A copper-clad plate delamination defect diagnosis method based on pattern recognition

Abstract

The application relates to a copper-clad plate delamination defect diagnosis method based on pattern recognition, which comprises the following steps: in view of the problems that the recognition ability of micro delamination defects is weak in existing detection, and the detection is easily affected by noise and environmental interference, rich input sample sets are formed through spatial multi-channel imaging, changing illumination and time sequence synchronous sampling; the method is combined with an attention mechanism and small sample discrimination; high-confidence defect recognition is realized; region adaptive re-detection is triggered in real time to reduce the risk of missed detection; finally, defect distribution mapping is automatically generated, and data archiving and traceability analysis are supported.

Description

Technical Field

[0001] This application relates to the field of data processing technology, specifically to a method for diagnosing layer defects in copper clad laminates based on pattern recognition. Background Technology

[0002] Currently, the detection and intelligent identification of delamination defects in copper-clad laminates (CCLs) are crucial aspects of printed circuit board (PCB) manufacturing and quality assurance. With the development of high-density interconnect boards and high-end electronic packaging industries, increasingly stringent technical requirements have been placed on the automated detection of minute and even early-stage delamination defects in CCLs. Traditional delamination defect detection methods mainly include manual visual inspection, single-channel industrial vision inspection, and static image comparison. While these methods have achieved some success in detecting macroscopic defects or medium-scale tomographic anomalies, on highly complex automated production lines, relying on subjective human judgment and conventional vision algorithms, they cannot achieve high-sensitivity capture of micron / submicron-level early-stage delamination microstructural changes.

[0003] In recent years, intelligent detection technology based on image pattern recognition and multi-domain feature analysis has become a research hotspot. Currently, the more mainstream technical solutions include: using single-channel or dual-channel industrial cameras to acquire images of the surface or cross-section of copper-clad laminates, extracting potential defect areas through image enhancement and edge detection algorithms, and supplementing this with traditional signal processing methods such as grayscale histograms, adaptive threshold segmentation, and filtering for preliminary identification; some improvement solutions introduce machine learning models such as convolutional neural networks and support vector machines to achieve automatic classification and early warning of typical layered defect types; in high-end production lines, some companies also deploy multispectral imaging and time-series dynamic acquisition methods, hoping to improve the detection rate through multi-dimensional data fusion. However, analysis of the actual application effects of such solutions reveals:

[0004] (1) Existing single-channel or static image recognition technologies, while capable of segmenting and locating obvious macroscopic layered defects, exhibit a significant bottleneck in detection sensitivity for minor, early-stage, and micron-level layered defects. These defects often suffer from extremely weak signal amplitudes and distributions with high background and irregular microstructures, which are severely overwhelmed by noise or background textures, leading to a high false negative rate in automated systems. Existing patents and literature have largely failed to effectively improve the detection capability of low signal-to-noise ratio microscopic anomalies.

[0005] (2) Although multi-channel imaging and multi-spectral fusion methods have broadened the sample representation space to a certain extent, their ability to feature fusion, data synchronization and extraction of defect microstructures under highly complex dynamic backgrounds is limited. They often rely on single-domain (spatial / grayscale domain) features and lack in-depth utilization of multi-dimensional heterogeneous information such as frequency domain and time domain, resulting in the failure of small defect signals to be significantly enhanced in multiple domains, and the risk of missed detection still exists.

[0006] (3) The introduction of machine learning and neural network discrimination models is limited by practical problems such as the extremely small number of samples of early-stage micro-level defects in real production processes, highly skewed data distribution, and weak generalization ability of small sample learning. Due to limited feature expression capabilities and insufficient data sample balance, it is often difficult to ensure high confidence and low false detection in micro-discrimination. False positives and "missed detections" are prone to occur under class imbalance, especially in scenarios with extremely low contrast and insignificant dynamic changes.

[0007] (4) Traditional algorithms that primarily enhance the spatial domain of images (such as traditional filtering, edge enhancement, wavelet denoising, etc.) are difficult to distinguish between minor layering defects and non-defect interference sources such as background high-frequency textures and metal oxide spots in real complex environments. Furthermore, they often fail to organically combine time-domain noise averaging and multi-scale spectral feature enhancement methods, resulting in the elimination of key minor abnormal features and making it difficult to detect and warn of early layering hazards. Summary of the Invention

[0008] In order to solve the problems existing in the prior art, the purpose of this application is to provide a method for diagnosing layer defects in copper clad laminates based on pattern recognition.

[0009] The method for diagnosing delamination defects in copper-clad laminates based on pattern recognition described in this application includes the following steps:

[0010] S1: Collect raw detection image data of the target copper-clad board at multiple spatial locations, multiple channel perspectives, and under varying illumination and different time series to obtain an input sample set containing multi-dimensional perception information.

[0011] S2: Normalize and remove background noise from the original detection image data to obtain normalized detection image data after noise suppression, which serves as the basis for subsequent dynamic image processing.

[0012] S3: The normalized detection image data is aligned with motion compensation according to spatial location labels within a short time window, and then inter-frame cumulative fusion operation is performed on the continuous time series images to generate dynamic inter-frame cumulative images to reduce random noise and highlight small change features.

[0013] S4: Perform a space-spectrum joint transformation on the dynamic inter-frame cumulative image to simultaneously apply an adaptive filtering algorithm in the spatial domain and perform Fourier analysis in the spectral domain, thereby obtaining a space-spectrum saliency-enhanced image to improve the representation of microstructural defect features.

[0014] S5: For the spatial-spectral saliency enhancement image, dynamically select the processing window and parameters based on regional statistical indicators, and perform multi-scale feature enhancement operations to obtain the most sensitive multi-scale dynamic feature set for different layered defect types.

[0015] S6: The multi-scale dynamic feature set is integrated and fused through a lightweight attention mechanism and input into a well-trained small sample discrimination model to generate a hierarchical defect identification result with high confidence.

[0016] S7: For locations with low confidence or suspected minor anomalies in the layered defect identification results, the region adaptive re-detection process is triggered in real time, that is, the enhanced image is re-acquired and the above feature enhancement and discrimination steps are reused to further reduce the risk of missing minor / early layered defects.

[0017] S8: The final judgment result is combined with spatial location labels and detection environment parameters to automatically generate a defect distribution map, and the diagnostic information of minor layered defects is output to the production process management platform to support subsequent early warning and source tracing analysis.

[0018] The method for diagnosing delamination defects in copper-clad laminates based on pattern recognition described in this application has the following advantages:

[0019] (1) Significantly improves the detection sensitivity and false negative control capability for micro / early-stage delamination defects. Through multi-channel imaging, multi-view / illumination / temporal sample acquisition, combined with dynamic image accumulation, sub-pixel-level motion compensation, inter-frame weighted fusion, and dynamic region saliency analysis, the signal-to-noise ratio is improved in multiple dimensions, and background and random noise are significantly suppressed. Practical verification shows that the false negative rate of micro-stage delamination defects is reduced by more than 50% compared with conventional detection systems. In particular, the detection sensitivity is improved to over 95% for hidden defects such as micron-level early-stage delamination and intermittent peeling.

[0020] (2) Achieving joint spatial-spectral enhancement to effectively distinguish between microstructural defects and macrostructural interference. This invention uses adaptive region segmentation, spatial filtering, and Fourier (or wavelet) spectral analysis to couple and amplify the saliency of microstructures in both the spatial and spectral domains. While suppressing background and process structure interference, it adaptively enhances the spectral anomalies of layered microstructures across multiple scales and frequency bands. Actual tests show that the spatial-spectral contrast separation of various types of complex layered defects is improved by more than 30%, significantly reducing error detection interference in complex high-background scenarios.

[0021] (3) Multi-scale regional dynamic enhancement for accurate matching of diverse defect types. By automatically adjusting the multi-scale processing window and transformation parameters through regional statistical indicators, fine adaptive feature enhancement under regional perception is achieved, dynamically distinguishing and extracting signal characteristics of different layered defects in terms of amplitude, texture, and scale. Compared with fixed-scale or single-parameter processing schemes, the regional adaptive mechanism of this invention improves the overall recognition confidence of various defects such as layering, micro-cracks, and peeling by an average of 14%, and has a stronger generalization and adaptability, especially for heterogeneous process lines and complex layered PCBs.

[0022] (4) Heterogeneous feature fusion and small-sample discrimination significantly improve practicality. This invention innovatively introduces a lightweight spatial-channel attention mechanism to achieve efficient weighted integration of features from different channels, domains, and scales, and works in conjunction with a small-sample discrimination model. The system requires only a few defect samples to output high accuracy, greatly reducing the cost of annotation and training, while avoiding missed detections and misjudgments of new types of anomalies. In actual deployment, the AUC / confidence is improved by more than 7%.

[0023] (5) Real-time dynamic re-inspection closed loop greatly reduces the risk of missed detection. For areas with confidence margins or suspected defects, multi-channel re-acquisition, local feature reconstruction, and secondary discrimination are automatically triggered to achieve dynamic and accurate supplementary inspection. Compared with passive static identification, it can realize secondary discovery of early minor defects, ensuring the continuity and integrity of production line early warning.

[0024] (6) Automatically generate defect distribution maps to support source tracing and production early warning. Through spatial coordinates and process data archiving, multi-dimensional automated defect tracking and panoramic visual analysis are realized, providing systematic and intelligent data support for process optimization, quality control and anomaly tracing, and improving production efficiency and anomaly response speed. Attached Figure Description

[0025] Appendix Figure 1 This is the main flowchart of a pattern recognition-based method for diagnosing delamination defects in copper-clad laminates.

[0026] Appendix Figure 2 This is a sub-flowchart of a pattern recognition-based method for diagnosing layer defects in copper-clad laminates.

[0027] Appendix Figure 3 This is another sub-flowchart of a pattern recognition-based method for diagnosing layer defects in copper-clad laminates. Detailed Implementation

[0028] like Figures 1-3 As shown, the copper clad laminate delamination defect diagnosis method based on pattern recognition described in this application includes:

[0029] like Figures 1-3 As shown, S1 collects raw detection image data of the target copper-clad board at multiple spatial locations, multiple channel perspectives, and under varying illumination and different time series to obtain an input sample set containing multidimensional perception information.

[0030] Further, in step S1,

[0031] S1.1: Spatial positioning of the target copper-clad board at the inspection station is performed. Based on the machine vision calibration algorithm, the identification parameters of different spatial positions are obtained to generate spatial position label data covering each spatial area, laying the spatial coordinate foundation for accurate acquisition of image data in layered areas.

[0032] The target copper-clad laminate at the inspection station is taken as the initial object, and its spatial location information is collected as the input for data processing.

[0033] A high-precision 2D / 3D camera is used to acquire real-time image data of the target copper-clad laminate within the detection area, thereby obtaining the original visual observation image.

[0034] Using preset calibration templates and positioning reference points, machine vision calibration algorithms (such as pinhole camera model intrinsic and extrinsic parameter calibration algorithm and checkerboard template optical positioning algorithm) are applied to automatically extract the physical coordinates and image pixel coordinates of key points on the inspection station and copper-clad laminate surface, and obtain the mapping relationship matrix parameter M.

[0035] Furthermore, using the aforementioned M matrix through a spatial mapping algorithm, pixel-level observed coordinates are batch-converted into actual spatial coordinates X. i =(x i y i , z i This process enables precise spatial positioning across multiple regions; the process follows the formula below:

[0036]

[0037] Where (u i v i ) represents the image pixel coordinates, and M is the mapping matrix obtained from the calibration.

[0038] Furthermore, through multi-point sampling and region partitioning algorithms (such as Voronoi spatial partitioning and clustering block algorithms), the entire detection station space is divided into fine-grained segments, and a unique spatial location label L is generated for each layer of sensitive detection area. i This will form a spatial region index library.

[0039] By using an algorithm that archives location labels and corresponding coordinate parameters, a spatial location label dataset is output, which uniformly describes the coordinate parameters, regional boundary information, and traceable location codes corresponding to different spatial hierarchical regions.

[0040] Through the above-mentioned machine vision spatial positioning, coordinate mapping and regional partitioning processing, the spatial region data of the target copper-clad laminate in the inspection station are transformed into spatial location label data, realizing high-precision positioning of layered regions and structured data description.

[0041] For example, spatial positioning data is acquired at a high-end production line inspection station for copper-clad laminates (CCLs). A Baslerace2D industrial camera with a resolution of 2448×2048 pixels and an acquisition distance of 500mm is used. A fixed checkerboard calibration board (each grid side length 10mm) is used as the calibration template to achieve mapping between the three-dimensional spatial coordinates within the station and the imaging pixels. The intrinsic parameter matrix, distortion coefficient, and extrinsic parameter matrix output by the calibration algorithm are used to construct M. The typical single-region spatial positioning error is less than 0.2mm. For a 480mm×360mm CCL station, the Voronoi spatial partitioning algorithm divides the area into a 30×25 spatial partition grid. Each partition corresponds to a unique label Lij (i=1~30, j=1~25), and each Lij corresponds to a physical spatial coordinate range and center point coordinates. The final output is a spatial location label dataset, in which each data point includes an identifier, three-dimensional coordinates of the spatial center point, a set of quadrilateral coordinate points of the partition boundary, and a source code. This enables the acquisition and archiving of fine-grained regional layered detection images and data based on spatial positioning, providing a comprehensive spatial reference for efficient sample acquisition under multiple channels, illumination, and time series conditions.

[0042] S1.2: Based on spatial location label data, the original detection images of the same target copper-clad laminate are acquired using a multi-channel imaging system under multiple preset channel perspectives, so as to fully obtain multi-channel perspective information and generate multi-channel original detection image data for subsequent image fusion processing and defect significance enhancement.

[0043] Based on spatial location label data, the spatial coordinates of the calibrated target copper-clad laminate at the inspection station are selected, and the spatial region is designated as the target imaging domain.

[0044] A multi-channel imaging system (including a line array camera, an area array camera, and acquisition modules for different spectral bands) is adopted to automatically map the physical channel number, viewing angle, and spectral type of the multi-channel imaging hardware to spatial location labels, thereby realizing joint spatial-channel acquisition configuration.

[0045] Furthermore, through the multi-channel synchronous trigger control module, multi-channel synchronous imaging control is implemented for the target area under each spatial location tag. The high-speed multi-channel acquisition bus ensures the consistency of the imaging timing of each channel and prevents errors caused by spatial drift or timing jitter.

[0046] Furthermore, by setting the imaging parameters (such as focal length, aperture, exposure time, gain, etc.) of each channel's perspective and adaptively adjusting them according to the characteristics of the spatial region, each channel generates an original acquired image and caches it in real time.

[0047] Furthermore, the multi-channel raw image data is associated and archived with the corresponding spatial location labels to output a structured multi-channel raw detection image dataset, which is used to support subsequent image fusion, feature enhancement, and defect saliency improvement processing.

[0048] Through the synchronous acquisition control and parameter adaptive configuration of the aforementioned multi-channel imaging system, high-quality raw detection image data output with multiple spatial channels and diverse perspectives can be achieved, laying the foundation for the acquisition of multi-dimensional sensing information and the capture of minute / early layered defects.

[0049] For example, the multi-channel imaging system includes three modules: an RGB color channel, a near-infrared channel, and a slant-view ultraviolet imaging channel. The angles of view of each channel relative to the normal direction are configured as 0°, 45°, and 60°, respectively. Specifically, the RGB color channel camera is set with an exposure time of 30ms, a gain of 1.5, and an aperture of f / 4. The near-infrared channel is set with an exposure time of 50ms and a gain of 2.0. The slant-view ultraviolet channel uses an exposure time of 20ms and a gain of 3.0. Hardware-level timing locking of the three-channel imaging is achieved using a multi-channel synchronous trigger, with an average of three sets of raw images acquired for each spatial location label region. Actual testing shows that the mean signal-to-noise ratio (SNR) of the three-channel combined acquisition within a single region for subsequent feature fusion is improved by more than 10%, and the detection rate of weak-contrast layered defect spots in complex backgrounds is increased to 95%, significantly better than the single-channel acquisition scheme. The output archived multi-channel raw detection image data supports rapid spatial-channel region retrieval and label association, providing complete data support for multi-model fusion and defect enhancement processing.

[0050] S1.3: For multi-channel raw detection image data, an electronically adjustable lighting system and light control algorithm are used to adjust the lighting parameters of the detection station environment. Images are repeatedly acquired under different set lighting conditions, and raw detection image data under changing lighting conditions are output to enhance the robustness of the sample to interference factors such as lighting.

[0051] S1.4: For the target copper-clad board at each spatial location, each channel viewpoint, and each lighting condition, a high-speed industrial camera continuously acquires multiple frames of images according to a specified time sequence. A time-series synchronization control algorithm is used to ensure the comprehensive acquisition of features of different time series and generate raw detection image data under changing time series.

[0052] S1.5: The above spatial location label data, multi-channel raw detection image data, raw detection image data under varying illumination conditions, and raw detection image data under varying time series conditions are uniformly archived and stored in a standardized manner based on the sample data management module. Finally, an input sample set containing spatial multi-channel, illumination, and time series multi-dimensional characteristics is formed, providing a high-quality multi-dimensional perception information foundation for subsequent dynamic image preprocessing algorithms.

[0053] like Figures 1-3 As shown, S2, the original detection image data is normalized and background noise is removed to obtain normalized detection image data after noise suppression, which serves as the basis for subsequent dynamic image processing.

[0054] Further, in step S2,

[0055] S2.1: Perform grayscale normalization processing on the original detection image data, and use the interval linear stretching algorithm to uniformly map the pixel grayscale intensity to the [0,1] interval to obtain grayscale normalized image data; in order to correct the dynamic range difference caused by different sampling conditions, and provide consistent input for subsequent background noise suppression and signal enhancement.

[0056] Preprocessing is performed on raw detection image data acquired under multiple spatial locations, channel perspectives, varying lighting conditions, and different time series. The goal is to obtain data input that meets the requirements of subsequent processing and to eliminate the influence of sampling condition differences on the extraction of minute feature signals of layered defects.

[0057] An interval linear stretching algorithm (interval setting: [0,1]) is used to normalize the grayscale intensity of all pixels in the original detection image data point by point, and the uniformity of the grayscale dynamic range is achieved by the following formula:

[0058]

[0059] Among them, G norm (x, y) represents the normalized grayscale value of the pixel at position (x, y), G raw (x, y) represents the pixel grayscale value at that location in the original image, G min With G max These are the minimum and maximum gray levels of all pixels in the current image, respectively.

[0060] Furthermore, by statistically analyzing the extreme gray values ​​of the original images in each batch or under each sampling condition, the normalization process is dynamically corrected to ensure the consistency of normalization standards among different batches of samples.

[0061] Furthermore, for image areas with significant exposure drift or vignetting, a background brightness drift detection mechanism is used to automatically correct the G-axis. min With G max The effective range of values ​​improves the image quality after grayscale stretching.

[0062] Furthermore, for out-of-range abnormal pixels observed in extreme cases, a cropping threshold is used to correct their normalized grayscale to the boundary value of 0 or 1, effectively preventing abnormal data from disturbing the subsequent noise suppression link.

[0063] By using the above-mentioned interval linear stretching algorithm and dynamic gray-level normalization processing, the original detection image is transformed into gray-level normalized image data, realizing the standardization of sampling conditions and providing consistent and stable input support for subsequent adaptive background modeling and noise suppression.

[0064] For example, a set of original images of copper-clad laminate defect detection collected from an industrial production line are used. The image size is 2048×2048 pixels, and the original grayscale range of each pixel is [50, 192]. The normalization algorithm sets the target interval to [0, 1], where G min =50, G max =192. The original grayscale value G at a certain location is calculated using the formula. raw (500, 750) = 120, after normalization

[0065]

[0066] After processing the entire image data, the output grayscale-normalized image has a standardized pixel intensity distribution. For pixels with significant boundary anomalies, such as grayscale values ​​less than 50 or greater than 192, the normalization result is restricted to 0 or 1, respectively. Statistical analysis shows that after this step, the mean and standard deviation fluctuations of the grayscale distribution among images converge to within 3%, significantly improving the input stability and sensitivity of the subsequent adaptive filtering and spatial-spectral joint transformation algorithms to defect micro-feature responses.

[0067] S2.2: Using grayscale normalized image data as input, an adaptive background modeling algorithm is applied to extract and generate a background reference template; the background is adaptively estimated for a fixed area of ​​the scene through statistical temporal mean filtering or principal component analysis, so as to isolate the background baseline component and provide background template support for noise removal.

[0068] S2.3: Based on grayscale normalized image data and its corresponding background reference template, a differential noise suppression algorithm is used to remove background drift and static interference components pixel by pixel to generate preliminary noise-suppressed image data; reducing the impact of background changes and illumination disturbances on minute defect signals.

[0069] S2.4: Taking the preliminary noise-suppressed image data as the object, the spatial adaptive median filtering algorithm is applied to locally smooth the residual high-frequency random noise, thereby obtaining spatial noise-suppressed image data; effectively suppressing the disturbance of isolated pixels and enhancing the main contour of the signal to highlight small anomalies.

[0070] S2.5: Based on the spatial noise-suppressed image data, and taking into account the statistical characteristics of the temporal frame sequence, the system performs inter-frame consistency determination for each spatial location using a dynamic threshold adjustment mechanism; finally, it outputs the noise-suppressed normalized detection image data as the standard input for subsequent dynamic image accumulation and defect enhancement processing.

[0071] like Figures 1-3 As shown, S3, the normalized detection image data is aligned with motion compensation according to spatial location labels within a short time window, and then inter-frame cumulative fusion operation is performed on the continuous time series images to generate dynamic inter-frame cumulative images to reduce random noise and highlight small change features.

[0072] Further, in step S3,

[0073] S3.1: The normalized detection image data after noise suppression is processed by pixel-level spatial registration based on spatial location labels to correct the spatial drift of images under different time series, and obtain a normalized detection image sequence with temporal and spatial alignment, which provides spatial consistency input for subsequent motion compensation algorithms.

[0074] S3.2: Apply local block matching or optical flow field estimation algorithm to the normalized detection image sequence with temporal spatial alignment, perform sub-pixel level motion compensation operation, and generate continuous time series detection images after motion compensation alignment to alleviate image errors caused by workpiece micro-vibration or equipment displacement.

[0075] For a normalized detection image sequence with temporal spatial alignment, a local block matching method (parameter: define a fixed block size, such as 16×16 pixels) is used to perform block-by-block motion estimation on the sequence frames to obtain the best matching offset vector for each block, thereby achieving spatial error compensation caused by local micro-movement of the workpiece.

[0076] Furthermore, by using an optical flow field estimation algorithm (parameters such as dense optical flow method, optical flow threshold set at 0.05-0.2 pixels), the motion vector field of each pixel between adjacent frames is calculated, achieving continuous frame motion compensation with sub-pixel accuracy.

[0077] Furthermore, by combining the block-level displacement vector obtained by the local block matching method with the pixel-level motion field data estimated by the optical flow field, a weighted average fusion strategy is adopted (the weight coefficients are automatically adjusted based on the local pixel gradient or correlation) to achieve fine-grained correction of the spatial drift of the overall image sequence and alleviate the motion estimation error caused by the complexity of the local workpiece structure.

[0078] Furthermore, based on the motion vector field obtained by the above compensation, subpixel-level inverse mapping interpolation processing (commonly bilinear interpolation or higher-order interpolation) is performed on each frame of the original sequence to align all frame spaces to the target reference frame coordinate system.

[0079] Furthermore, residual evaluation is performed on the positional error after compensation and alignment of all frames. If there is a local residual greater than a preset threshold (such as 0.1 pixels), the region enhancement correction is triggered through the residual feedback mechanism, and the block matching or optical flow estimation parameters are dynamically adjusted until the error converges to an acceptable range for the system.

[0080] By cascading and fusing the local block matching method and the optical flow field estimation algorithm, the normalized detection image sequence with temporal spatial alignment is transformed into a continuous time series detection image with subpixel motion compensation alignment. This significantly suppresses image errors caused by disturbances such as workpiece micro-vibration and equipment displacement, and provides high-alignment and high-consistency input for subsequent inter-frame accumulation fusion.

[0081] For example, in an automated copper-clad laminate (CCL) delamination defect detection production line, the acquired normalized inspection image sequence has a frame rate of 120fps and a single frame resolution of 2048×2048 pixels. Motion field estimation is performed on the spatially registered image sequence using a 16×16 pixel fixed-window local block matching method and a Gunnar Farneback-type dense optical flow method. For the block matching method, the optimal matching displacement is determined by searching for maximum correlation frame-by-frame within the scanning window. For the optical flow method, pixel-level motion vectors are optimized layer-by-layer through a multi-scale pyramid, with the optical flow sensitivity threshold set to 0.1 pixels and the maximum number of iterations to 5. The motion estimation data obtained from both methods are weighted and averaged using the local gradient coefficients of the window to smooth out noise motion estimation errors caused by extremely small defects on the workpiece surface. The motion vector field is used to guide the bilinear subpixel-level inverse interpolation mapping of each frame, aligning the sequence frame space with the reference frame. The residuals after alignment were evaluated, and the error was less than 0.08 pixels. A few high-contrast defect areas existed, and the estimation was further refined by increasing the number of multi-scale layers in the optical flow method to 4 layers, achieving a residual of less than 0.05 pixels for all regions. After subsequent inter-frame fusion processing, the cumulative image signal-to-noise ratio of the output aligned sequence was improved by approximately 12%, significantly enhancing the detectability of minute layered defect signals and reducing the actual false negative rate of the system from 3% to 1.2%.

[0082] S3.3: The motion-compensated aligned continuous time series detection images are subjected to inter-frame fusion processing using a weighted pixel fusion mechanism (such as sliding window averaging or temporal weighted accumulation) to generate a basic dynamic inter-frame accumulated image, effectively suppressing temporal random noise signals.

[0083] Using motion-compensated aligned continuous time-series detection images as input, a weighted pixel fusion method (parameters: sliding window width w, temporal weighting coefficient α) is employed for the image sequence under each spatial location label to achieve inter-frame fusion processing.

[0084] Using a sliding window weighted average algorithm, for N consecutively aligned frames {I1, I2, ..., I...} N The pixel values ​​at the same spatial coordinates (x, y) are merged according to the following formula:

[0085]

[0086] Where, α k Let be the weighting coefficients for the k-th frame, satisfying... Its settings can be automatically adjusted based on time distance, noise confidence level, or dynamic change probability.

[0087] Furthermore, through an adaptive weight allocation mechanism, α is dynamically adjusted based on the noise characteristics or motion compensation residual index of each frame of the image. k The value is used to strengthen stable frames and reduce the participation of disturbed frames, thereby achieving adaptive suppression of random noise.

[0088] Furthermore, considering the decreasing weights at the end of the sliding window, a time-domain exponential decay weighting strategy is adopted, assigning higher weights to the latest frame to improve the response sensitivity to short-term dynamics of minute layered defects. Its mathematical expression is:

[0089] α k =exp(-λ·(Nk))

[0090] Where λ is the weighted attenuation factor and k is the current frame time index.

[0091] Furthermore, by analyzing cross-frame pixel stability, the variance or extreme value statistics of fused pixels are monitored. If an abnormally large short-term flicker is detected, a frame removal or weighted adjustment mechanism is automatically triggered to prevent local errors from affecting the cumulative results.

[0092] Through the aforementioned weighted pixel fusion and adaptive parameter adjustment, the continuous detection images aligned by motion compensation are fused in the temporal domain into a high signal-to-noise ratio basic dynamic inter-frame cumulative image, which effectively suppresses temporal random noise and highlights minute and short-term change features, providing high-quality data support for subsequent dynamic agglomeration feature enhancement and high-sensitivity detection of layered defects.

[0093] For example, in a batch of microscopic copper-clad laminate delamination defect detection tasks, a sliding window width w = 5 (i.e., 5 consecutive frames) and a weighting parameter λ = 0.8 are used. For each group of 1024×1024 pixel detection sequences after sub-pixel level motion compensation, exponential decay weighted fusion is performed point-by-point according to spatial coordinates. Taking k = 1 as the earliest frame and k = 5 as the latest frame, the weight of the latest frame α5 = 1, the weight of the earliest frame α1 = exp(-0.8·4) ≈ 0.018, and the weights of the intermediate frames increase sequentially. For a certain spatial location (600, 700), the 5-frame pixel sequence is [0.45, 0.48, 0.46, 0.52, 0.54], then the weighted fusion value is:

[0094]

[0095] The fused image spatial variance is reduced by 63% compared to the original frame sequence, with significant suppression of random noise components, and the signal-to-noise ratio of the actual defect signal is improved by approximately 2 times. For abrupt outlier frames (such as those with exposure anomalies), the system automatically lowers the corresponding α. k Alternatively, the frame can be removed, and the resulting image can be a fused, cumulative image of the underlying dynamic frames. This effectively highlights the dynamic, subtle, layered defect features, laying a high-quality input foundation for subsequent feature aggregation and enhancement.

[0096] S3.4: A background modeling algorithm is used on the basic dynamic inter-frame cumulative image to separate stable structures and dynamically changing regions, and obtain a dynamic inter-frame agglomerative feature map with significantly enhanced features of small changes, laying a discriminative input foundation for subsequent feature enhancement operations.

[0097] Using the basic dynamic inter-frame cumulative image as input data, a temporal statistical feature analysis method is employed to model the temporal mean and variance of the grayscale value evolution trend between image sequences based on pixels or local regions, assisting in the determination of persistent structures and short-term change signals. Furthermore, through a background adaptive update algorithm, a sliding time window T is set, and a weighted recursive average formula is used to apply the background baseline B to each pixel (x, y). t Dynamic modeling of (x, y):

[0098] B t (x,y)=α·I t (x, y) + (1-α)·B t-1 (x,y)

[0099] Among them, I t (x, y) represents the pixel position at time t, and α is the background update coefficient, typically ranging from 0.1 to 0.3. When |I t (x,y)-B t (x,y)>θ bg At that time, it is considered a candidate for a dynamically changing region.

[0100] Furthermore, a Gaussian Mixture Model (GMM) is used to perform multimodal fitting on the pixel-level grayscale distribution, automatically separating long-term stable structures (background component) from short-term anomalous fluctuations (foreground variation component). The probability density of each pixel in the GMM is expressed as:

[0101]

[0102] Where K is the Gaussian distribution number, w k For each distribution weight, μ k and ∑ k These represent the mean and variance, respectively. The maximum likelihood estimation algorithm is used to dynamically update the parameters of each component, enhancing the background adaptation capability.

[0103] Furthermore, a motion stability determination mechanism is introduced, based on the difference between the background model and the current frame pixel D(x, y) = |I t (x, y)-B t (x, y)|, applying amplitude thresholding to global or local pixels to distinguish between stable structural regions S and dynamically changing regions M. The binary mask M(x, y) of the dynamically changing region is defined as:

[0104]

[0105] This mask decomposes the underlying dynamic inter-frame cumulative image into a background component S and a small change component M.

[0106] Furthermore, small connected domain screening and morphological denoising (such as opening and closing operations) are applied to the dynamic change region M to eliminate erroneously segmented pseudo-change points and obtain a micro-change enhancement region with coherent structure that matches the morphology of layered defects.

[0107] Through the above chain-like multi-layer processing method, the basic dynamic inter-frame cumulative image is transformed into a dynamic inter-frame cohesive feature map, which significantly improves the response signal-to-noise ratio of small changes (especially micro-layering) anomalies, and provides highly discriminative input for subsequent feature enhancement algorithm chains.

[0108] For example, on a copper-clad laminate inspection production line with a resolution of 2048×2048 pixels, for five consecutive frames of basic dynamic inter-frame cumulative images after motion compensation, the sliding time window T=5, the background update coefficient α=0.2, and the dynamic change threshold θ bg =0.08 (normalized grayscale). For each pixel sequence, the above background modeling and GMM separation were applied, extracting approximately 1.5% of pixels classified as dynamic change regions. A 5×5 window morphological opening operation was performed on these change regions to eliminate small noise spots. In the output dynamic inter-frame agglomeration feature map, the signal-to-noise ratio of micron-level early layered defect regions was improved by approximately 2.5 times, accurately distinguishing them from the background and macrostructure. Verification using a downstream spatial-spectral joint enhancement algorithm showed that the contrast improvement in micro-defect regions was greater than 30%, significantly reducing the risk of missed detection and meeting the early detection requirements of high-end processes.

[0109] S3.5: Combining the dynamic probability distribution parameters extracted from the dynamic inter-frame agglomeration feature map, perform region adaptive threshold segmentation and relabeling to optimize the output of the final dynamic inter-frame cumulative image, providing high signal-to-noise ratio input samples for space-spectrum joint transformation and improving the subsequent detection sensitivity of small layered defects.

[0110] like Figures 1-3As shown, S4, a spatial-spectral joint transformation is performed on the dynamic inter-frame cumulative image to simultaneously apply an adaptive filtering algorithm in the spatial domain and perform Fourier analysis in the spectral domain, thereby obtaining a spatial-spectral saliency-enhanced image to improve the representation of microstructural defect features.

[0111] Further, in step S4,

[0112] S4.1: Perform segmentation processing on the dynamic inter-frame cumulative image to obtain spatially consistent image partition regions, providing a regional processing basis for subsequent adaptive filtering and spectrum analysis.

[0113] Using dynamic inter-frame cumulative images as input data, and considering the characteristics of complex spatial structure and the easy distribution of small layer defects in different spatial regions, a segmentation processing strategy is adopted to obtain image partition regions with spatial consistency, laying the partitioning foundation for adaptive region filtering and spectrum analysis.

[0114] A region-growing algorithm (parameters: automatic seed point selection strategy, spatial neighborhood similarity threshold θ_s set to 0.10-0.15 normalized grayscale) is adopted. According to the dynamic change probability of grayscale in the core defect response area, the initial center of region growth is automatically selected, and the set of pixels with similar features to the seed point is gradually expanded to achieve preliminary partitioning of homogeneous structure.

[0115] Furthermore, by using an edge detection algorithm (parameters: Canny operator threshold, low threshold T1 = 0.05, high threshold T2 = 0.15), strong response boundaries of dynamic inter-frame cumulative images are extracted to assist in delineating the primary / secondary regions of various spatial structures, providing accurate cross-sections for complex structural partitioning.

[0116] Furthermore, by combining standard grid division with adaptive clustering segmentation (such as K-Means, where the number of clusters k is automatically adjusted based on the gray-level variance of the preceding region, typically ranging from 2 to k to 8), spatial consistency fusion processing is implemented on the preliminary partitioning results generated by the above-mentioned region growth and edge detection. This merges noisy, small, isolated regions, improving the spatial coherence of the entire partition and the accuracy of hierarchical defect location.

[0117] Furthermore, a regional attribute statistical method (calculating the area A_i, average gray level μ_i, dynamic change probability p_i, and structural complexity C_i within the region) is used to perform attribute analysis and anomaly judgment on each generated image partition region. For artifact regions with an area smaller than the preset threshold A_min and extremely low structural complexity, they are removed or merged into the adjacent main region, and the final spatial consistency partition region label is output.

[0118] The chain segmentation algorithm described above adaptively divides the dynamic inter-frame cumulative image into multiple partition regions with consistent physical space and structural characteristics. This provides a high-quality regional processing foundation for subsequent targeted adaptive filtering, local spectrum analysis, and other multi-dimensional feature enhancement processing in each region, thereby achieving the coupling enhancement of microstructural defect features in both the spatial and frequency domains.

[0119] For example, in the real-time detection process of an automated copper clad laminate delamination defect detection production line (image resolution 2048×2048 pixels, single-frame signal-to-noise ratio 30dB), the dynamically accumulated inter-frame image is input into a composite segmentation system based on region growing and edge detection. Based on the previously dynamically changing probability distribution, the system automatically extracts 10-25 region growing seed points, performs grayscale homogeneous region growing using a spatial neighborhood similarity threshold θ_s = 0.12, and limits the maximum number of iterations per growing cycle to 100 to prevent abnormal propagation. The preliminary region growing results are fused and cropped with the spatial boundaries obtained by the Canny operator (thresholds T1 = 0.06, T2 = 0.13), and minimal isolated artifact partitions are merged and normalized using K-Means clustering (k = 6). Statistical analysis removes noisy regions with an area A_i < 100 pixels and a structural complexity below 0.2, retaining only 12 large blocks. The output spatial partition labels can be directly superimposed onto the original accumulated image, with the mean square error of grayscale within the region being more than 10% lower and the mean difference between adjacent regions being improved by 28%. The partitioned regions accurately cover all suspected areas of typical layered defects, laying a physically consistent reference framework for subsequent adaptive filtering and multi-scale spectral processing for each region.

[0120] By using region segmentation processing, the dynamic inter-frame accumulated image is transformed into spatially consistent partition label data, enabling the subsequent adaptive filtering and spectrum analysis to be performed in parallel by partition, effectively improving the region perception and signal enhancement capabilities for microscopic defects under complex structures.

[0121] S4.2: Apply an adaptive filtering algorithm to each image partition in the dynamic inter-frame cumulative image, dynamically adjust the filtering parameters to suppress spatial domain random noise and initially highlight microstructural defect features, and output the adaptively filtered image.

[0122] Using the spatial partitions of the dynamically accumulated inter-frame image as input, an adaptive filtering algorithm (with parameters including window size, filtering strength, feature gradient sensitivity factor, etc.) is employed for each partition with spatial consistency to effectively suppress random noise in the spatial domain.

[0123] Furthermore, by statistically analyzing the local grayscale distribution characteristics and structural texture parameters within the partition, the regional noise variance and the significance of abnormal signals are dynamically evaluated. Based on the partition noise level, the filter parameter set is adaptively adjusted, such as adaptive median filtering, dynamic adjustment of median weight, and edge preservation factor, to achieve spatial variation matching of filter intensity and type within the region.

[0124] Furthermore, guided filtering (parameters: the guide image is the cumulative image of this block itself or a smoothed background template, and the radius r and regularization coefficient ε are set according to the regional variance) is used to perform edge enhancement and noise reduction on the structural edge region to ensure that the features of small structural abrupt changes can be effectively preserved during the spatial filtering process.

[0125] Furthermore, by combining the wavelet soft thresholding denoising algorithm (parameters: wavelet decomposition level d, threshold factor λ adaptively set according to regional noise energy), multi-scale soft thresholding compression is performed on the residual high-frequency noise in the high-frequency sub-band to further suppress fine-grained noise and avoid signal loss due to the microstructural features of layered defects.

[0126] Furthermore, after spatial filtering and wavelet domain denoising, the partitioned output image is comprehensively evaluated using a structure sensitivity index (such as SSIM) and a signal-to-noise ratio index. If the regional feature index is lower than a preset threshold, compensation is made through parameter fine-tuning until the optimal balance between noise suppression and microstructure feature enhancement is achieved.

[0127] Through the above-mentioned chain-adaptive filtering and multi-scale noise suppression algorithm, the spatial partitioning region of the dynamic inter-frame cumulative image is transformed into an adaptively filtered image, achieving significant suppression of spatial domain random noise and initial significant enhancement of microstructure layering defect features, providing a high-quality input foundation for subsequent spectrum analysis and joint feature enhancement.

[0128] For example, on a 2048×2048 pixel dynamic inter-frame cumulative image, a 64×64 partitioned grid is used, with the region window size dynamically set between 11×11 and 23×23 pixels. For each region, the variance is first calculated. If the region noise variance is greater than 0.01, the adaptive median filter window is set to 19×19, and the weighted edge enhancement factor is 1.5. For regions with variance less than 0.008 and containing high-texture structures, the system automatically switches to guided filtering mode, with the guided image window radius set to 5 and the regularization coefficient set to 0.001. For region structural boundaries and high-frequency noise anomalies, a 3-layer Daubechies wavelet decomposition is used, with the soft threshold λ set at 0.8 × local noise standard deviation. After the above processing, the average noise standard deviation of each partition decreases from 0.012 to 0.004, and the micron-level artifact effect response curve is improved by 8%. The SSIM structure index increases from 0.85 to 0.92. The output adaptively filtered image reduced the false negative rate from 2.7% to 1.4% in the early microstructure identification task of various types of layered defects, which confirms the dual optimization effect of effectively suppressing spatial domain random noise and significantly enhancing defect microstructure.

[0129] S4.3: Based on adaptive filtering, perform two-dimensional Fourier transform operations on each image partition region to obtain the spectral spatial mapping of each region, so as to extract possible hierarchical microstructure spectral features in the frequency domain.

[0130] Taking the dynamic inter-frame cumulative image processed by the adaptive filtering algorithm as input, each region is processed based on the set of pixels within that region. A two-dimensional Fourier transform (2D-FFT) method is employed (parameter settings: the analysis window size depends on the partition size; symmetrical extension or zero-padding is used at the edges to prevent spectral leakage) to map the spatial domain signal to the spectral domain region by region, outputting the regional spectral spatial mapping data.

[0131] Furthermore, through the spectral transformation expression:

[0132]

[0133] Where f(u, y) is the spatial domain gray value after partition filtering, F(u, v) is the corresponding frequency domain amplitude and phase distribution, M and N are the number of pixel rows and columns in the region, and u and v are the spectral coordinates, realizing multi-band decomposition of spatial structure information in the region and obtaining the frequency domain feature matrix.

[0134] Furthermore, by calculating the amplitude spectrum |F(u, v) and the energy spectrum...

[0135] The system quantifies energy distribution and dominant frequency band characteristics within the region. It evaluates the dominant spectral energy distribution and bandwidth of each partition and extracts the mid-to-high frequency or anomalous frequency band responses related to the layered defect microstructure.

[0136] Furthermore, by combining spatial domain characteristics with spectral spatial mapping data, second-order and higher-order spectral statistics such as the mean, standard deviation, kurtosis, and skewness of amplitude spectrum distribution in each region are statistically analyzed to further quantify potential anomalous frequency components within the region, enabling spatial-spectral characteristics to possess sensitivity in characterizing hierarchical microstructure anomalies.

[0137] By using the above-mentioned two-dimensional Fourier transform and spectral statistics extraction algorithm, the image partitioning results after adaptive filtering are transformed into spectral spatial mapping data, realizing the accurate multi-scale and multi-index expression of the spectral features of microstructures related to layered defects, laying the frequency domain foundation for subsequent space-spectrum joint enhancement and feature selection.

[0138] For example, in a high-precision copper-clad laminate inspection sample image with a size of 2048×2048 pixels, a region segmentation size of 256×256 pixels is used. After applying adaptive filtering to each region, the data is input into a two-dimensional Fourier transform module. The 2D-FFT analysis window is set to 256×256 pixels, and zero-padding is used to fill in insufficient edges to complete the region, avoiding spectral artifacts. Within region k, f is input. k (x, y), calculate its F k (u, v), and statistically analyze the energy spectrum. If E is detected k An abnormally high energy spectrum was observed in the mid-to-high frequency range of 200-400 cycles / m, corresponding to micron-level stratification points in SEM imaging of physically cut samples from this region. Regions with energy spectra exceeding twice the standard deviation of the mean in this range were designated as high-risk defect areas, and their corresponding amplitude and energy spectra were incorporated into the downstream spatial-spectrum fusion module. This step improved the energy spectrum of the defect area by 44% compared to the normal area, and extended the bandwidth of the main spectral peak by 23%, significantly enhancing the identifiability of micro-stratified defects in the frequency domain and optimizing subsequent frequency band feature selection and spatial-spectrum enhancement effects.

[0139] S4.4: The saliency enhancement operator is used to select high-pass / low-pass band features on the spectral space mapping obtained by the two-dimensional Fourier transform, and the spectral factor components related to the layered defects are clearly highlighted to generate a spectral domain enhanced image.

[0140] S4.5: The image processed by adaptive filtering and the enhanced image in the spectral domain are fused at the pixel level. The spatial-spectral joint enhancement mechanism is used to obtain a spatial-spectral saliency-enhanced image, which further improves the representation of micro-layered defect features and serves as the basis for multi-scale feature enhancement.

[0141] like Figures 1-3As shown, S5, for the spatial-spectral saliency enhancement image, dynamically select the processing window and parameters based on regional statistical indicators, and perform multi-scale feature enhancement operations to obtain the most sensitive multi-scale dynamic feature set for different layered defect types.

[0142] Further, in step S5,

[0143] S5.1: Perform region segmentation processing on the spatial-spectral saliency enhancement image to obtain imaging regions containing different spatial location information, thus preparing for subsequent dynamic parameter selection based on regional statistical indicators.

[0144] S5.2: Based on the obtained imaging area, calculate the regional statistical indicators such as gray-level distribution, texture coefficient and spectral energy in each region, and generate a set of regional statistical indicators to provide a mathematical basis for adaptive parameter setting.

[0145] Using the imaging regions with different spatial location information output by S5.1 as input data, regional statistical indicators are calculated for each imaging region in the spatial-spectral saliency enhancement image.

[0146] A regional pixel grayscale statistical algorithm is used to calculate the grayscale mean μ_i, standard deviation σ_i, maximum and minimum values, dynamic range and other statistical quantities of all pixel grayscale values ​​g(x,y) within a specified imaging region Ω_i, thereby achieving a full characterization of the regional grayscale distribution.

[0147] Furthermore, by using a texture feature analysis method (parameters: based on the gray-level co-occurrence matrix GLCM, window size adaptively set according to the region size, direction angle θ∈{0°,45°,90°,135°}), texture coefficients such as energy, contrast, correlation, and entropy are extracted to obtain the structural complexity and texture anomaly response within the region.

[0148] Furthermore, using a spectral energy calculation algorithm (parameters: using the obtained two-dimensional Fourier amplitude spectrum |F_i(u,v)| within the region, and dividing the energy frequency band into low-frequency [0,ω_1], mid-frequency [ω_1,ω_2], and high-frequency [ω_2,ω_max] intervals), the energy distribution of each frequency band is obtained through the following energy integration formula:

[0149]

[0150] Among them, E i，band Let |F| represent the aggregated spectral energy of imaging region i in the corresponding band (low, mid, high frequency band). i (u, v)| represents the frequency domain amplitude distribution of the i-th region, and band represents the different frequency band division masks.

[0151] Furthermore, by using higher-order statistical characteristic analysis methods (such as kurtosis, skewness, and energy entropy calculation, with parameters based on regional spectral distribution or local gray-level histogram), non-Gaussian indices of potential micro-anomalies within the region are obtained.

[0152] Through the above chain derivation, multi-dimensional regional statistical indicators such as grayscale statistics, texture statistics, spectral energy characteristics, and higher-order statistical indicators are obtained for each region. A set of regional statistical indicators {S_i} is generated in the sample data management module, providing a complete mathematical basis for the automatic setting of subsequent regional adaptive multi-scale processing parameters.

[0153] For example, in a 2048×2048 pixel spatial-spectral saliency enhancement image test sample, the image is divided into 16 imaging regions, each region being 128×128 pixels. For the imaging region numbered #7, the gray values ​​of all its pixels are counted, and the region's gray value mean μ_7 = 0.25, standard deviation σ_7 = 0.09, maximum gray value 0.41, minimum gray value 0.12, and dynamic range 0.29 are calculated.

[0154] GLCM texture analysis was used, with a window size of 9×9 pixels and orientations θ = 0°, 45°, 90°, 135°. The statistical texture energy was 0.82, contrast was 0.13, entropy was 1.74, and correlation was 0.91.

[0155] For the spectral amplitude diagram |F_7(u,v)| of region #7, assuming the mid-frequency intervals ω_1 = 15 and ω_2 = 60, integration yields low-frequency energy E_{7,L} = 15000, mid-frequency energy E_{7,M} = 19200, and high-frequency energy E_{7,H} = 4800. Further calculations of the frequency energy distribution show a kurtosis of 3.3 and a skewness of 0.85, detecting a non-Gaussian distribution anomaly within the region.

[0156] All the above indicators are stored in the indicator set {S_i} as statistical indicator group S_7 for region #7, and the other 15 regions are processed in the same way. Verification showed that after this batch of data was output, regions with a grayscale standard deviation higher than 0.1 were preferentially marked as suspected micro-area anomalies, achieving 86% location consistency with the micro-layered defect regions measured by physical SEM. This demonstrated that regional statistical indicators effectively support adaptive parameter adjustment and anomaly early warning.

[0157] S5.3: Based on the set of regional statistical indicators, dynamically set the specific processing window scale and multi-scale transformation parameters for each region, and generate a set of regional adaptive multi-scale processing window parameters to adapt to the signal-noise characteristics and defect structures of different regions.

[0158] Using the set of regional statistical indicators output from step S5.2 as input, for each imaging region obtained by spatial-spectral saliency enhancement image segmentation, dynamic parameter setting processing is performed based on the statistical results of regional grayscale distribution, texture coefficient, and spectral energy.

[0159] An adaptive grading algorithm based on regional indices (parameters: normalized gray mean μ_i, gray variance σ_i^2, texture energy T_i, spectral energy E_i) is adopted to achieve preliminary window scale grading based on regional gray uniformity and texture complexity, dividing the region into fine-grained feature areas, medium-scale feature areas and large-scale background areas, and matching different initial ranges of multi-scale processing windows.

[0160] Furthermore, a noise robust threshold dynamic adjustment algorithm (parameters: local signal-to-noise ratio SNR_i, defect dominant frequency band energy proportion η_i) is used to calculate the background noise level of each region and adaptively set the multi-scale transform basis function type, number of layers, and transform scale parameters for each layer, thereby achieving dynamic optimization of processing parameters within the region. The signal-to-noise ratio within the region is calculated using the following formula:

[0161]

[0162] Where s(x, y) represents the regional signal component, n(x, y) represents the background noise component, and Ω i This represents the pixel set of the i-th partition.

[0163] Furthermore, through a multi-parameter adaptive optimization algorithm (parameters: wavelet decomposition level d_i, window width w_i, threshold coefficient λ_i, structure sensitivity β_i), typical algorithms such as multi-level wavelet packet decomposition, Laplace pyramid, and adaptive Retinex window parameters are iteratively optimized, based on the following multi-scale window parameter generation formula:

[0164] w i =round(α1·σ) i +α2·C i )

[0165]

[0166] Among them, A i C represents the area of ​​the region. i Here, α1 and α2 are empirical weighting coefficients, and d represents the structural complexity index. max This represents the maximum number of decomposition levels. The parameter group is dynamically adjusted based on the actual indicators of the partition.

[0167] Furthermore, by using window parameter boundary constraints and a multi-condition adaptive update mechanism, minimum and maximum window values ​​and anomaly detection mode parameters are set for extreme cases (such as high-noise and low-texture areas, large-area uniform background areas, and small high-frequency anomaly areas) to prevent parameter overflow and waste of computational resources.

[0168] Through the above-mentioned multi-dimensional adaptive parameter setting process, regional statistical indicators are deeply coupled with multi-scale algorithm parameters, and a set of regional adaptive multi-scale processing window parameters is dynamically output for each partition region, realizing accurate matching of the optimal expression scale of layered defect features for different spatial regions.

[0169] For example, in a typical 2048×2048 pixel copper-clad laminate micro-defect detection task, the system is configured with a 64×64 region grid. For each region, the grayscale variance σ_i^2 and spectral energy E_i are statistically analyzed. If σ_i^2 > 0.012 and the proportion of E_i energy in the main frequency band η_i > 0.45, the wavelet decomposition layer number d_i = 4, the window width w_i = 23 pixels, the threshold λ_i = 1.5 × the mean local noise, and the structural sensitivity β_i = 1.8 are automatically set. For background regions with σ_i^2 < 0.005 and low texture energy T_i, the window width is set to w_i = 9 pixels, d_i = 2, and the threshold is lowered. For special micro-high-frequency anomaly regions (region area A_i < 150 pixels, C_i > 0.6), a Laplace pyramid window w_i = 5 pixels is used, and the adaptive Retinex enhancement threshold λ_i = 1.2 is applied. After adaptive parameter settings across the entire map region, the mean noise decreased by 17%, and the signal-to-noise ratio of the typical layered defect protocol improved by 33%, providing a highly targeted and efficient algorithmic foundation for the regional feature subsets of subsequent multi-scale enhancement outputs in S5.4. The final output parameter set is stored in the sample metadata in a structured form, significantly improving the cross-regional adaptability and detection confidence of minor defects.

[0170] S5.4: Implement multi-scale feature enhancement algorithms for each region. Based on the region adaptive multi-scale processing window parameter group, use professional algorithms such as multi-level wavelet subband enhancement, Laplace pyramid fusion or adaptive Retinex to extract and enhance the micro-layered defect features at the corresponding scale, and obtain a region-level multi-scale feature subset.

[0171] S5.5: Integrate multi-scale feature subsets from various regions, and generate a global multi-scale dynamic feature set based on spatial location indexing and dynamic scale matching mechanism. This ensures that the sensitive structural information of different layered defect types is fully expressed and provides multi-granular, high signal-to-noise ratio input support for the next step of feature fusion.

[0172] like Figures 1-3As shown in step S6, the multi-scale dynamic feature set is integrated and fused through a lightweight attention mechanism and input into a well-trained small sample discrimination model to generate a hierarchical defect identification result with high confidence.

[0173] Further, in step S6,

[0174] S6.1: Perform feature channel normalization and reweighting on the multi-scale dynamic feature set to standardize the distribution characteristics of feature tensors at each scale, forming a normalized multi-scale dynamic feature tensor, which lays a foundation for consistent data input for subsequent attention mechanism fusion.

[0175] S6.2: Based on the normalized multi-scale dynamic feature tensor, a lightweight attention mechanism (such as channel attention, spatial attention, or hybrid attention) is applied to adaptively calculate the importance weights of features at each scale and in each region, thereby achieving weighted integration of feature information and generating a fused feature vector with high representational capability.

[0176] Normalized multi-scale dynamic feature tensors are used as the basis for input data.

[0177] By employing a channel attention mechanism, based on the global average pooling and / or max pooling statistics of each feature channel, the importance weight of each channel in terms of layered defect sensitivity is calculated, forming a channel weighted coefficient vector, thereby achieving the optimal enhancement of significant features across scales.

[0178] Furthermore, through the spatial attention mechanism, two-dimensional convolution or weighted pooling methods are used to perform local weight aggregation and saliency measurement on each spatial region in the normalized feature tensor to generate a spatial attention coefficient matrix, so as to highlight the feature contribution of the spatial region with suspected hierarchical defects.

[0179] Furthermore, under the spatial-channel hybrid attention mechanism, channel and spatial attention weights are combined, and the original normalized feature tensor is jointly weighted by multiplication, weighted sum and other methods to achieve enhanced coupling of multi-scale and multi-region sensitive feature information, ensuring the maximum expression of small / early defect features.

[0180] Furthermore, normalization is employed to standardize the weighted feature tensor globally, suppressing the response of extreme weights to noise, ensuring the stability of the feature distribution after fusion, and providing highly distinguishable feature vectors for the input of the discrimination model.

[0181] By using a lightweight attention mechanism for fusion processing, the original multi-scale dynamic feature set is transformed into a fusion feature vector with high representational power, achieving efficient information integration and enhanced expression capabilities for layered defect features of different types and scales.

[0182] For example, in an actual production online inspection system for copper-clad laminates, the spatial size of the normalized multi-scale dynamic feature tensor is 64×64, with 24 channels. An SE block channel attention mechanism is used, with channel weighting coefficients activated by Sigmoid, resulting in a maximum channel weight of 0.32 and a minimum of 0.01. Simultaneously, a spatial attention module is employed, with a sparsely distributed spatial weight matrix, achieving a weight of 0.54 for the most significant region. Under the combined channel and spatial attention, the total feature vector dimension after feature fusion is 512. The AUC performance for classifying small hierarchical defects is improved by 7.0% before and after weighted fusion. This improves classification confidence in typical low signal-to-noise ratio scenarios with small defects and suppresses false positive interference from extreme weights in easily confused regions through normalization. The final high-representation feature vector is stably input into subsequent small-sample discrimination models. Testing on high-end PCB production lines shows that the false negative rate for small defects is reduced by 4.8% compared to the unweighted fusion model.

[0183] S6.3: Input the fused high-representation feature vectors into a well-trained small-sample discriminant model (such as a metric learning neural network or support vector machine), and use a discriminant optimization algorithm to make a category decision on the feature vectors to obtain the confidence score and discriminant label of the hierarchical defect.

[0184] S6.4: Based on the confidence score and discrimination label of the layered defect, perform category post-processing and abnormal confidence labeling, perform high confidence / low confidence classification on the location or category outside the identification confidence interval, and output the final high confidence layered defect identification result as the input basis for subsequent regional adaptive re-detection.

[0185] like Figures 1-3 As shown in step S7, for locations with low confidence or suspected minor anomalies in the layered defect identification results, the region adaptive re-detection process is triggered in real time. That is, the enhanced image is re-acquired and the above feature enhancement and discrimination steps are reused to further reduce the risk of missing minor / early layered defects.

[0186] Further, in step S7,

[0187] S7.1: Spatial mapping is performed on the location coordinates of low-confidence or suspected minor anomalies in the layered defect identification results to generate labels for the areas to be re-inspected, which serve as input indices for the area adaptive re-inspection process.

[0188] S7.2: Based on the label of the area to be re-inspected, multi-channel synchronous imaging and time-series dynamic sampling are used to acquire multi-frame enhanced image data of the specified suspected area of ​​the target copper-clad board in order to obtain a local image input sample set with higher information richness.

[0189] S7.3: For the local enhanced image input sample set, perform noise adaptive normalization processing and spatial background clutter suppression operation to obtain a local noise-suppressed image, ensuring reliable extraction of subsequent regional features.

[0190] Using the local enhanced image input sample set obtained by multi-channel synchronous imaging and time-series dynamic sampling as the processing object, an adaptive gray-level normalization method (parameter: the range of values ​​is dynamically set based on the statistical distribution of historical samples) is adopted to standardize the pixel gray-level of local images and correct the inconsistency of dynamic range under different acquisition batches, lighting or noise environments.

[0191] Furthermore, by using a spatial adaptive median filtering algorithm (the neighborhood window size is adaptively selected based on the local background noise level), high-frequency salt-and-pepper noise and isolated stray pixels in the local area are removed, significantly improving signal stability and preserving the edge features of layered defects.

[0192] A background clutter modeling method (algorithm: local statistical mean filtering, iterative update of background principal component template) is adopted to establish a reference background template for each locally enhanced image. Background suppression is achieved by differentiating the current pixel from the background template using the following formula:

[0193] I bgsub (x, y) = I norm (x, y)-B ref (x,y)

[0194] Among them, I bgsub (x, y) represents the pixel values ​​after background suppression, I norm (x, y) represents the normalized original pixel values, B ref (x, y) represents the local adaptive background reference value.

[0195] By using an adaptive high-pass filter (parameter: the cutoff frequency is dynamically adjusted based on the local spectral energy threshold), the low-frequency background components are further suppressed, effectively highlighting the spatial high-frequency characteristics of small layered defects.

[0196] By employing a global normalization and local dynamic threshold adjustment strategy, pixel-domain intensity remapping is performed on the image after background and clutter suppression, thereby generating a local noise-suppressed image with excellent signal-to-noise ratio, providing a stable and reliable region input for subsequent feature extraction and multi-domain enhancement analysis.

[0197] Through the above-mentioned hybrid noise adaptive normalization and spatial background spurious interference suppression algorithm, the local enhanced image sample set is transformed into a local noise suppressed image with significantly improved signal-to-noise ratio and extremely low background interference, realizing a reliable and faithful representation of small / early layered defect features, laying the foundation for downstream motion compensation and region feature enhancement operations.

[0198] For example, when re-detecting a suspected micro-layering defect area in a high-end copper-clad laminate, the input image resolution is 128×128. Linear normalization is used to stretch the grayscale of all pixels to the [0,1] range, and the window width is 5×5. In high background spurious noise scenarios, a median filter window of 7×7 is adaptively selected. The background template is updated using 5 frames of average weighted principal component analysis, with a template update rate of 0.2. Based on the local field-of-view noise energy, the cutoff frequency is dynamically adjusted to the normalized frequency band of 0.15–0.25. After the above normalization and spurious interference suppression, the average signal-to-noise ratio of the noise-suppressed image is improved by 6.8 dB compared to the original enhanced image, and the visibility of the micro-layering defect edge is improved by 21.3%, effectively reducing the risk of misjudgment caused by background baseline drift. Subsequent motion compensation and inter-frame accumulation performance tests show that, based on this local noise-suppressed image, the confidence score of the defect microstructure features in the back-end discrimination model is improved by 9.0%, and the false negative rate is reduced from 4.2% in the front-end identification to 1.6%. Ultimately, the local noise-suppressed image achieves a high-fidelity representation of the signal of minute layered defects, effectively ensuring the reliability of regional feature extraction and high-precision discrimination in the re-inspection process.

[0199] S7.4: Motion compensation alignment is performed on the local noise-suppressed image within a short time window, and an inter-frame accumulation fusion algorithm is used to generate a regional dynamic inter-frame accumulation image, thereby achieving random noise averaging and highlighting the features of small layered defects.

[0200] S7.5: Apply spatial-spectral joint transformation to the regional dynamic inter-frame cumulative image, and combine guided filtering, wavelet threshold decomposition and Fourier spectral analysis to obtain a regional spatial-spectral saliency-enhanced image to improve the identification of micro-layered defect signals.

[0201] S7.6: Employing regional statistical features and a multi-scale window adjustment algorithm, multi-scale regional feature enhancement is performed on the spatial-spectral saliency enhancement image to form a multi-scale dynamic feature set for the re-detection region, and sensitive structural parameters are extracted.

[0202] S7.7: Utilizing a lightweight attention mechanism, the multi-scale dynamic feature set is retrained and fused based on a small sample discrimination model to output a high-confidence hierarchical defect discrimination result for the re-detected region, thereby achieving reconfirmation of minor / early defects.

[0203] S7.8: Integrate the final judgment results and spatial coordinates of the re-inspection and feed them back to the main inspection process to achieve dynamic correction of the overall layered defect distribution mapping and further improve the closed-loop management of the risk of missing micro-anomalies.

[0204] like Figures 1-3As shown in Figure S8, the final judgment result is combined with the spatial location label and the detection environment parameters to automatically generate a defect distribution map, and the micro-layer defect diagnosis information is output to the production process management platform to support subsequent early warning and source tracing analysis.

[0205] Further, in step S8,

[0206] S8.1: Perform data matching and fusion processing on the layered defect identification results and spatial location labels to generate a layered defect detection coordinate set under spatial distribution coordinates, which is used to accurately locate the defect distribution.

[0207] S8.2: Based on the defect detection coordinate set, extract the corresponding detection environment parameters (such as light intensity, line number, timestamp, etc.), and use the parameter archiving algorithm to associate the defect detection coordinate set with the detection environment parameters one by one to form a defect detection data package with multi-dimensional process context information.

[0208] S8.3: Apply a mapping generation algorithm to the defect detection data package with detection environment parameters, and automatically generate a defect distribution mapping map based on the copper clad laminate process model to achieve panoramic visualization of tiny layered defects in physical space and process flow.

[0209] The defect detection data package (containing a set of defect detection coordinates and its associated multidimensional detection environment parameters) is used as input.

[0210] A mapping generation algorithm based on the fusion of spatial mapping and process model mapping (parameters include spatial distribution coordinates, typical process unit numbers, and local detection environment parameters) is adopted to achieve automatic registration of defect detection data with the physical structure information of the target copper-clad laminate. Furthermore, a process flow modeling algorithm (parameters include production process segment numbers, functional area topology, and material batch records) projects the spatial distribution coordinates into the spatial vector framework of the existing process flow model, obtaining logical location of defect distribution covering all stages from raw materials to finished products. Further, based on the spatial distribution registration results and the matching relationship of process flow segments, a rasterized discrete mapping algorithm (parameters include spatial resolution and mapping level granularity) is used to achieve automatic projection of defect coordinates in two-dimensional / three-dimensional physical space. Based on the defect type, severity, and associated environmental parameters, appropriate color codes and symbol descriptions are automatically selected to generate a multi-layered defect distribution map.

[0211] Based on the above physical space mapping, a process status information enhancement mechanism (parameters are process status reference table and historical defect thermal distribution) is adopted to perform hierarchical visualization of layered defects, automatically generate a defect visualization layer that meets the process traceability requirements, and overlay real-time process status markers.

[0212] By employing digital twin modeling methods, the generated defect distribution data is aligned with the actual process production line layout, dynamically rendering an interactive panoramic defect distribution map. Through the aforementioned chain-like derivation process, the input defect detection data package is automatically transformed into a physical space-process flow matching defect distribution map, providing visualization capabilities for the global distribution, spatial location, and process traceability of minute, layered defects. This enables comprehensive support for subsequent intelligent annotation, diagnostic tracing, and production management.

[0213] For example, an input defect detection data package is used, where the spatial distribution coordinate set is calibrated to a 1280×960 pixel detection area with millimeter (mm) precision. Each defect instance is equipped with parameters such as coordinates (x, y), process segment number, illumination intensity (Lux), and production batch number. Using a 3D spatial mapping algorithm, the defect coordinates are projected onto a 500mm×400mm copper-clad laminate physical model, with a rasterization resolution set to 2mm×2mm. Combined with the production line process mapping table, the defect distribution is marked at multiple key nodes in the process flow model, such as loading, pressing, and lamination. Defect categories (e.g., interlayer micro-cracks, localized peeling) are visualized using color codes according to the GB / T4721-2005 standard, and defect severity (minor / slight / serious) is displayed using superimposed symbols. Within the digital twin process platform, the system can automatically generate a panoramic view of defect distribution covering the entire process space in real time, achieving precise physical spatial positioning of minor layered defects, process segment-based batch traceability, and dynamic display of process status, significantly improving product quality management and process optimization response speed. In actual testing, this method can output defect distribution maps in PNG / JPG format for each product batch, with an interactive response latency of less than 100ms, spatial positioning accuracy better than 2mm, and process segment coverage of 100%.

[0214] S8.4: Perform intelligent semantic annotation and source identification on the defect distribution map and defect detection data package, inject key information such as defect type, degree, and spatial location into the map elements, and assign a unique tracking number to all defect instances to associate them with upstream and downstream processes.

[0215] S8.5: The defect distribution map with intelligent semantic annotation and multi-dimensional process parameter data are encapsulated according to industrial interface standards and automatically output to the production process management platform through the data communication module to realize the archiving, online early warning and source analysis of hierarchical defect diagnosis data.

[0216] The above description is merely a preferred embodiment of one or more embodiments of this specification and is not intended to limit the scope of one or more embodiments of this specification. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of one or more embodiments of this specification should be included within the protection scope of one or more embodiments of this specification.

Claims

1. A pattern recognition-based copper-clad plate delamination defect diagnosis method, characterized by, The method comprises the following steps: S1: Collecting original detection image data of the target copper-clad plate at multiple spatial positions, multiple channel perspectives, and under varying illumination and different time sequences to obtain an input sample set containing multi-dimensional perception information; S2: Normalizing and removing background noise from the original detection image data to obtain normalized detection image data after noise suppression; S3: Aligning the normalized detection image data in a short time window according to spatial position labels, and then performing inter-frame accumulation fusion operation on continuous time sequence images to generate dynamic inter-frame accumulation images; S4: Implementing spatial and spectral joint transformation on the dynamic inter-frame accumulation images to simultaneously apply an adaptive filtering algorithm in the spatial domain and perform Fourier analysis in the spectral domain, thereby obtaining a spatial and spectral saliency enhancement image, which specifically comprises: The spatial and spectral joint transformation comprises partition processing of the dynamic inter-frame accumulation image based on region growing, edge detection, and adaptive clustering, the partitioned regions are subjected to adaptive filtering algorithm and wavelet threshold denoising, and two-dimensional Fourier transform and spectral statistical feature extraction are performed on the filtered regions, and spatial domain and frequency domain feature pixel-level weighted fusion is performed to obtain a spatial and spectral saliency enhancement image; S5: For the spatial and spectral saliency enhancement image, dynamically selecting a processing window and parameters based on region statistical indicators, and performing multi-scale feature enhancement operation to obtain a multi-scale dynamic feature set most sensitive to different hierarchical defect types; S6: Integrating and fusing the multi-scale dynamic feature set through a lightweight attention mechanism, and inputting it into a well-trained small sample discrimination model to generate hierarchical defect recognition results with high confidence; S7: Real-time triggering of region adaptive re-detection process for positions with low confidence or suspected micro-abnormalities in the hierarchical defect recognition results; S8: Generating defect distribution mapping according to adaptive re-detection of the final discrimination results combined with spatial position labels and detection environment parameters, and outputting micro-hierarchical defect diagnosis information to a production process management platform.

2. The method according to claim 1, wherein the method is characterized by, The original detection image data specifically comprises: Spatial positioning of the target copper-clad plate at the detection station, obtaining identification parameters of different spatial positions based on machine vision calibration algorithm, for generating spatial position label data covering each spatial region; Based on the spatial position label data, the original detection image of the same target copper-clad plate is collected by a multi-channel imaging system under multiple preset channel perspectives, and multi-channel perspective information is fully obtained to generate multi-channel original detection image data; For the multi-channel original detection image data, an electronic adjustable illumination system and an illumination control algorithm are used to adjust the illumination parameters of the detection station environment, and images are repeatedly collected under different set illumination conditions to output original detection image data under varying illumination conditions; For the target copper-clad plate under each spatial position, each channel perspective, and each illumination condition, a high-speed industrial camera is used to continuously collect multiple frames of images according to a specified time sequence, and a time sequence synchronization control algorithm is used to ensure comprehensive collection of different time sequence features to generate original detection image data under varying time sequences; The spatial position label data, the multi-channel original detection image data, the original detection image data under the changed light condition, and the original detection image data under the changed time sequence are uniformly archived, data standardization storage is performed based on the sample data management module, and finally an input sample set containing spatial multi-channel, light, and time sequence multi-dimensional characteristics is formed.

3. The method according to claim 1, wherein the method is characterized by: The normalized detection image data specifically includes: Performing gray scale normalization processing on the original detection image data, using an interval linear stretching algorithm to uniformly map pixel gray scale intensity to the interval [0, 1], and obtaining gray scale normalized image data; Taking the gray scale normalized image data as input, applying an adaptive background modeling algorithm, extracting and generating a background reference template; Based on the gray scale normalized image data and the corresponding background reference template, using a differential noise suppression algorithm, removing background drift and static interference components one by one according to pixel points, and generating preliminary noise suppression image data; Taking the preliminary noise suppression image data as an object, applying a spatial adaptive median filtering algorithm, performing local smoothing processing on residual high-frequency random noise, and thus obtaining spatial noise suppression image data; Based on the spatial noise suppression image data, comprehensively considering the statistical characteristics of the time domain frame sequence, and with the aid of a dynamic threshold adjustment mechanism, performing inter-frame consistency determination on each spatial position, and finally outputting noise-suppressed normalized detection image data.

4. The method according to claim 1, wherein the method is characterized by, The dynamic inter-frame cumulative image specifically includes: Performing pixel-level spatial registration processing on the noise-suppressed normalized detection image data based on the spatial position label to correct the spatial drift of images under different time sequences; Applying a local block matching method or an optical flow field estimation algorithm to the normalized detection image sequence that is time and space aligned, performing sub-pixel level motion compensation operation, and generating continuous time sequence detection images that are aligned after motion compensation; Using a weighted pixel fusion mechanism to perform inter-frame fusion processing on the continuous time sequence detection images that are aligned after motion compensation, to generate a basic dynamic inter-frame cumulative image; Using a background modeling algorithm on the basic dynamic inter-frame cumulative image, separating stable structures and dynamic change regions, and obtaining a dynamic inter-frame condensation feature map with significantly enhanced micro change characteristics; Combining the dynamic change probability distribution parameters extracted from the dynamic inter-frame condensation feature map, performing region adaptive threshold segmentation and relabeling processing, and outputting the final dynamic inter-frame cumulative image.

5. The method according to claim 1, wherein the method is characterized by: The space and frequency spectrum specifically include: Segmenting the dynamic inter-frame cumulative image to obtain image partition regions with spatial consistency, providing a regional processing basis for subsequent adaptive filtering and spectral analysis; Applying an adaptive filtering algorithm to each image partition region in the dynamic inter-frame cumulative image, dynamically adjusting the filtering parameters to suppress spatial domain random noise and preliminarily highlight microstructure defect features, and outputting an adaptive filtering processing image; Performing two-dimensional Fourier transform operation on each image partition region based on the adaptive filtering processing image, obtaining the frequency spectrum space mapping of each region, and extracting hierarchical microstructure spectral features in the frequency domain; The frequency spectrum space mapping obtained by two-dimensional Fourier transform is subjected to high-pass or low-pass band feature selection by using a saliency enhancement operator, thereby explicitly highlighting the frequency spectrum factor components related to the hierarchical defects, and generating a frequency spectrum domain enhanced image; The image subjected to adaptive filtering and the frequency spectrum domain enhanced image are subjected to pixel-level weighted fusion, and a spatial and frequency spectrum saliency enhanced image is obtained by using a spatial and frequency spectrum joint enhancement mechanism.

6. The pattern recognition-based copper-clad plate delamination defect diagnosis method of claim 1, wherein, The multi-scale feature enhancement includes calculation of gray scale distribution, texture coefficient and frequency spectrum energy statistical indexes of the spatial and frequency spectrum saliency enhanced image according to regions, adaptive setting of the size of a multi-scale processing window, the number of transform layers and threshold coefficients according to the region statistical indexes, matching of wavelet packet decomposition, Laplace pyramid and adaptive Retinex different enhancement algorithms, and extraction of multi-scale dynamic features of each region.

7. The method according to claim 2, wherein the method is characterized by: The multi-channel imaging system includes at least three channels of RGB visible light, near infrared and ultraviolet, and the exposure time, gain and aperture parameters of each channel can be automatically adjusted according to the spatial region, and the channels are synchronously locked by using a hardware-level time sequence.

8. The method according to claim 1, wherein the method is characterized by: In the multi-scale feature enhancement, the adaptive setting method of the region feature parameters includes gray scale variance, frequency spectrum energy and signal-to-noise ratio.

Images