Particle defect detection system and method based on algorithm fusion and adaptive adjustment

Abstract

The invention discloses a particle defect detection system and method based on algorithm fusion and adaptive adjustment, and relates to the technical field of semiconductor manufacturing detection, the particle defect detection system comprises an image acquisition module, an algorithm fusion processing module, an adaptive adjustment module and a defect identification module, preprocessing the original data; carrying out deep mining on the obtained low-resolution image features, and outputting a high-resolution image after algorithm fusion; environment data are received, and various parameters of the system are adjusted in a self-adaptive mode according to the collected data; performing particle defect identification and classification on the high-resolution image in combination with the parameters subjected to self-adaptive adjustment; the technical problems of insufficient particle defect detection precision and poor environmental adaptability in the prior art are solved, and it is ensured that reconstructed image details can be clear under different environmental conditions.

Description

Technical Field

[0001] This invention relates to the field of semiconductor manufacturing inspection technology, specifically to a particle defect detection system and method based on algorithm fusion and adaptive adjustment, for detecting particle defects on the surface of micro-devices such as semiconductor wafers and electronic components. Background Technology

[0002] In semiconductor manufacturing, particle defect detection is a critical step in ensuring product quality. Existing particle defect detection systems typically employ a single super-resolution algorithm to process the detected images, such as traditional interpolation algorithms or single deep learning algorithms. This makes it difficult to achieve high-precision image reconstruction across different inspection objects and complex environments. Furthermore, current technologies lack mechanisms to automatically optimize algorithm parameters based on the material, shape, and ambient lighting of the inspection object. This leads to the loss of image details and limited detection accuracy in highly reflective materials, complex-shaped components, or environments with fluctuating lighting. For example, when inspecting highly reflective wafer surfaces, existing algorithms cannot effectively handle image blurring caused by reflections, resulting in missed detection of minute particle defects. Summary of the Invention

[0003] To address the technical problems of insufficient accuracy and poor environmental adaptability in particle defect detection in existing technologies, this invention provides a particle defect detection system and method based on algorithm fusion and adaptive adjustment, achieving high-precision particle defect detection in complex scenarios.

[0004] On the one hand, the particle defect detection system based on algorithm fusion and adaptive adjustment includes an image acquisition module for acquiring low-resolution images of the object to be detected, an algorithm fusion processing module for deeply mining the image features acquired by the image acquisition module to output high-resolution images, an adaptive adjustment module for receiving environmental data and adaptively adjusting various system parameters according to the acquired data, and a defect identification module that combines the adaptively adjusted parameters to identify and classify particle defects in the high-resolution image, thereby improving the recognition accuracy of micro-textures and nano-sized particles.

[0005] On the other hand, the particle defect detection method based on algorithm fusion and adaptive adjustment includes the following steps:

[0006] S1. Acquire low-resolution images of the target object and perform raw data preprocessing;

[0007] S2. After deep mining and algorithmic fusion of the low-resolution image features obtained in S1, a high-resolution image is output.

[0008] S3. Receive environmental data and adaptively adjust various system parameters based on the collected data;

[0009] S4. Combine the adaptively adjusted parameters to identify and classify particle defects in the high-resolution image.

[0010] Technical effects:

[0011] The system in this invention first performs real-time analysis of the material, shape, and ambient lighting of the object to be detected. Then, based on the analysis results, it dynamically selects and optimizes the parameters of the super-resolution algorithm and the preprocessing parameters of the CNN model to ensure that the details of the reconstructed image are clear under different environmental conditions, solving the problem of poor environmental adaptability and effectively improving the recognition rate and detection accuracy of particle defects. Unlike the single network architecture in existing technologies, the algorithm fusion processing module of this invention performs in-depth mining of image features through dual-branch processing. When a low-resolution image is acquired, the algorithm automatically starts. One branch, the FSR algorithm, is responsible for the time domain, mainly targeting moving particles to quickly improve the image frame rate. The other branch, the deep learning algorithm, is responsible for the spatial domain, extracting multi-scale features from details, enhancing the detection of minute defects, and achieving a balance between frame rate and accuracy. Attached Figure Description

[0012] Figure 1 This is a diagram showing the overall system architecture in an embodiment of the present invention.

[0013] Figure 2 This is a flowchart of the S2 algorithm fusion process in an embodiment of the present invention.

[0014] Figure 3 This is a flowchart of the adaptive adjustment process of method S3 in an embodiment of the present invention. Detailed Implementation

[0015] like Figure 1 As shown, the image acquisition module in the particle defect detection system based on algorithm fusion and adaptive adjustment in this embodiment includes a line laser source, a multi-channel optical sensor, and an image acquisition card. The line laser illuminates the surface of the object to be detected in an oblique or normal incident manner, the multi-channel optical sensor captures the scattered light signal, and the image acquisition card converts the analog signal into a digital image.

[0016] Furthermore, the algorithm fusion processing module integrates the FSR algorithm, a deep learning-based super-resolution algorithm, and a data fusion unit. The FSR algorithm quickly improves the image frame rate to generate a frame rate optimized image, which is then input into the super-resolution algorithm to generate a detail-enhanced image. The data fusion unit can weightedly fuse the frame rate optimized image and the detail-enhanced image to output a high-resolution image.

[0017] More preferably, the deep learning-based super-resolution algorithm is a multi-scale feature extraction network, which consists of five residual layers. Each residual layer is followed by an attention module. The first residual layer includes a 3×3 convolutional layer and an activation function, which can extract the basic features of the frame rate optimized image. The second to fifth residual layers include a bottleneck structure of skip connections, thereby outputting a detail-enhanced image.

[0018] Furthermore, the adaptive adjustment module includes an environmental perception sensor and a parameter optimization unit. The environmental perception sensor includes a spectral sensor capable of measuring the reflectivity ρ of the material of the object being detected; a morphology sensor capable of calculating the surface curvature κ of the object being detected; a light sensor capable of measuring the local light intensity I of the object being detected; and a temperature and humidity sensor capable of collecting surrounding environmental parameters. The parameter optimization unit uses a fuzzy logic algorithm to dynamically adjust the light compensation parameters, edge enhancement coefficients, and preprocessing parameters of the CNN model in the defect recognition module of the super-resolution algorithm in the algorithm fusion processing module based on the collected environmental parameters, in order to compensate for environmental influences. The fuzzy logic algorithm addresses environmental uncertainties (such as nonlinear changes in light intensity) by replacing precise mathematical models with fuzzy rules, thereby improving system adaptability.

[0019] It also includes a main unit and an electrical control cabinet. The main unit integrates a laser system and a precision motion platform to ensure synchronization between motion and illumination. For example, the timing of motion start / stop and laser triggering is matched to avoid image blurring. It works with the image acquisition module to complete a full surface scan of the object under test. The electrical control cabinet provides power supply support and PLC logic control to ensure the real-time operation of the algorithm fusion processing module. It also receives various environmental parameters and drives external hardware devices to execute parameter adjustment commands from the adaptive adjustment module, thus forming a closed-loop control.

[0020] This embodiment of the particle defect detection method based on algorithm fusion and adaptive adjustment includes the following steps:

[0021] S1. Acquire low-resolution images of the target object and perform raw data preprocessing;

[0022] S2. After deep mining and algorithmic fusion of the low-resolution image features obtained in S1, a high-resolution image is output.

[0023] S3. Receive environmental data and adaptively adjust various system parameters based on the collected data;

[0024] S4. Combine the adaptively adjusted parameters to identify and classify particle defects in the high-resolution image.

[0025] Furthermore, S1 specifically triggers a laser light source, synchronously activating a multi-channel optical sensor to collect scattered light signals; the image acquisition card converts the analog signal into a digital image, and automatically adds a timestamp and position coordinate information to each frame of the image.

[0026] Furthermore, such as Figure 2 As shown, S2 specifically involves performing a 2x downsampling on the low-resolution image, applying the FSR algorithm iteratively three times to reconstruct and generate a frame rate optimized image; inputting the super-resolution algorithm to perform bicubic interpolation upsampling, and outputting a detail-enhanced image; the data fusion unit calculates the gradient magnitudes of the frame rate optimized image and the detail-enhanced image, generates a weight mask based on the gradient map to highlight edge regions, and uses the weighted fusion formula: I_fused=w×I_DL+(1-w)×I_FSR,

[0027] Wherein, I_fused represents the high-resolution image output after weighted fusion by the data fusion unit; I_DL is the detail-enhanced image; I_FSR is the frame rate-optimized image; w is the gradient weight, with a value range of [0,1], which can highlight the edge regions of the image. I_DL is given a higher weight in the edge regions (w is closer to 1), and I_FSR is given a higher weight in the non-edge regions (w is closer to 0), thereby balancing the frame rate and detail accuracy; finally, histogram equalization is applied to enhance the global contrast and output a high-resolution image.

[0028] Furthermore, such as Figure 3 As shown, in step S3, the reflectivity ρ of the material of the object to be detected is measured by a spectral sensor, the surface curvature κ of the object to be detected is calculated by a morphology sensor, the local light intensity I of the object to be detected is measured by a light sensor, and the ambient temperature and humidity parameters are collected by a temperature and humidity sensor, with a range of 20±2℃ and 40±5%RH; the fuzzy logic mapping function takes the universe of discourse of the environmental parameters as input and outputs the variable α.

[0029] The reflectivity ρ is measured in the range of [0.1, 0.9] with an accuracy of ±0.02, defined as low [0.1, 0.4], medium [0.4, 0.6], and high [0.6, 0.9].

[0030] The surface curvature κ is measured in the range [0, 0.1] with a resolution of 0.01 mm^-1, and is defined as flat [0, 0.03], slightly curved [0.03, 0.07], and bent [0.07, 0.1].

[0031] Local illumination intensity I measurement range [100, 1000], accuracy ±5%, defined as dark [100, 400], medium [400, 700], bright [700, 1000];

[0032] When reflectivity ρ, surface curvature κ, and local illumination intensity I are all at a moderate level, it indicates that the image quality is stable and no special enhancement is required; α = 1.0.

[0033] When the reflectivity ρ is high and the surface curvature κ is slightly curved or bent, the processing of the reflective area needs to be enhanced, and α < 1.0. The core logic is that high reflectivity will cause reflective interference, and the processing of the reflective area needs to be enhanced by reducing the value of α (such as suppressing halo and enhancing details). The flatter the surface curvature, the more uniform the reflection, and the higher the processing intensity.

[0034] For example, ρ=0.9, κ=0.04, α=0.9; ρ=0.9, κ=0.1, α=0.7.

[0035] When the reflectance ρ is low and the local illumination intensity I is dark or medium, it is necessary to enhance the processing of low contrast areas, α > 1.0; the core logic is that low reflectance or low illumination will lead to low image contrast, and it is necessary to enhance low contrast areas by increasing the α value (such as increasing brightness or strengthening edges). The more significant the low reflectance or low illumination, the higher the processing intensity.

[0036] For example, ρ=0.2, I=200, α=1.3; ρ=0.2, I=600, α=1.1.

[0037] The above are just some examples. For complex scenarios, those skilled in the art can determine the variable α through the core logic to achieve adaptive optimization in different scenarios.

[0038] Furthermore, step S4 uses a CNN model to identify and classify particulate defects. The classification categories include particles, scratches, dents, bumps, contamination, and normal. The enhanced high-resolution image is segmented into 512×512 overlapping sub-blocks, and each sub-block is processed in parallel. Defects are detected using a sliding window with a step size of 256 pixels. Non-maximum suppression (NMS) with a threshold of 0.5 is applied, and overlapping detection boxes are merged to generate a list of defect candidate regions. After extracting the geometric features of the defect candidate regions, grayscale features are calculated. Secondary classification is performed based on the grayscale features to improve the accuracy of small-sized defect identification. Finally, a support vector machine is applied to post-process the CNN model results to improve classification accuracy.

[0039] Adjust the super-resolution algorithm parameters in the algorithm fusion processing module based on the calculated variable α:

[0040] Illumination compensation parameter β = 0.5 + 0.5 × α; Edge enhancement coefficient γ = 1.0 + 0.3 × (α - 1);

[0041] And the preprocessing parameters of the CNN model in the defect identification module:

[0042] Normalized standard deviation σ = 0.5 + 0.2 × α; threshold offset ΔT = 0.1 × (α - 1).

[0043] The image acquisition module then loads the new parameters and obtains the optimized detection image, determining whether the accuracy meets the standard. If not, it returns to the parameter optimization unit for iterative optimization. A two-way feedback mechanism is employed here: the parameters of each device in the environmental perception sensor constrain the acquisition parameters to avoid hardware overload; simultaneously, if the accuracy fails to meet the standard, environmental data is re-acquired for optimization, ensuring the algorithm's robustness.

[0044] The following detailed explanation uses the detection of surface particle defects on a 300mm wafer as an example.

[0045] First, the wafer under test is transferred to the main unit's detection position via the wafer transfer module (EFEM). The wafer positioning system performs coarse alignment using a calibrator with an accuracy of ±50μm. The main unit's precision motion platform moves the wafer under test to the scanning start position to complete the wafer loading and positioning. According to relevant technical specifications, a line laser source with a wavelength of 266nm and a power of 3mW is configured. The scanning speed is set to 150mm / s, the line width to 50μm, and the scanning interval to 10μm. The multi-channel optical sensor is configured in dark field focusing mode with a numerical aperture NA = 0.9, completing the initial settings of relevant scanning parameters.

[0046] The trigger line laser source illuminates the wafer under test at a 45° oblique angle, and simultaneously activates the multi-channel optical sensor to acquire the scattered light signal at a 12-bit depth with a frame rate of 30fps. The image acquisition card converts the analog signal into a digital image with a resolution of 512×512 pixels, and automatically adds a timestamp and position coordinate information to each frame. The raw data is preprocessed, and flat field correction is performed to eliminate the non-uniformity of the multi-channel optical sensor response. Medium-range filtering is applied to remove salt-and-pepper noise with a window size of 3×3. The image is normalized to the [0,1] floating-point range to complete the low-resolution image acquisition.

[0047] The low-resolution image is downsampled by a factor of 2 to 256×256 as the original image. The FSR algorithm is applied with a learning rate of 0.001, and the reconstruction is completed after 3 iterations to generate a frame rate optimized image. The 256×256×3 frame rate optimized image is input into a multi-scale feature extraction network and upsampled to a high-resolution image of 2048×2048 through bicubic interpolation, focusing on enhancing features such as wafer surface steps and particles. The data fusion unit calculates the gradient magnitude of the frame rate optimized image and the detail enhancement image, generates a weight mask based on the gradient map to highlight edge regions, and adopts the weighted fusion formula: I_fused=w×I_DL+(1-w)×I_FSR, where w is the gradient weight. Histogram equalization is applied to enhance global contrast.

[0048] The spectral sensor detected that the wafer is made of silicon with a reflectivity ρ = 0.3; the morphology sensor obtained that the wafer surface curvature κ < 0.01 and the local illumination intensity I = 500 Lux. The parameter optimization module calculated α = 0.7 and the illumination compensation parameter β = 1.2, and automatically adjusted the feature extraction strategy of the algorithm for the silicon surface.

[0049] The CNN model detects defects in high-resolution images. During detection, it first extracts geometric features of candidate defect regions, including area, perimeter, roundness, and aspect ratio, and calculates grayscale features such as average grayscale, standard deviation, and contrast. Then, it performs secondary classification based on these features to improve the accuracy of identifying small defects. Simultaneously, it applies a Support Vector Machine (SVM) to post-process the initial classification results from the CNN model, further enhancing classification accuracy. In the output stage, the system labels the defect location and type, generates labeled detection result images, and accurately calculates the defect size with an accuracy of ±0.05μm, capable of detecting defects ≥100nm. Finally, it generates a detection report conforming to Klarf format, containing basic information such as wafer ID, detection time, and detection parameters, as well as a defect list (coordinates, size, type, confidence level) and statistical information (total number of defects, classification distribution, defect density). It successfully identified particulate defects ≥30nm in size, achieving a 40% improvement in detection accuracy compared to existing technologies.

[0050] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. A particle defect detection system based on algorithm fusion and adaptive adjustment, characterized in that, It includes an image acquisition module for acquiring low-resolution images of the object to be detected, an algorithm fusion processing module for deep mining of image features acquired by the image acquisition module to output high-resolution images, an adaptive adjustment module for receiving environmental data and adaptively adjusting various system parameters based on the acquired data, and a defect identification module that, in combination with the adaptively adjusted parameters, identifies and classifies particle defects in the high-resolution image, thereby improving the accuracy of identifying micro-textures and nanoscale particles.

2. The particle defect detection system based on algorithm fusion and adaptive adjustment according to claim 1, characterized in that, The algorithm fusion processing module integrates the FSR algorithm, a deep learning-based super-resolution algorithm, and a data fusion unit. The FSR algorithm quickly improves the image frame rate to generate a frame rate optimized image, which is then input into the super-resolution algorithm to generate a detail-enhanced image. The data fusion unit can weightedly fuse the frame rate optimized image and the detail-enhanced image to output a high-resolution image.

3. The particle defect detection system based on algorithm fusion and adaptive adjustment according to claim 1, characterized in that, The adaptive adjustment module includes an environmental sensing sensor and a parameter optimization unit. The environmental sensing sensor includes a spectral sensor that can measure the reflectivity ρ of the material of the object being detected; a morphology sensor that can calculate the surface curvature κ of the object being detected; a light sensor that can measure the local light intensity I of the object being detected; and a temperature and humidity sensor that can collect surrounding environmental parameters. The parameter optimization unit uses a fuzzy logic algorithm to dynamically adjust the light compensation parameters, edge enhancement coefficients, and preprocessing parameters of the CNN model in the algorithm fusion processing module of the super-resolution algorithm, as well as the preprocessing parameters of the CNN model in the defect recognition module, based on the collected environmental parameters, in order to compensate for environmental influences.

4. The particle defect detection system based on algorithm fusion and adaptive adjustment according to claim 1, characterized in that, It also includes a main unit and an electrical control cabinet. The main unit integrates a laser system and a precision motion platform, which work with the image acquisition module to complete the full surface scanning of the object under test. The electrical control cabinet provides power supply support and PLC logic control to ensure the real-time operation of the algorithm fusion processing module. At the same time, it receives various environmental parameters and drives external hardware devices to execute adaptive adjustment module parameter adjustment instructions, thereby forming a closed-loop control.

5. A particle defect detection method based on algorithm fusion and adaptive adjustment, characterized in that, The particle defect detection system based on algorithm fusion and adaptive adjustment as described in any one of claims 1 to 4 includes the following steps: S1. Acquire low-resolution images of the target object and perform raw data preprocessing; S2. After deep mining and algorithmic fusion of the low-resolution image features obtained in S1, a high-resolution image is output. S3. Receive environmental data and adaptively adjust various system parameters based on the collected data; S4. Combine the adaptively adjusted parameters to identify and classify particle defects in the high-resolution image.

6. The particle defect detection method based on algorithm fusion and adaptive adjustment according to claim 5, characterized in that, S1 is specifically a trigger line laser source that synchronously activates a multi-channel optical sensor, enabling the multi-channel sensor to acquire scattered light signals; the image acquisition card converts the analog signal into a digital image, and automatically adds a timestamp and position coordinate information to each frame of the image.

7. The particle defect detection method based on algorithm fusion and adaptive adjustment according to claim 5, characterized in that, Specifically, S2 involves downsampling the low-resolution image by a factor of 2, applying the FSR algorithm iteratively three times to reconstruct and generate a frame rate optimized image; inputting the super-resolution algorithm to perform bicubic interpolation upsampling, and outputting a detail-enhanced image; the data fusion unit calculates the gradient magnitudes of the frame rate optimized image and the detail-enhanced image, generates a weight mask based on the gradient map to highlight edge regions, and uses the weighted fusion formula: I_fused=w×I_DL+(1-w)×I_FSR, Wherein, I_fused represents the high-resolution image output after weighted fusion by the data fusion unit; I_DL is the detail-enhanced image; I_FSR is the frame rate-optimized image; w is the gradient weight, with a value range of [0,1], which can highlight the edge regions of the image, assigning higher weights to I_DL in edge regions and higher weights to I_FSR in non-edge regions, thereby balancing frame rate and detail accuracy; finally, histogram equalization is applied to enhance global contrast and output a high-resolution image.

8. The particle defect detection method based on algorithm fusion and adaptive adjustment according to claim 5, characterized in that, In S3, the reflectivity ρ of the material of the object to be detected is measured by a spectral sensor, the surface curvature κ of the object to be detected is calculated by a morphology sensor, the local light intensity I of the object to be detected is measured by a light sensor, and the ambient temperature and humidity parameters are collected by a temperature and humidity sensor, with a range of 20±2℃ and 40±5%RH; the fuzzy logic mapping function takes the universe of discourse of the environmental parameters as input and outputs the variable α. The reflectivity ρ is measured in the range of [0.1, 0.9] with an accuracy of ±0.02, defined as low [0.1, 0.4], medium [0.4, 0.6], and high [0.6, 0.9]. The surface curvature κ is measured in the range [0, 0.1] with a resolution of 0.01 mm^-1, and is defined as flat [0, 0.03], slightly curved [0.03, 0.07], and bent [0.07, 0.1]. Local illumination intensity I measurement range [100, 1000], accuracy ±5%, defined as dark [100, 400], medium [400, 700], bright [700, 1000]; When reflectivity ρ, surface curvature κ, and local illumination intensity I are all at a moderate level, it indicates that the image quality is stable and no special enhancement is required; α = 1.

0. When the reflectivity ρ is high and the surface curvature κ is slightly curved or bent, it is necessary to enhance the treatment of the reflective area, α < 1.0; When the reflectivity ρ is low and the local illumination intensity I is dark or medium, it is necessary to enhance the processing of low contrast areas, and α > 1.

0.

9. The particle defect detection method based on algorithm fusion and adaptive adjustment according to claim 5, characterized in that, The S4 method uses a CNN model to identify and classify particulate defects, including particles, scratches, dents, bumps, contamination, and normal defects. The enhanced high-resolution image is segmented into 512×512 overlapping sub-blocks, and each sub-block is processed in parallel. Defects are detected using a sliding window with a step size of 256 pixels, non-maximum suppression is applied, and a threshold of 0.5 is used. Overlapping detection boxes are merged to generate a list of candidate defect regions. Geometric features of the candidate defect regions are extracted, and grayscale features are calculated. Secondary classification is performed based on these grayscale features to improve the accuracy of small-sized defect identification. Support vector machines are then applied to post-process the CNN model results to further enhance classification accuracy.

10. The particle defect detection method based on algorithm fusion and adaptive adjustment according to claim 8 or 9, characterized in that, Adjust the super-resolution algorithm parameters in the algorithm fusion processing module based on the calculated variable α: Illumination compensation parameter β = 0.5 + 0.5 × α; Edge enhancement coefficient γ = 1.0 + 0.3 × (α - 1); And the preprocessing parameters of the CNN model in the defect identification module: Normalized standard deviation σ = 0.5 + 0.2 × α; Threshold offset ΔT = 0.1 × (α - 1); The image acquisition module then loads the new parameters and acquires the optimized detection image, determines whether the accuracy meets the standard, and if not, returns to the parameter optimization unit for iterative optimization.

Images