does not detect any metal elements), the system will trigger the manual re-inspection process, and professional personnel will further confirm the true situation of the defects.

[0078] By combining the results of optical detection and energy spectrum analysis, it is possible to more accurately judge the type of defects and reduce the false alarm rate; when the two detection methods are inconsistent, manual re-inspection is introduced as the final confirmation means to ensure the reliability of the detection results.

[0079] 3. Generate a visual inspection report. The specific implementation process includes the following steps: Report generation, the system automatically generates a detailed visual inspection report, which contains the following information: Defect location, accurately marking the specific position coordinates of each detected defect; Defect type, the type of defect determined according to optical detection and energy spectrum analysis (such as metal residues, transparent film defects, etc.); Confidence level, providing a confidence score for each defect detection result to help users evaluate the credibility of the results; Element composition, listing the main element compositions in the defect area confirmed by energy spectrum analysis; Report output, these reports can be distributed to relevant personnel in electronic form for subsequent analysis and decision support.

[0080] Visual reports intuitively display the detection results and their related details, facilitating quick understanding and communication. By providing detailed data support, it helps management make fact-based decisions, optimizing production processes and quality management procedures.

[0081] By continuously monitoring the performance degradation indicators of optical components, when a performance decline is detected, it automatically switches to a redundant optical path and initiates a cleaning process, ensuring the continuous and efficient operation of the detection system. Meanwhile, an energy spectrum analysis module is integrated to verify the elemental composition of the defective area, and when there is a conflict between the detection results and the energy spectrum data, it triggers a manual re-inspection process, enhancing the accuracy and reliability of the detection results. Finally, a visual detection report containing the defect location, type, confidence level, and elemental composition is generated, providing strong support for subsequent analysis and decision-making. This series of measures not only improves the stability and accuracy of the detection system but also significantly reduces the risks of false alarms and missed detections, providing comprehensive protection for the quality control of integrated circuit masks.

[0082] In the embodiment of this application, when the energy spectrum analysis module is integrated with the multi-spectral detection system, a timing synchronization trigger mechanism is adopted. The method further includes:

[0083] When the optical detection determines a high-confidence defect, the energy spectrum analysis module scans the elemental composition of the same area. Among them, when the multi-spectral detection system determines a certain area as a high-confidence defect (for example, the confidence level exceeds 90%), the energy spectrum analysis module (such as EDS) is immediately started through the timing synchronization trigger mechanism; this mechanism ensures that the optical detection and the elemental composition scanning can analyze the same area in the shortest time, reducing sample changes or environmental impacts caused by time differences. By immediately starting the energy spectrum analysis, the optical detection results can be verified in a timely manner, improving the detection efficiency and accuracy.

[0084] If there is a conflict between the defect type detected by the optical detection and the elemental distribution detected by the energy spectrum analysis module, a three-dimensional topography scan is initiated, and the reflectivity deviation of the optical image is corrected through the topography data, and finally a fusion verification report is generated.

[0085] Among them, once the optical detection identifies a potential high-confidence defect, the energy spectrum analysis module will conduct a detailed elemental composition scan of this area; for example, if the optical detection believes that there is metal residue in this area, the EDS will check whether this area actually contains metal elements.

[0086] If the defect type detected by optical inspection (such as metal residue) is inconsistent with the element distribution detected by the energy spectrum analysis module (for example, no metal elements are detected by EDS), an additional three-dimensional topography scanning program is initiated. Using a white light interferometer or other three-dimensional topography measurement equipment, a detailed surface topography scan is performed on the suspected defect area. By analyzing the minute undulations and structural features of the surface, the reflectivity deviation of the optical image caused by surface topography differences can be corrected.

[0087] Based on the optical image, energy spectrum analysis results, and three-dimensional topography data, the system automatically generates a fusion verification report. This report not only includes the original optical inspection results and energy spectrum analysis results but also incorporates the surface structure information provided by the three-dimensional topography data to help comprehensively understand the nature of the defect.

[0088] The integration of the energy spectrum analysis module and the multispectral detection system adopts a timing synchronization trigger mechanism to ensure that when the optical inspection determines a high-confidence defect, the energy spectrum analysis module can immediately perform an elemental composition scan of the same area. When there is a conflict between the optical inspection results and the energy spectrum analysis, a three-dimensional topography scan is initiated. The reflectivity deviation of the optical image is corrected using the surface topography data, and a fusion verification report is generated. This process not only improves the immediacy and accuracy of detection but also solves the misjudgment problem that may be caused by a single detection method through a multi-level data verification mechanism, providing more accurate and comprehensive support for the quality control of integrated circuit mask plates.

[0089] In the embodiment of the present application, the method further includes:

[0090] S511, when it is detected that the performance of the optical component degrades, the detection accuracy loss after switching to the redundant optical path is predicted through a fault impact assessment model. If the loss is lower than a preset threshold, the switch is automatically made.

[0091] Among them, when it is monitored that the performance degradation index of the optical component (such as the decrease in the transmittance of the filter or the degree of lens surface contamination) exceeds the preset threshold, the system first predicts the possible detection accuracy loss after switching to the redundant optical path through a fault impact assessment model. If the prediction result shows that the detection accuracy loss after switching to the redundant optical path is lower than the preset threshold (for example, the accuracy loss does not exceed 5%), the system will automatically switch to the redundant optical path to continue the detection. By predicting the detection accuracy loss after switching to the redundant optical path, the system can make a more informed decision, avoid unnecessary downtime for maintenance, and ensure the accuracy of the detection results.

[0092] S512, if the loss exceeds the threshold, a multi-band collaborative detection mode is triggered. The detection is maintained by enhancing the weight of the non-degraded band and introducing an adaptive noise reduction algorithm, and the degradation rate of the degraded component is recorded to predict the maintenance cycle.

[0093] Among them, if the predicted result shows that the detection accuracy loss after switching to the redundant optical path exceeds the preset threshold, the system will not immediately switch to the redundant optical path, but start the multi-band collaborative detection mode. In this mode: enhance the weight of the non-degraded band, the system will increase the weight of those bands that have not been affected to compensate for the lack of function of the affected bands. For example, if the ultraviolet band is greatly affected by pollution, the usage frequency and importance of the near-infrared band will be increased; introduce an adaptive noise reduction algorithm. In order to further improve the detection quality, during the multi-band collaborative detection process, the system will apply an adaptive noise reduction algorithm to dynamically adjust the filtering parameters, reduce noise interference, and improve image clarity and contrast.

[0094] The system will also record the degradation rate of the degraded components in real time to facilitate the prediction of future maintenance cycles. For example, by analyzing the change trend of the filter transmittance over time, predict when it will reach an unacceptable level and arrange the maintenance plan in advance.

[0095] When the performance degradation of the optical component is detected, the system predicts the detection accuracy loss after switching to the redundant optical path through the fault impact assessment model and decides whether to automatically switch according to the loss situation. If the loss exceeds the threshold, the multi-band collaborative detection mode is triggered, and the detection quality is maintained by enhancing the weight of the non-degraded band and introducing an adaptive noise reduction algorithm. In addition, the system also records the degradation rate of the degraded components to predict the maintenance cycle. This series of measures not only improves the self-regulation ability of the system, ensures a high level of detection accuracy even in the case of performance degradation, but also provides a scientific basis for subsequent maintenance work, thus improving the reliability and efficiency of the entire system.

[0096] It should be noted here that during the dynamic update of the defect type - spectrum mapping table, the method also includes:

[0097] S410, establish a performance evaluation system for band combinations, and generate corresponding multi-dimensional performance index vectors through the defect detection rate, environmental interference intensity, false alarm / missed alarm times, and system resource occupancy rate collected in real time.

[0098] Among them, the generation of the multi-dimensional performance index vector: establish a dynamically updated performance evaluation system for each band combination. This system generates corresponding multi-dimensional performance index vectors by collecting the following data in real time: the defect detection rate, the success rate of a specific band combination in detecting a certain type of defect; the environmental interference intensity, the sensitivity of the band combination to environmental changes (such as temperature, humidity, vibration, etc.); the false alarm / missed alarm times, the number of false alarms and missed alarms generated when using this band combination; the system resource occupancy rate, the occupancy of the system computing resources by the band combination during use.

[0099] S420 uses an adaptive weight adjustment algorithm to perform dynamic priority sorting on band combinations based on a multi-dimensional performance index vector. The weight adjustment rules include:

[0100] If the detection confidence of a certain band combination exceeds 90% continuously among similar defects, its priority will be increased, and performance migration with other inefficient bands will be triggered. This means that this efficient band combination will be more applied to subsequent detection tasks, and at the same time, its successful strategy will be tried to be applied to other poorly performing bands to improve their performance.

[0101] For those band combinations with a false alarm rate exceeding the preset threshold and being particularly sensitive to environmental interference, the system will reduce their weights or mark them as "to be verified status". In this case, the system will also enable the collaborative detection mode of alternative band combinations to reduce the errors caused by these bands.

[0102] When a band combination marked as "to be verified status" continuously shows false alarms, the system will start the knowledge transfer process across defect types. This means that the system will extract data from the successful cases of this band combination in the detection of other defect types and inject it into the training set of the current defect type. The purpose of this is to enrich the training samples of the current defect type and may help identify and correct the reasons for the failure of the current band combination under specific conditions.

[0103] By establishing a performance evaluation system for band combinations, collecting and analyzing multi-dimensional performance index vectors (including defect detection rate, environmental interference intensity, false alarm / missed alarm times, and system resource occupancy rate) in real time, and using an adaptive weight adjustment algorithm to perform dynamic priority sorting on band combinations. The rules include increasing the priority of efficient band combinations and triggering performance migration, as well as reducing the weights of band combinations with high false alarm rates and sensitive to environmental interference or marking them as to be verified status, and enabling the collaborative detection mode of alternative band combinations. When a band combination marked as to be verified status continuously has false alarms, start the knowledge transfer process across defect types and inject the data of the corresponding band combination in the detection of other defect types into the training set of the current defect type. This series of measures not only improves the self-learning and adaptation ability of the system, but also significantly enhances the flexibility and accuracy of the detection strategy.

[0104] In addition, the method also includes:

[0105] S310 extracts the semantic feature maps of each band image through a deep learning model and constructs a cross-band feature association map.

[0106] Among them, a deep learning model (such as a convolutional neural network) is used to extract semantic feature maps from the images of each band. These feature maps capture the unique information of the images in different bands. For example, the surface contamination features in the ultraviolet band or the material thickness differences in the near-infrared band; based on the extracted semantic feature maps, a cross-band feature correlation map is constructed. This map shows the correlation and interaction of features between different bands, which helps to understand how each band jointly describes the same physical phenomenon.

[0107] By deeply analyzing the semantic features of each band and establishing the associations between them, the characteristics of the target object can be more comprehensively understood, providing a basis for subsequent precise registration and fusion. And the feature correlation map can help the system better identify and match the same feature points in different bands, thus improving the effect of global consistency correction.

[0108] S320, the optical flow method is used to calculate the sub-pixel displacement field of adjacent band images, and global consistency correction is performed in combination with the feature correlation map.

[0109] Among them, the optical flow method is used to calculate the sub-pixel displacement field between adjacent band images. This method can accurately locate the movement in the image below the pixel level, ensuring that even small changes can be captured; using the previously constructed cross-band feature correlation map, global consistency correction is performed on the calculated displacement field. This step ensures the spatial consistency of all band images and avoids offset problems caused by mechanical vibration or other factors.

[0110] Through sub-pixel displacement field calculation and global consistency correction, highly accurate registration of multi-band images can be achieved, significantly improving the quality of the fused image; in addition, an effective correction mechanism reduces image distortion that may be caused by cumulative errors, ensuring the authenticity and accuracy of the final fused image.

[0111] S330, in the fusion stage, the signal-to-noise ratio of each band is dynamically weighted through an attention mechanism, and the weight value is negatively correlated with the current environmental interference intensity of the band to generate the corresponding fused image.

[0112] Among them, in the fusion stage, an attention mechanism is introduced to dynamically weight the signal-to-noise ratio of each band, and the weight value is negatively correlated with the current environmental interference intensity of the band, that is, the stronger the environmental interference of a band, the lower its weight; for example, in the case of large vibrations, the weight of the ultraviolet band sensitive to vibrations is reduced. According to the dynamically adjusted weight values, the band images are fused into a high-quality fused image.

[0113] By means of the attention mechanism, the weights of each band are dynamically adjusted, enabling the adaptive optimization of the fusion result according to real-time environmental conditions. This ensures that the fused image retains key information while minimizing noise interference as much as possible. This mechanism allows the system to flexibly adapt to different environmental conditions and generate high-quality fused images whether in an ideal environment or in the face of complex on-site conditions, improving the adaptability and robustness of the system.

[0114] Semantic feature maps of each band image are extracted through a deep learning model, and a cross-band feature correlation map is constructed. The optical flow method is used to calculate the sub-pixel displacement field of adjacent band images and global consistency correction is performed in combination with the feature correlation map. In the fusion stage, the signal-to-noise ratio of each band is dynamically weighted by the attention mechanism, and the weight value is negatively correlated with the current environmental interference intensity of the band, thereby generating the corresponding fused image. This series of measures enhances feature understanding and matching accuracy, achieves high-precision multi-band image registration, and optimizes the quality of the fused image through the dynamic weighting mechanism, ensuring clear and accurate detection results under various environmental conditions.

[0115] Suppose we are conducting quality inspections on a batch of newly produced integrated circuit mask plates, with a focus on metal residue defects. Such defects can seriously affect the function and performance of the chip, so high-precision detection methods are required to ensure no metal residue.

[0116] Step S100: First, the mask plate is preliminarily scanned using a multi-spectral camera to obtain initial images containing multiple bands such as visible light, near-infrared, and ultraviolet. By analyzing these images through a deep learning model, the system identifies problem areas where metal residue may exist, and based on the pre-trained "defect type - spectral mapping table", the near-infrared band (for detecting metal thickness) and the deep ultraviolet band (for identifying metal surface contamination) are selected as the optimal band combination for further analysis.

[0117] Step S200: After determining the target bands, considering that temperature and vibration changes in the production workshop may affect the imaging quality, the system monitors the temperature, humidity, and vibration parameters in real time and calculates the corresponding optical path offset. Based on this data, the system dynamically adjusts the deflection angle of the micromirror to compensate for imaging offset and performs secondary multi-spectral image acquisition. The result of this is high-quality, non-offset secondary multi-spectral images, providing a basis for subsequent precise analysis.

[0118] Step S300: Next, the FPGA performed adaptive filtering on the secondary multi-spectral image and completed the precise registration of the multi-band images, generating a fused high-contrast image. The coordinates of the candidate defect regions were extracted from this image, and an object detection model integrating a spectral attention mechanism was loaded on the GPU to precisely locate these regions. For the detected metal residue defects, the system output the detailed bounding box coordinates and confidence scores.

[0119] Step S400: Based on the results of this detection, the system evaluated the performance of the selected band combination. Since the near-infrared plus deep-ultraviolet band combination showed excellent performance in detecting metal residues, with a continuous detection rate reaching 95%, the system automatically increased the priority of this combination in future similar detection tasks. At the same time, any false alarm or missed detection cases were also recorded for future optimization.

[0120] Step S500: Finally, after the optical inspection confirmed the metal residue defects, the system immediately triggered the energy spectrum analysis module to verify the elemental composition of the same region, ensuring the accuracy of the detection results. Combining the data of the optical image and the energy spectrum analysis, the system generated a detailed visual inspection report, including information such as the defect location, type, confidence level, and elemental composition.

[0121] Through the above process, this solution not only improves the detection accuracy of key defects such as metal residues, but also enhances the self-learning and adaptation capabilities of the system. Especially when encountering complex environmental conditions, it can ensure the reliability of the detection results through intelligent adjustment and compensation measures. Ultimately, this helps to ensure that each integrated circuit mask plate meets the highest quality standards, thereby improving the product quality and efficiency of the entire production line.

[0122] An embodiment of this application discloses a multi-spectral collaborative detection system for integrated circuit mask plate defects. Referring to Figure 2 , it includes:

[0123] A candidate defect type recognition module 001, which preliminarily scans the mask plate through a multi-spectral camera to obtain the corresponding initial multi-band image, and is used to identify the corresponding candidate defect types based on the initial multi-band image. The candidate defect types include metal residues, transparent thin film defects, and phase distortions; dynamically selects the optimal band combination from a pre-trained defect type - spectrum mapping table according to the candidate defect type. If the candidate defect type is a metal residue defect, the near-infrared band and the deep-ultraviolet band are activated. If the candidate defect type is a transparent thin film defect, the extreme-ultraviolet band and the visible light band are activated. If the candidate defect type is a phase distortion defect, the ultraviolet band and the mid-infrared band are activated; switches to the target band combination within 10 microseconds through a liquid crystal tunable filter to obtain the multi-spectral image corresponding to high confidence.

[0124] The offset calculation module 002 determines the corresponding temperature change amount, humidity change amount, and vibration acceleration component in real time based on the temperature and humidity sensor and the vibration sensor, and calculates the offsets of the optical path in the X-axis and Y-axis based on the temperature change amount and the vibration acceleration component. The offset has a linear relationship with the temperature change amount and the vibration acceleration component; dynamically adjusts the deflection angle of the micromirror according to the offset and the focal length of the optical system to compensate for the imaging offset in real time; after compensation, acquires the corresponding secondary multi-spectral image by secondary acquisition of the image with the multi-spectral camera;

[0125] The high-contrast image generation module 003 performs filtering on the secondary multi-spectral image and dynamically adjusts the filtering neighborhood radius through a field-programmable gate array, completes multi-band image registration in the FPGA, generates a fused high-contrast image, and extracts the coordinates of the candidate defect regions; loads the target detection model in the graphics processor, the target detection model integrates the spectral attention mechanism, locates the coordinates of the candidate defect regions, outputs the defect bounding box coordinates and the defect confidence level. If the detected defect confidence level is lower than the preset threshold, triggers the band re-selection mechanism and switches to the supplementary band combination for re-inspection;

[0126] The update module 004 is used to update the defect type - spectrum mapping table based on the real-time detection results. If there is a band combination with a continuous detection rate exceeding 90% in the same type of defect, increases the corresponding priority; if there is a band combination with a false alarm rate exceeding 5%, reduces the corresponding weight or temporarily disables it; when the environmental fluctuation exceeds the limit, automatically switches to the anti-interference mode, turns off the bands sensitive to environmental impacts, and triggers the full-scan calibration process of the MEMS micromirror. Among them, the real-time detection results include the defect type, position coordinates, confidence level, environmental parameters, and offset;

[0127] The performance degradation index monitoring module 005 is used to monitor the performance degradation index corresponding to the optical component in real time. The performance degradation index includes the decrease in the transmittance of the filter and the degree of lens surface contamination. When any index exceeds the preset threshold, automatically switches to the redundant optical path and triggers the cleaning program; and integrates the energy spectrum analysis module to verify the element composition of the defect region. If there is a conflict between the optical detection result and the energy spectrum data, triggers the manual re-inspection process; generates a visual detection report, marking the defect position, type, confidence level, and element composition.

[0128] An embodiment of the present application also discloses an integrated circuit mask defect multi-spectral collaborative detection system, including a processor, and a program of the integrated circuit mask defect multi-spectral collaborative detection method described in any one of the above is run in the processor.

[0129] An embodiment of the present application also discloses a storage medium, storing a program of the integrated circuit mask defect multi-spectral collaborative detection method described in any one of the above.

[0130] Although the embodiments of the present application have been shown and described above, it can be understood that the above embodiments are exemplary and should not be construed as limiting the present application. Those of ordinary skill in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of the present application.

Claims

1. A multi-spectral collaborative detection method for integrated circuit mask defects, characterized in that, Including: Preliminarily scan the reticle through a multispectral camera to obtain the corresponding initial multi-band image, identify the corresponding candidate defect types based on the initial multi-band image, and the candidate defect types include metal residue, transparent thin film defect, and phase distortion; dynamically select the optimal band combination from the pre-trained defect type - spectrum mapping table according to the candidate defect types. If the candidate defect type is a metal residue defect, activate the near-infrared band and the deep ultraviolet band. If the candidate defect type is a transparent thin film defect, activate the extreme ultraviolet band and the visible light band. If the candidate defect type is a phase distortion defect, activate the ultraviolet band and the mid-infrared band; switch to the target band combination within 10 microseconds through a liquid crystal tunable filter to obtain the multi-spectral image corresponding to high confidence; Based on the temperature and humidity sensors and vibration sensors, determine the corresponding temperature change, humidity change, and vibration acceleration component in real time. Calculate the offsets of the optical path in the X-axis and Y-axis based on the temperature change and the vibration acceleration component. The offsets are linearly related to the temperature change and the vibration acceleration component; dynamically adjust the deflection angle of the micromirror according to the offsets and the focal length of the optical system to compensate for the imaging offset in real time; after compensation, collect the image twice through the multispectral camera to obtain the corresponding secondary multi-spectral image; Execute filtering on the secondary multi-spectral image and dynamically adjust the filtering neighborhood radius through a field-programmable gate array. Complete the registration of the multi-band image in the FPGA to generate a fused high-contrast image, and extract the coordinates of the candidate defect area; load the target detection model in the graphics processing unit. The target detection model integrates a spectral attention mechanism to locate the coordinates of the candidate defect area and output the coordinates of the defect bounding box and the defect confidence. If the detected defect confidence is lower than the preset threshold, trigger the band re-selection mechanism and switch to the supplementary band combination for re-inspection; Update the defect type - spectrum mapping table based on the real-time detection results. If there is a band combination with a continuous detection rate exceeding 90% among the same type of defects, increase the corresponding priority; if there is a band combination with a false alarm rate exceeding 5%, reduce the corresponding weight or temporarily disable it; when the environmental fluctuation exceeds the limit, automatically switch to the anti-interference mode, turn off the bands sensitive to environmental impacts, and trigger the full-scan calibration process of the MEMS micromirror. Among them, the real-time detection results include defect type, position coordinates, confidence, environmental parameters, and offsets; Monitor the performance degradation indicators corresponding to the optical components in real time. The performance degradation indicators include the decrease in the transmittance of the filter and the degree of lens surface contamination. When any indicator exceeds the preset threshold, automatically switch to the redundant optical path and trigger the cleaning program; and integrate an energy spectrum analysis module to verify the elemental composition of the defect area. If there is a conflict between the optical detection result and the energy spectrum data, trigger the manual re-inspection process; generate a visual detection report, marking the defect position, type, confidence, and elemental composition.

2. The multi-spectral collaborative detection method for integrated circuit mask defects according to claim 1, wherein The method further includes: Analyze the temperature and humidity changes and vibration frequencies in real time through a multi-sensor fusion algorithm to construct a dynamic interference model. Among them, the optical path offset caused by the temperature change is corrected through a non-linear compensation curve, and the offset caused by vibration is compensated after separating the high-frequency noise and low-frequency trend through an adaptive filter; The deflection angle of the microscope is dynamically adjusted by a stepwise iterative algorithm. After each iteration, a verification image is collected and the deflection parameters are fed back for optimization.

3. The multi-spectral collaborative detection method for integrated circuit mask defects according to claim 1, characterized in that Regarding the band reselection mechanism, the method further includes: When the defect confidence level is lower than the threshold, the historical defect database is called to match the current defect characteristics. If the match is successful, the historical optimal band combination is directly activated. If the match is unsuccessful, a backup band combination that is closest to the current defect spectral response is selected through a spectral-feature correlation map. The spectral-feature correlation map is generated based on a pre-trained deep learning model and includes a multi-dimensional similarity matrix of defect types and band combinations.

4. The multi-spectral collaborative detection method for integrated circuit mask defects according to claim 1, wherein When the energy spectrum analysis module is integrated with the multi-spectral detection system, a timing synchronization trigger mechanism is adopted. The method further includes: When the optical detection determines a high-confidence defect, the energy spectrum analysis module performs an elemental composition scan of the same area. If there is a conflict between the defect type detected by the optical detection and the elemental distribution detected by the energy spectrum analysis module, three-dimensional topography scanning is initiated. The reflectivity deviation of the optical image is corrected through the topography data, and finally a fusion verification report is generated.

5. The multi-spectral collaborative detection method for integrated circuit mask defects according to claim 4, wherein The method further includes: When it is detected that the performance of the optical component degrades, the detection accuracy loss after switching to the redundant optical path is predicted through a fault impact assessment model. If the loss is lower than the preset threshold, it is automatically switched. If the loss exceeds the threshold, a multi-band collaborative detection mode is triggered. The detection is maintained by enhancing the weight of the non-degraded band and introducing an adaptive noise reduction algorithm. At the same time, the degradation rate of the degraded component is recorded to predict the maintenance period.

6. The multi-spectral collaborative detection method for integrated circuit mask defects according to claim 1, wherein During the dynamic update of the defect type - spectrum mapping table, the method further includes: Establish a band combination performance evaluation system, and generate corresponding multi-dimensional performance index vectors based on the defect detection rate, environmental interference intensity, false alarm / missed alarm times, and system resource occupancy rate collected in real time. An adaptive weight adjustment algorithm is adopted to perform dynamic priority sorting on the band combinations based on the multi-dimensional performance index vectors. Among them, the weight adjustment rules include: If there is a band combination whose detection confidence level continuously exceeds 90% in the same type of defect, the corresponding priority is increased and the performance migration with other low-efficiency bands is triggered. If there is a band combination whose false alarm rate exceeds the threshold and whose environmental interference sensitivity is higher than the average value, the corresponding weight is reduced or it is marked as a pending verification state, and a collaborative detection mode of an alternative band combination is enabled. When the band combination marked as a pending verification state continuously generates false alarms, a cross-defect type knowledge transfer process is initiated, and the data of the corresponding band combination in the detection of other defect types is injected into the training set of the current defect type.

7. The multi-spectral collaborative detection method for integrated circuit mask defects according to claim 1, wherein During the multi-band image registration process, the method further includes: Extract the semantic feature maps of each band image through a deep learning model to construct a cross-band feature correlation map. Use the optical flow method to calculate the sub-pixel displacement field of adjacent band images, and perform global consistency correction in combination with the feature correlation map. In the fusion stage, the signal-to-noise ratio of each band is dynamically weighted through an attention mechanism. The weight value is negatively correlated with the current environmental interference intensity of the band, and the corresponding fusion image is generated.

8. An integrated circuit mask defect multispectral collaborative detection system, characterized in that, Including: Candidate defect type identification module: Initially scan the reticle through a multispectral camera to obtain the corresponding initial multi-band image, and identify the corresponding candidate defect types based on the initial multi-band image. The candidate defect types include metal residue, transparent thin film defects, and phase distortion; Dynamically select the optimal band combination from the pre-trained defect type - spectrum mapping table according to the candidate defect type. If the candidate defect type is a metal residue defect, activate the near-infrared band and the deep ultraviolet band. If the candidate defect type is a transparent thin film defect, activate the extreme ultraviolet band and the visible light band. If the candidate defect type is a phase distortion defect, activate the ultraviolet band and the mid-infrared band; Switch to the target band combination within 10 microseconds through a liquid crystal tunable filter to obtain the multi-spectral image corresponding to high confidence; Offset calculation module: Based on the temperature and humidity sensor and the vibration sensor, determine the corresponding temperature change, humidity change, and vibration acceleration component in real time. Calculate the offsets of the optical path in the X-axis and Y-axis based on the temperature change and the vibration acceleration component. The offsets are linearly related to the temperature change and the vibration acceleration component; Dynamically adjust the deflection angle of the micromirror according to the offsets and the focal length of the optical system to compensate for the imaging offset in real time; After compensation, collect images twice through the multispectral camera to obtain the corresponding secondary multi-spectral image; High-contrast image generation module: Perform filtering on the secondary multi-spectral image through a field-programmable gate array and dynamically adjust the filtering neighborhood radius. Complete the registration of the multi-band image in the FPGA to generate a fused high-contrast image and extract the coordinates of the candidate defect area; Load the target detection model in the graphics processor. The target detection model integrates a spectral attention mechanism to locate the coordinates of the candidate defect area, output the defect bounding box coordinates and the defect confidence. If the detected defect confidence is lower than the preset threshold, trigger the band re-selection mechanism and switch to the supplementary band combination for re-inspection; Update module: Update the defect type - spectrum mapping table based on the real-time detection results. If there is a band combination with a continuous detection rate exceeding 90% among the same type of defects, increase the corresponding priority; If there is a band combination with a false alarm rate exceeding 5%, reduce the corresponding weight or temporarily disable it; When the environmental fluctuation exceeds the limit, automatically switch to the anti-interference mode, turn off the bands sensitive to environmental impacts, and trigger the full-scan calibration process of the MEMS micromirror. Among them, the real-time detection results include defect type, position coordinates, confidence, environmental parameters, and offsets; Performance degradation index monitoring module: Used to monitor the performance degradation index corresponding to the optical components in real time. The performance degradation index includes the decrease in the transmittance of the filter and the degree of lens surface contamination. When any index exceeds the preset threshold, automatically switch to the redundant optical path and trigger the cleaning procedure; And integrate an energy spectrum analysis module to verify the elemental composition of the defect area. If there is a conflict between the optical detection result and the energy spectrum data, trigger the manual re-inspection process; Generate a visual inspection report, marking the defect position, type, confidence, and elemental composition.

9. A multi-spectral collaborative detection method for integrated circuit mask defects, characterized in that, It includes a processor, and a program of the multi-spectral collaborative detection method for integrated circuit mask defects as described in any one of claims 1-7 runs in the processor.

10. A storage medium, characterized in that, It stores a program of the multi-spectral collaborative detection method for integrated circuit mask defects as described in any one of claims 1-7.

Images