Film lamination defect detection method and system based on image processing

Abstract

The invention relates to the technical field of image processing, and discloses a film lamination defect detection method and system based on image processing. The method comprises the following steps: acquiring and preprocessing a film-conducting layer interface by using a polarization enhanced imaging device to obtain an enhanced image; applying double-layer gradient edge detection to extract the edge of a fitting area, and marking a micro-defect sensitive area; an interface smoothness index and a fitting uniformity feature are obtained through multi-scale morphology feature extraction; inputting the features into a support vector classifier to identify bubbles-wrinkles; performing three-dimensional reconstruction on an identification result to generate a bubble distribution and stress field graph; and finally evaluating the fitting quality, and outputting an integrity score. According to the invention, automatic accurate detection and quantitative evaluation of tiny defects of the thin film-conducting layer bonding interface are realized.

Description

Technical Field

[0001] This application relates to the field of image processing technology, and in particular to a method and system for detecting film bonding defects based on image processing. Background Technology

[0002] With the rapid development of flexible electronics, OLED displays, and wearable devices, the bonding process between thin films and conductive layers plays an increasingly important role in electronic product manufacturing. Existing technologies for detecting defects in thin film bonding mainly include manual visual inspection, ultrasonic testing, X-ray imaging, and traditional machine vision techniques. Manual visual inspection relies on the inspector's experience, is inefficient, and is easily affected by subjective factors; while ultrasonic testing can detect internal bubbles, its resolution is limited and its detection speed is slow; X-ray imaging requires sophisticated equipment, is expensive, and poses radiation risks; traditional machine vision techniques mainly rely on grayscale contrast and simple threshold segmentation, typically employing a single image acquisition mode, supplemented by edge detection and morphological processing for image analysis. These methods, to a certain extent, meet production needs, especially showing good detection results for obvious large-area defects.

[0003] However, existing technologies have significant shortcomings in the detection and quantitative assessment of minute defects. First, traditional optical imaging methods struggle to capture tiny bubbles and wrinkles at the thin-film-conductive layer interface, especially defects smaller than 100 micrometers, which are often overlooked. Second, conventional image processing algorithms have limited ability to distinguish between different types of defects (such as bubbles and wrinkles), leading to a high false positive rate. Third, existing detection methods are mostly limited to two-dimensional planar analysis, unable to effectively reconstruct and quantify the three-dimensional morphology and depth information of defects. Finally, the lack of a systematic defect evaluation system makes it difficult to standardize and quantify the bonding quality, resulting in significant differences in results between different inspection equipment and operators, affecting continuous optimization of production processes and quality control. As electronic products develop towards higher precision and higher reliability, these technological limitations have become bottlenecks restricting product yield and performance improvement. Summary of the Invention

[0004] This application provides a method and system for detecting defects in thin film bonding based on image processing, which can be used to achieve automated and accurate detection and quantitative evaluation of minute defects (especially microbubbles and microwrinkles) at the interface of thin film-conductive layer bonding.

[0005] In a first aspect, this application provides a method for detecting thin film bonding defects based on image processing. The method includes: performing differential optical acquisition of the thin film-conductive layer interface using a polarization-enhanced imaging device, and preprocessing the original image using interface reflection enhancement filtering to obtain an interface-enhanced preprocessed image; extracting edges from the bonding area using a dual-layer gradient edge detection algorithm based on the interface-enhanced preprocessed image to obtain a micro-defect sensitive area marker map; extracting features from the segmented area using a multi-scale morphology feature extractor based on the micro-defect sensitive area marker map to obtain an interface smoothness index and bonding uniformity feature vector; inputting the bonding uniformity feature vector into a support vector classifier for bubble-wrinkle recognition to obtain an interface abnormal area mapping result; performing three-dimensional reconstruction of microbubbles and microwrinkles using a contour extraction algorithm based on the interface abnormal area mapping result to obtain an interlayer bubble distribution and wrinkle stress field map; and evaluating the thin film-conductive layer bonding quality using an interface integrity evaluation model based on the interlayer bubble distribution and wrinkle stress field map to obtain a bonding interface integrity score.

[0006] Secondly, this application provides an image processing-based thin-film bonding defect detection system, the image processing-based thin-film bonding defect detection system comprising: The acquisition module is used to perform differential optical acquisition of the thin film-conductive layer interface through a polarization-enhanced imaging device, and to preprocess the original image through interface reflection enhancement filtering to obtain an interface-enhanced preprocessed image. The analysis module is used to extract the edges of the bonding area based on the interface enhancement preprocessed image using a two-layer gradient edge detection algorithm, and obtain a micro-defect sensitive area marker map; The extraction module is used to extract features from the segmented region based on the micro-defect sensitive area marker map using a multi-scale morphology feature extractor, and obtain the interface smoothness index and the fitting uniformity feature vector. The recognition module is used to input the uniformity feature vector of the fit into the support vector classifier to identify bubbles and wrinkles, and obtain the mapping result of abnormal areas of the interface. The reconstruction module is used to perform three-dimensional reconstruction of microbubbles and microfolds based on the mapping results of abnormal interface areas and the contour extraction algorithm, so as to obtain the distribution of interlayer bubbles and the stress field map of folds. The evaluation module is used to evaluate the bonding quality of the thin film-conductive layer based on the interlayer bubble distribution and wrinkle stress field diagram using an interface integrity evaluation model, and obtain a bonding interface integrity score.

[0007] A third aspect of the present invention provides a computer device, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor invokes the instructions in the memory to cause the computer device to perform the above-described image processing-based thin-film bonding defect detection method.

[0008] A fourth aspect of the present invention provides a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the above-described image processing-based thin-film bonding defect detection method.

[0009] The technical solution provided in this application effectively improves the optical contrast and visibility of micro-defects at the interface by using a polarization-enhanced imaging device for differentiated optical acquisition combined with interface reflection enhancement filtering preprocessing. This solves the problem that traditional imaging is unable to capture small interface anomalies, especially achieving a detection rate of up to 95% for defects at the 30μm level. The dual-layer gradient edge detection algorithm organically integrates the Canny and Sobel operators and achieves effective extraction of edge information through adaptive threshold segmentation, significantly improving the detection sensitivity for irregular edges. This allows for the accurate marking of small defect areas, and the edge positioning accuracy is improved to the sub-pixel level. The multi-scale morphological feature extractor performs feature analysis on the segmented region, integrating geometric morphological parameters, gray-level statistical features, and wavelet decomposition features to construct a comprehensive defect feature description system. The interface smoothness index and the fitting uniformity feature vector can quantitatively characterize the severity of the defect. The feature extraction dimension reaches 32 dimensions, providing a rich feature space for accurate classification. The support vector classifier achieves accurate identification of bubble-wrinkle types based on artificial intelligence technology. Training with a large number of labeled samples, the model employs a nonlinear kernel function to map a multidimensional feature space and automatically optimizes the decision boundary, achieving a classification accuracy of over 94%. This demonstrates the crucial role of machine learning algorithms in defect identification. Furthermore, the maximum margin characteristic unique to the support vector machine model makes the classification results more robust, effectively addressing noise interference and feature fluctuations in the production environment. The contour extraction algorithm, combined with the depth mapping matrix, enables the 3D reconstruction of microbubbles and microwrinkles, transforming 2D image information into 3D spatial structures and providing key parameters such as defect volume and depth, offering significant guidance for stress analysis and quality prediction. The interface integrity evaluation model comprehensively considers various defect characteristics and indicators, establishing a standardized scoring mechanism that transforms complex defect information into an intuitive integrity score. This model integrates expert knowledge and data-driven intelligent evaluation methods, not only objectively evaluating the bonding quality of the current sample but also tracking process change trends, providing precise basis for parameter adjustments during production. This achieves closed-loop control of defect detection and process optimization, ultimately significantly improving the yield and consistency of thin-film bonded products. Attached Figure Description

[0010] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 This is a schematic diagram of an embodiment of the thin film bonding defect detection method based on image processing in this application. Figure 2 This is a schematic diagram of one embodiment of the thin film bonding defect detection system based on image processing in this application. Figure 3 This is a schematic block diagram of the structure of the computer device in an embodiment of the present invention. Detailed Implementation

[0012] This application provides a method and system for detecting film bonding defects based on image processing. The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0013] For ease of understanding, the specific process of the embodiments of this application is described below. Please refer to [link / reference]. Figure 1 One embodiment of the thin-film bonding defect detection method based on image processing in this application includes: Step S101: Differential optical acquisition of the thin film-conductive layer interface is performed by a polarization enhancement imaging device, and the original image is preprocessed by interface reflection enhancement filtering to obtain an interface enhancement preprocessed image. Step S102: Based on the interface enhancement preprocessed image, the edge of the bonding area is extracted using a dual-layer gradient edge detection algorithm to obtain a micro-defect sensitive area marker map; Step S103: Based on the micro-defect sensitive area marker map, the segmented region is extracted using a multi-scale morphology feature extractor to obtain the interface smoothness index and the bonding uniformity feature vector. Step S104: Input the bonding uniformity feature vector into the support vector classifier to perform bubble-wrinkle recognition and obtain the interface abnormal area mapping result. Step S105: Based on the mapping results of the interface anomaly region, the microbubbles and microfolds are reconstructed in three dimensions using the contour extraction algorithm to obtain the interlayer bubble distribution and fold stress field map. Step S106: Based on the interlayer bubble distribution and wrinkle stress field diagram, evaluate the bonding quality of the thin film-conductive layer using the interface integrity evaluation model to obtain the bonding interface integrity score.

[0014] It is understood that the executing entity of this application can be an image processing-based thin film bonding defect detection system, or it can be a terminal or a server; no specific limitation is made here. This application's embodiments use a server as an example for illustration.

[0015] Specifically, when performing differential optical acquisition of the thin-film-conductive layer interface, the polarization-enhanced imaging device uses a polarized light source with a 45° polarization angle to illuminate the sample surface. This angle maximizes the reflection contrast at the thin-film-conductive layer interface. A line scan camera captures changes in reflected light intensity at a set scanning frequency to obtain the original image of the interface reflection. Gaussian filtering is first applied to the original image to eliminate random noise; the filter kernel size is typically set to 5×5 pixels, smoothing the image while preserving edge information. Adaptive histogram equalization is then performed, dividing the image into 8×8 sub-blocks. Histogram equalization is performed individually within each sub-block to extract the interface reflection contrast features. The interface reflection enhancement filter uses a Gabor filter bank with eight filters in different directions and a frequency parameter of 0.5 to enhance the brightness variation characteristics at the interface between the thin film and the conductive layer. In the geometric correction and compensation stage, bilinear interpolation is used to eliminate lens distortion and shooting angle deviation. Finally, a nonlinear contrast adjustment function is used for dynamic range expansion to obtain the interface-enhanced preprocessed image.

[0016] When extracting edges in the bonding region, the dual-layer gradient edge detection algorithm first applies the Canny and Sobel operators in parallel for primary edge extraction. The Canny operator includes four steps: Gaussian filtering, gradient calculation, non-maximum suppression, and dual-threshold detection, suitable for detecting fine edges; the Sobel operator calculates horizontal and vertical gradients, suitable for detecting strong edges. The two sets of edge images are fused using a logical AND operation, retaining edge points detected by both operators. The adaptive threshold segmentation algorithm calculates the optimal segmentation threshold using the Otsu method, dividing the image into edge and non-edge regions. Dual-layer gradient calculation calculates the first and second gradients at each pixel, recording the gradient directions to form a gradient feature map. Region growing starts from the detected edge starting point and expands the region according to gradient direction and gradient similarity, connecting discontinuous edges less than 5 pixels apart. Watershed segmentation marking uses the complete edge contour as the boundary line of the watershed algorithm to divide the film bonding region, conductive layer region, and interface region, obtaining a micro-defect sensitive region marking map. When extracting features from segmented regions, the multi-scale morphology feature extractor first assigns a unique ID to each region in the micro-defect sensitive region labeling map, establishing a region lookup table. Geometric morphological parameters are calculated for each region, including area (total number of pixels within the region), perimeter (number of boundary point sets), equivalent radius, and shape factor. Gray-level histograms are extracted, and first-order statistics (mean, standard deviation, skewness, and kurtosis) and second-order statistics (energy, contrast, correlation, and entropy) are calculated. The interface smoothness index is obtained by fitting the interface line with a quadratic polynomial, calculating the curvature, and integrating the curvature sequence. Wavelet decomposition uses Haar wavelets to perform a three-level decomposition of the region image, extracting low-frequency and high-frequency components, and constructing a multi-scale spectral feature library. The interface line is divided into 20 equal-length segments, and the local smoothness index is calculated. A weighted average is used to synthesize the fitting uniformity index. Feature dimensionality reduction uses principal component analysis to reduce the original feature space to 32 dimensions and performs normalization to obtain the fitting uniformity feature vector.

[0017] When using a support vector classifier for bubble-wrinkle recognition, the feature vectors are first standardized using Z-score to ensure uniformity. An SVM classifier with an RBF kernel is then built, optimizing the penalty and kernel parameters. A scalar distance between the sample and the classification boundary is calculated, where the normal vector and bias term are obtained from classifier training. Two-stage classification first uses a threshold to categorize regions as normal or abnormal, then uses another threshold to further categorize abnormal regions as bubble-type defects or wrinkle-type defects. Confidence scores reflect classification reliability, and a threshold of 0.75 filters out low-confidence regions. High-reliability recognition results are mapped back to the original image using the original coordinates, with bubble regions marked in red and wrinkle regions marked in yellow, forming an interface abnormality region mapping result. When using a contour extraction algorithm for 3D reconstruction of microbubbles and microwrinkles, bubble defect regions and wrinkle defect regions are first separated through connected component analysis. A contour tracking algorithm is used to extract bubble boundary segments, calculating the bubble area and roundness. A depth mapping matrix is ​​constructed based on gray-level gradient information, establishing the correspondence between gray-level values ​​and depth. The skeleton extraction algorithm obtains the fold centerline through morphological refinement and measures the width distribution perpendicular to the centerline. The thin plate model calculates the stress distribution and establishes a mapping relationship between local deformation and stress. Bubble location and volume data are integrated into a three-dimensional spatial distribution model, and the fold stress value matrix data is expressed using a heatmap, with red representing high-stress areas and blue representing low-stress areas.

[0018] When evaluating the bonding quality of the thin film-conductive layer using the interface integrity evaluation model, the following values ​​are extracted from the interlayer bubble distribution: bubble density (number of bubbles per unit area), bubble size distribution (frequency distribution of bubbles of each size), maximum bubble diameter, and total bubble volume. The following values ​​are extracted from the wrinkle stress field diagram: total wrinkle length, average wrinkle width, peak wrinkle stress, and wrinkle direction consistency (proportion of wrinkle length to total length in the main direction). A mapping relationship between these features and the integrity score is established, and a weighted defect score is calculated by balancing the weights of bubble and wrinkle factors. The severity level of the defect is determined based on the weighted defect score: 90≤Q≤100 is excellent, 80≤Q<90 is good, 60≤Q<80 is medium, and Q<60 is poor, ultimately yielding the bonding interface integrity score.

[0019] For example, during the lamination process of a flexible substrate film for an OLED display, a polarization-enhanced imaging device acquires the original image at a 45° polarization angle, with an image resolution of 4096×3072 pixels. After Gaussian filtering and adaptive histogram equalization, the interface contrast is significantly improved. The Gabor filter bank enhances the interface line features. In the dual-layer gradient detection, the low threshold of the Canny operator is set to 50, and the high threshold is set to 150; the threshold of the Sobel operator is set to 40. Three main lamination regions are detected, among which region #2 has a shape factor of 0.78 and an interface smoothness index of 2.34, indicating an anomaly. After feature dimensionality reduction, the data is input into a support vector classifier, which identifies a cluster of microbubbles in region #2, with a maximum bubble diameter of 86 micrometers and a depth of 12 micrometers; and a micro-wrinkle in region #3 with a length of 234 micrometers and an average width of 25 micrometers. After comprehensive evaluation, the bubble factor score is 15, the wrinkle factor score is 12, and the lamination interface integrity score is 79, which belongs to the medium lamination quality, requiring adjustment of lamination process parameters to improve the lamination effect.

[0020] In this embodiment, differential optical acquisition using a polarization-enhanced imaging device combined with interface reflection enhancement filtering preprocessing effectively improves the optical contrast and visibility of micro-defects at the interface, solving the problem that traditional imaging is unable to capture small interface anomalies, especially achieving a detection rate of up to 95% for defects at the 30μm level; the dual-layer gradient edge detection algorithm organically integrates the Canny and Sobel operators and achieves effective extraction of edge information through adaptive threshold segmentation, significantly improving the detection sensitivity for irregular edges, enabling precise marking of small defect areas, and improving edge localization accuracy to the sub-pixel level; the multi-scale morphological feature extractor performs feature analysis on the segmented region, integrating geometric morphological parameters, gray-level statistical features, and wavelet decomposition features to construct a comprehensive defect feature description system, where the interface smoothness index and fitting uniformity feature vector can quantitatively characterize the severity of defects, and the feature extraction dimension reaches 32 dimensions, providing a rich feature space for accurate classification; the support vector classifier achieves accurate identification of bubble-wrinkle types based on artificial intelligence technology, and this classifier... The model, trained on a large number of labeled samples, employs a nonlinear kernel function to map a multidimensional feature space and automatically optimizes the decision boundary, achieving a classification accuracy of over 94%. This demonstrates the crucial role of machine learning algorithms in defect identification. Furthermore, the maximum margin characteristic unique to the support vector machine model makes the classification results more robust, effectively addressing noise interference and feature fluctuations in the production environment. The contour extraction algorithm, combined with the depth mapping matrix, enables the 3D reconstruction of microbubbles and microwrinkles, transforming 2D image information into 3D spatial structures and providing key parameters such as defect volume and depth, offering significant guidance for stress analysis and quality prediction. The interface integrity evaluation model comprehensively considers various defect features and indicators, establishing a standardized scoring mechanism that transforms complex defect information into an intuitive integrity score. This model integrates expert knowledge and data-driven intelligent evaluation methods, not only objectively evaluating the bonding quality of the current sample but also tracking process change trends, providing precise basis for parameter adjustments during production. This achieves closed-loop control of defect detection and process optimization, ultimately significantly improving the yield and consistency of film-bonded products.

[0021] In one specific embodiment, the process of performing step S101 may specifically include the following steps: (1) By irradiating the thin film-conductive layer interface with a polarized light source with a polarization angle of 45°, the change in the intensity of reflected light is captured by a line scan camera to obtain the original image of the interface reflection; (2) Based on the original image of the interface reflection, random noise is eliminated by Gaussian filtering algorithm to generate a noise-reduced image; (3) Apply adaptive histogram equalization to the noise-reduced image, extract the interface reflection contrast features, and form a contrast-enhanced image; (4) The contrast enhancement image is fed into the interface reflection enhancement filter to enhance the brightness change characteristics at the junction of the thin film and the conductive layer and generate an interface layer protrusion map. (5) Perform geometric correction and compensation on the interface hierarchy highlighting diagram to eliminate lens distortion and shooting angle deviation, and form a geometrically accurate interface diagram; (6) The geometrically accurate interface map is dynamically expanded using a nonlinear contrast adjustment function to obtain an interface enhancement preprocessed image.

[0022] Specifically, the process involves illuminating the thin-film-conductive layer interface with a polarized light source set at a polarization angle of 45°, and then capturing the changes in reflected light intensity using a line scan camera to obtain the original image of the interface reflection. The polarized light source generates light waves with a specific vibration direction; the 45° polarization angle is chosen to maximize the contrast between reflected and transmitted light at the thin-film and conductive layer interface. When polarized light illuminates the thin-film-conductive layer interface, the difference in refractive index between the two materials causes specific reflection and refraction phenomena at the interface. These optical properties are directly related to the interface bonding quality. The line scan camera is equipped with a polarization filter matched to the light source, typically set to a scanning frequency of 4000 lines / second and a resolution of 4096 pixels / line. This allows it to capture the changes in reflected light intensity line by line, forming an original image matrix containing interface reflection information. Based on the original image of the interface reflection, a Gaussian filtering algorithm is used to eliminate random noise, generating a denoised image. The Gaussian filtering algorithm uses a two-dimensional Gaussian function as the filter kernel to perform convolution operations on the image. The kernel size is typically set to 5×5 pixels, and the standard deviation is set to 1.2. The Gaussian function exhibits a bell-shaped distribution in the spatial domain, with the largest weight at the center and decreasing towards the edges. This characteristic allows it to effectively smooth image noise while preserving edge information. Specifically, the Gaussian filter kernel is slid across the original image pixel by pixel, calculating the weighted sum of pixel values ​​within the kernel's coverage area and their corresponding kernel weights. This sum replaces the value of the central pixel. After this processing, the value of each pixel is replaced by the weighted average of its neighboring pixels, smoothing out random noise and resulting in a denoised image.

[0023] Adaptive histogram equalization is applied to the denoised image to extract interface reflection contrast features, resulting in a contrast-enhanced image. Adaptive histogram equalization divides the image into 8×8 sub-blocks, performing histogram equalization independently within each sub-block. Then, bilinear interpolation is used to synthesize the complete image, avoiding the over-enhancing contrast problem that may occur with global histogram equalization. For each sub-block, a gray-level histogram is first calculated to represent the distribution of gray values. Then, a cumulative distribution function is calculated to map the original gray values ​​to new gray values, achieving contrast enhancement. Bilinear interpolation maintains a smooth transition between adjacent sub-blocks, avoiding block artifacts and ultimately forming a contrast-enhanced image.

[0024] The contrast-enhanced image is fed into an interface reflection enhancement filter to amplify the brightness variation characteristics at the interface between the thin film and the conductive layer, generating an interface layer highlighting map. The interface reflection enhancement filter employs a Gabor filter bank with eight filters in different directions, a frequency parameter of 0.5, and a bandwidth parameter of 1.0. Gabor filters are bandpass filters with good directional and frequency selectivity, suitable for detecting texture features at specific directions and scales. The convolution result of the input image with each Gabor filter in each direction is calculated, resulting in eight response images. These response images are then fused using a maximum value operation to enhance the brightness variation characteristics at the interface between the thin film and the conductive layer, generating the interface layer highlighting map. Geometric correction compensation is performed on the interface layer highlighting map to eliminate lens distortion and shooting angle deviation, forming a geometrically accurate interface map. Geometric correction compensation establishes a distortion model based on the coordinates of control points on a calibration board, including radial and tangential distortion. Radial distortion is caused by the lens shape, making straight lines in the image appear curved; tangential distortion is caused by the non-parallelism between the lens and the image sensor, leading to local image deformation. The distortion model is used to calculate the position of each pixel in the original image mapped to the corrected image, and a grayscale value is assigned to each pixel in the corrected image using bilinear interpolation. Simultaneously, a perspective transformation matrix is ​​used to correct for shooting angle deviations, ensuring that interface features present true geometric relationships in the image, thus forming a geometrically accurate interface map.

[0025] The geometrically accurate interface map is dynamically expanded using a nonlinear contrast adjustment function to obtain an interface-enhanced preprocessed image. This function employs an S-curve transform to stretch mid-tone grayscale regions and compress bright and dark areas, enhancing midtone details while avoiding loss of detail in highlights and shadows. The processed image exhibits an expanded dynamic range, highlighting subtle features of the thin-film-conductive layer interface, resulting in the final interface-enhanced preprocessed image.

[0026] For example, in the thin-film bonding inspection of an OLED display panel, the sample is first illuminated with a 45° polarized light source. A line scan camera acquires the original image at a speed of 4000 lines / second, with a resolution of 4096×3072 pixels and 16 bits of grayscale per pixel, recording the intensity distribution of reflected light at the thin-film-conductive layer interface. Next, a 5×5 Gaussian filter kernel is applied to the original image for noise reduction, effectively suppressing random noise introduced by the camera sensor and ambient light fluctuations, significantly improving the signal-to-noise ratio. Then, the image is divided into 8×8 sub-blocks for adaptive histogram equalization, each sub-block being 512×384 pixels. By calculating the grayscale distribution and cumulative distribution function of each sub-block, local contrast enhancement is achieved, making the interface reflection characteristics more prominent. Subsequently, an 8-directional Gabor filter bank is applied to process the image. The 8 filtered response images are then fused using a maximum value operation, significantly enhancing the brightness gradient value at the interface between the thin film and the conductive layer. Geometric correction is performed on the enhanced image to eliminate radial and tangential distortion, significantly reducing the positional deviation of the interface line in the image after correction. Finally, a nonlinear S-curve transform was used to expand the dynamic range of the image, outputting a 16-bit grayscale interface enhancement preprocessed image, laying the foundation for subsequent two-layer gradient edge detection. By comparing the images before and after processing, it was observed that the signal-to-noise ratio of the thin film-conductive layer interface in the preprocessed image was significantly improved, the edge sharpness was significantly enhanced, and the contrast of microbubbles and wrinkled areas was enhanced, greatly improving the accuracy and sensitivity of subsequent defect detection.

[0027] In one specific embodiment, the process of performing step S102 may specifically include the following steps: (1) Apply the Canny operator and the Sobel operator to the interface enhancement preprocessed image to perform primary edge extraction, and obtain two sets of edge images; (2) The two sets of edge images are fused by logical AND operation to generate preliminary edge detection results; (3) An adaptive threshold segmentation algorithm is used to process the preliminary edge detection results. A dynamic threshold is set to control the edge detection sensitivity and obtain the edge candidate region. (4) Apply two-layer gradient calculation to the edge candidate region to extract gradient direction and gradient magnitude features and form a gradient feature map; (5) Perform region growing on the gradient feature map, connect discontinuous edges, and construct a complete edge contour; (6) Divide the complete edge contour into watersheds, and divide the film bonding area, conductive layer area and interface area to obtain a micro-defect sensitive area marking map.

[0028] Specifically, when applying the Canny and Sobel operators to perform primary edge extraction on preprocessed images for interface enhancement, it's essential to first understand the working principles and characteristics of these two operators. The Canny operator comprises four steps: Gaussian filtering, gradient calculation, non-maximum suppression, and dual-threshold detection. It can detect fine edge structures and has strong noise resistance. In practical applications, Gaussian filtering uses a 5×5 kernel to smooth the image, then calculates the gradients in the horizontal and vertical directions, further removes non-edge points using non-maximum suppression, and finally uses a dual-threshold method (typically a low threshold of 50 and a high threshold of 150) to determine true edges. The Sobel operator is relatively simpler, using two 3×3 convolution kernels to calculate the gradients in the horizontal and vertical directions, then calculating the gradient magnitude and direction, and finally using a single threshold (typically set to 40) to determine whether a pixel is an edge. Each operator has its advantages: Canny is suitable for detecting fine edges, while Sobel is better suited for detecting strong edges. Combining the two can comprehensively capture edge information in the image. When fusing two sets of edge images using a logical AND operation to generate the preliminary edge detection result, pixel-level logical AND operations are performed on the edge images generated by Canny and Sobel respectively. The logical AND operation requires that corresponding pixels in both images be marked as edges (value 1) before the corresponding pixels in the resulting image are marked as edges; if a pixel at a corresponding position in either image is not an edge (value 0), then that position is also not an edge in the resulting image. This fusion method preserves strong edges recognized by both operators and filters out points that might be noise or false edges detected by a single operator, thereby improving the reliability of edge detection. The preliminary edge detection result is a binary image containing preliminary contour information of the thin-film-conductive layer interface region.

[0029] An adaptive thresholding algorithm is used to segment the initial edge detection results. A dynamic threshold is set to control the edge detection sensitivity. During the process of obtaining candidate edge regions, the Otsu method is used to automatically determine the optimal segmentation threshold. The Otsu method is an adaptive thresholding algorithm based on the image's gray-level histogram, which determines the optimal threshold by maximizing the inter-class variance. The processing includes: calculating the normalized gray-level histogram of the image, representing the proportion of each gray-level pixel in the total pixels; for each possible threshold, calculating the inter-class variance between the foreground (gray-level values ​​greater than the threshold) and the background (gray-level values ​​less than or equal to the threshold); and selecting the threshold that maximizes the inter-class variance as the optimal threshold. The adaptive nature is reflected in the fact that the threshold is dynamically adjusted according to the gray-level distribution of the input image, rather than using a fixed threshold. The initial edge detection results are then subjected to a secondary thresholding process using the optimal threshold to obtain candidate edge regions. A two-layer gradient calculation is applied to edge candidate regions to extract gradient direction and magnitude features, forming a gradient feature map. First, the first-order gradient (derivatives in the horizontal and vertical directions) is calculated on the original interface enhancement preprocessed image using the Sobel or Prewitt operators. Then, the gradient magnitude and direction are calculated. Next, the gradient is calculated again on the first-order gradient map to obtain the second-order gradient and its direction. The two-layer gradient feature fusion combines the first-order and second-order gradient information using a weighted combination, with the first-order gradient weight typically set to 0.7 and the second-order gradient weight set to 0.3. The gradient feature map contains the magnitude and direction information of the edges, providing crucial information for subsequent edge connectivity and region segmentation.

[0030] When performing region growing on the gradient feature map to connect discontinuous edges and construct a complete edge contour, a seed point is first determined, typically the edge point with the largest local gradient magnitude. Then, starting from the seed point, the region expands along both sides of the edge according to the gradient direction, checking whether adjacent pixels meet the growth conditions. These conditions include gradient direction similarity and a gradient magnitude threshold: the gradient direction of adjacent pixels differs from the gradient direction of the current edge point by less than a certain angle (usually 20 degrees), and the gradient magnitude is greater than a set threshold (usually 30% of the total gradient magnitude). Pixels meeting these conditions are added to the edge set as new expansion points. When the distance between two edge endpoints is less than a preset threshold (usually 5 pixels), they are connected by straight line or curve interpolation to form a continuous edge. After region growing, a complete edge contour is obtained. In the process of dividing the complete edge contour into thin film bonding area, conductive layer area, and interface area to obtain the micro-defect sensitive area marker map, the complete edge contour is first used as the boundary line of the watershed algorithm. Then, the distance from each non-edge pixel to the nearest edge pixel is calculated through distance transformation to form a distance map. Local maximum points are found in the distance map as marker points, and each marker point represents a region. The marker map is then compared with the boundary. Figure 1The input watershed algorithm expands the region from the marked point based on the gradient information of the image until adjacent regions meet or reach the boundary. Finally, based on the region features (such as location, shape, gray-level statistics, etc.), the segmented region is classified into thin film bonding region, conductive layer region and interface region, generating a micro-defect sensitive region marker map with different markers.

[0031] For example, in the thin-film bonding detection of an OLED display panel, after acquiring a preprocessed interface enhancement image with a resolution of 4096×3072 pixels, edge extraction is first performed using the Canny and Sobel operators respectively. The Canny operator uses a Gaussian kernel for smoothing, with a low threshold set to 50 and a high threshold set to 150, detecting delicate edge structures; the Sobel operator uses a threshold of 75 to detect the main strong edges. The two sets of edge images are fused using a logical AND operation, retaining the edges detected by both operators and effectively filtering out false edges generated by independent operators. The Otsu method is applied to the fused preliminary edge results to calculate the optimal threshold, resulting in a dynamic threshold of 82. The edge grayscale values ​​are then further segmented to form edge candidate regions. A two-layer gradient calculation is applied to the edge candidate regions. The first-order gradient is calculated using the Sobel operator, and the second-order gradient is calculated by applying the Sobel operator again based on the first-order gradient, resulting in a combined gradient feature map. When performing region growing on the gradient feature map, the point with the largest local gradient magnitude was selected as the initial seed. The growth conditions were set to a gradient direction difference of less than 20 degrees and a gradient magnitude greater than a set threshold. This connected multiple discontinuous edges, constructing a complete edge contour. Finally, a watershed algorithm was applied to the complete edge contour, using the edge as the watershed line and the internal region as a catchment basin. Multiple regions were marked and classified into thin film bonding areas, conductive layer areas, and interface areas based on their location and morphological features. In the marking map, the thin film bonding area is marked in red, the conductive layer area in green, and the interface area in blue. Analyzing the characteristics of the interface area revealed abnormal edge curvature and poor edge continuity in some areas, providing precise localization of micro-defect sensitive areas for subsequent multi-scale morphological feature extraction.

[0032] In one specific embodiment, the process of executing step S103 may specifically include the following steps: (1) Number and index the micro-defect sensitive area marking map, assign a unique identifier to each area, and establish an area retrieval table; (2) Read the boundary point set of each segmented region sequentially from the region retrieval table, and calculate the geometric morphological parameters of the region, specifically including area, perimeter, equivalent radius and shape factor; (3) For each segmented region, extract its gray-level histogram and calculate the first-order and second-order statistics to form a gray-level distribution descriptor; (4) Apply the curvature calculation algorithm to the thin film-conductive layer interface region, calculate the curvature value of each point on the interface line by fitting a quadratic polynomial, and perform integral calculation on the curvature value sequence on the interface line to obtain the interface smoothness index. (5) Wavelet decomposition is applied to each segmented region to extract low-frequency and high-frequency components at different scales in sequence, and a multi-scale spectral feature library is constructed. (6) Divide the thin film-conductive layer interface region into a fixed number of equal-length sub-segments, calculate the local smoothness index of each sub-segment, and synthesize the overall bonding uniformity index by weighted average; (7) Integrate all extracted morphological features, statistical features and multi-scale features, perform feature dimensionality reduction and normalization, and generate a fixed-dimensional fitting uniformity feature vector.

[0033] Specifically, the micro-defect sensitive area marker map is numbered and indexed, and a unique identifier is assigned to each area. During the process of building the area retrieval table, the micro-defect sensitive area marker map is scanned using a connected component analysis algorithm to identify all independent regions composed of interconnected pixels. Connected component analysis typically employs a two-pass scanning method. The first pass scans the image from left to right and from top to bottom, assigning temporary labels to each non-background pixel and recording the equivalence relationships between adjacent labels. The second pass scans the image again, merging equivalent labels to ensure that each connected component has a unique identifier. In this way, the thin film bonding area, conductive layer area, and interface area in the marker map each obtain different area IDs, which are stored in the area retrieval table. This table contains key information such as area ID, area type label, area center coordinates, and boundary coordinate sequence, providing a query basis for subsequent feature extraction. The boundary point set of each segmented region is sequentially read from the area retrieval table, and the geometric morphological parameters of the region are calculated, specifically including area, perimeter, equivalent radius, and shape factor. Each numbered region is processed one by one. The area is calculated by counting the total number of pixels within the region, while the perimeter is obtained by obtaining the total number of boundary pixels using a boundary tracking algorithm. The equivalent radius is the radius value when the region is assumed to be a perfect circle, calculated through area conversion. The shape factor is a dimensionless parameter describing the regularity of the region's shape, calculated through the relationship between area and perimeter; the closer the value is to 1, the closer the shape is to a circle, and the smaller the value, the more irregular the shape. These geometric morphological parameters together constitute a quantitative description of the region's shape, effectively identifying anomalous shape regions, such as irregular bubbles or wrinkles.

[0034] For each segmented region, its grayscale histogram is extracted, and first-order and second-order statistics are calculated to form a grayscale distribution descriptor. First, the grayscale values ​​of the corresponding region in the original interface enhancement preprocessed image are read, and the frequency of pixels at each grayscale level is counted to generate a grayscale histogram. First-order statistics include the mean (arithmetic mean of all pixel grayscale values ​​within the region), standard deviation (degree of grayscale value dispersion), skewness (degree of deviation of grayscale distribution from symmetry), and kurtosis (degree of sharpness of grayscale distribution). Second-order statistics are calculated based on the grayscale co-occurrence matrix and include energy (texture uniformity), contrast (degree of grayscale difference), correlation (degree of grayscale correlation), and entropy (randomness of grayscale distribution). These statistical features can comprehensively describe the pixel grayscale distribution within the region, providing important evidence for defect identification.

[0035] A curvature calculation algorithm is applied to the thin-film-conductive layer interface region. The curvature value of each point on the interface line is calculated using quadratic polynomial fitting, and the curvature value sequence on the interface line is integrated to obtain the interface smoothness index. First, the boundary point set of the interface region is extracted from the region lookup table and reordered according to spatial location to form an ordered interface line. For each point on the interface line and its neighborhood points, a quadratic polynomial curve is fitted using the least squares method to obtain curve parameters. The curvature value at each point on the fitted curve is calculated. The magnitude of the curvature value reflects the degree of bending of the interface at that point; areas with large curvature changes often correspond to locations with poor adhesion. By integrating the curvature value sequence across the entire interface line, the interface smoothness index, which characterizes the overall smoothness of the interface, is obtained. A larger value indicates a less smooth interface and a higher probability of defects. Wavelet decomposition is applied to each segmented region to sequentially extract low-frequency and high-frequency components at different scales. In the process of constructing a multi-scale spectral feature library, discrete wavelet transform is used to perform multi-level decomposition of the region image. Wavelet decomposition divides the image into low-frequency approximate components and high-frequency detail components in three directions (horizontal, vertical, and diagonal). A three-level decomposition is typically employed to generate multi-scale representations from coarse to fine. Low-frequency components reflect the overall grayscale distribution and large-scale structure of the region, while high-frequency components capture detailed features such as edges and textures. By extracting feature parameters such as energy, variance, and mean of each component, a complete multi-scale spectral feature library is constructed, enabling the analysis of regional characteristics at different scales and effectively identifying defects of various sizes.

[0036] The interface region between the thin film and the conductive layer is divided into a fixed number of equal-length segments. The local smoothness index (LSI) of each segment is calculated, and the overall bonding uniformity index is synthesized by weighted averaging. First, the interface line is divided into N segments (usually N=20) according to their arc length. For the j-th segment, the LSI_j is calculated using the following formula:

[0037] in and This represents the start and end arc length parameters of the j-th sub-segment. This represents the curvature value of the interface line at the arc length parameter s. This represents the standard deviation of the curvature values ​​within that segment. , and These are weighting coefficients, which control the importance of integral curvature, maximum curvature, and curvature variance in the evaluation, and their values ​​are typically [values ​​to be filled in]. =0.5、 =0.3、 =0.2. The overall fit uniformity index is obtained by weighted averaging of the local smoothness index of each segment. The calculation formula is:

[0038] in This indicates the uniformity bonding index. is the weight coefficient of the j-th sub-segment, which is related to the position and importance of the sub-segment. Generally, the weight of the central region is higher than that of the edge region.

[0039] In the process of integrating all extracted morphological, statistical, and multi-scale features, and performing feature dimensionality reduction and normalization to generate a fixed-dimensional, uniformly fitting feature vector, the various features are first organized into high-dimensional feature vectors according to a predefined format. Since different features have different dimensions and numerical ranges, normalization is required, typically using Z-score normalization or Min-Max normalization to map all feature values ​​to similar numerical ranges. Then, principal component analysis (PCA) is used for feature dimensionality reduction, retaining principal components that account for more than 95% of the total variance, usually compressing the original hundreds of dimensions to around 32 dimensions. Finally, the features are reorganized according to a predefined sorting rule to generate a fixed-dimensional, uniformly fitting feature vector, which serves as input to the subsequent support vector classifier.

[0040] For example, in the detection of thin-film bonding defects in a flexible OLED display panel, 28 regions were identified in the micro-defect sensitive area marker map, including 10 thin-film bonding areas, 8 conductive layer areas, and 10 interface areas. Each region was assigned a unique ID from 1 to 28. Geometric morphology parameter calculations were performed on the interface region with ID 15, revealing an area of ​​12468 pixels, a perimeter of 842 pixels, an equivalent radius of 63.1 pixels, and a shape factor of 0.22. This is significantly lower than the shape factor of normal regions (typically greater than 0.6), indicating an irregular shape and potential defects. Extracting the grayscale histogram of this region yielded a mean of 127.3, a standard deviation of 42.8, a skewness of 0.67, and a kurtosis of 3.24, showing significant differences from the statistical characteristics of normal regions. A quadratic polynomial fitting was performed on the interface line of this region. A maximum curvature of 5.7 was found at arc length s=120, significantly higher than the average curvature of 1.2. The integral yielded an interface smoothness index of 3.82, exceeding the warning threshold of 3.0. The interface line was divided into 20 equal-length segments. The 12th segment had the highest local smoothness index of 6.14, with a weight of 1.5. The weights of the other segments were all 1.0. The calculated uniformity fit index (UBI) was 2.76, significantly higher than the average of 0.95 in the normal fit area. Wavelet level 3 decomposition revealed that the high-frequency energy in the vertical direction of this region was three times higher than that of the normal region, indicating a structural anomaly in the vertical direction. Integrating all features and performing dimensionality reduction using PCA, a 32-dimensional fit uniformity feature vector was obtained, which showed high similarity to bubble-like defect features in the sample library.

[0041] In one specific embodiment, the process of executing step S104 may specifically include the following steps: (1) The uniformity of the fit feature vector is standardized by unifying the dimensions and normalizing the values ​​to generate a normalized feature vector; (2) Establish a sample training matrix based on the normalized feature vectors, optimize the support vector classifier parameters by gradient descent, and construct the feature space decision boundary; (3) Input the fitting uniformity feature vector of the sample to be detected into the trained support vector classifier and calculate the distance scalar between the sample and the classification boundary; (4) The sample area is classified in two stages according to the distance scalar size. First, the area is divided into normal area and abnormal area. Then, the abnormal area is further subdivided into bubble-type defects and wrinkle-type defects. (5) Calculate the confidence score for each identified defect area, and set a threshold to filter low confidence areas and retain high reliability identification results; (6) Map the high-reliability identification results back to the original image according to the original coordinate position, generate area markers with different color codes, and form the interface abnormal area mapping results.

[0042] Specifically, the standardization process unifies the dimensions and normalizes the values ​​of the fitting uniformity feature vectors to generate normalized feature vectors. This process first requires processing the values ​​from different measurement methods, which differ significantly in dimension and range. Common standardization methods include Z-score standardization and min-max standardization. Z-score standardization subtracts the mean of each feature value and divides it by the standard deviation, resulting in a feature distribution with a mean of 0 and a standard deviation of 1. Min-max standardization maps feature values ​​to a fixed interval (usually [0,1] or [-1,1]). Standardization ensures that features of different scales receive equal attention in the classifier, preventing certain features with larger values ​​from dominating the classification process. In the steps of building a sample training matrix based on the normalized feature vectors, optimizing the support vector classifier parameters using gradient descent, and constructing the feature space decision boundary, a large number of sample feature vectors of known categories (normal, bubble-type defects, wrinkled defects) are first collected to form the training matrix. Support Vector Classifiers (SVCs) are supervised learning algorithms that find the optimal hyperplane to separate samples of different classes in the feature space. The core task is to find the maximum margin hyperplane, maximizing the distance from each class of sample point to the hyperplane. Gradient descent iteratively optimizes the classifier parameters, including the penalty coefficient and kernel function parameters. The penalty coefficient controls the model's complexity and error tolerance; the kernel function maps the original features to a high-dimensional space, transforming linearly inseparable problems into linearly separable ones. Commonly used kernel functions include linear kernels, polynomial kernels, and radial basis function kernels. For thin-film bonding defect detection, radial basis function kernels typically perform best.

[0043] The fitting uniformity feature vector of the sample to be tested is input into a trained support vector classifier. When calculating the distance scalar between the sample and the classification boundary, the feature vector of the sample to be tested is first standardized in the same way as the training data, and then fed into the trained classifier. The classifier calculates the distance from the sample to the decision boundary, i.e., the decision function value. This distance scalar not only indicates which class the sample belongs to, but also reflects the degree of certainty of the classification: the farther the distance to the boundary, the more reliable the classification result. For multi-class classification problems, a one-to-one or one-to-many strategy is usually adopted, and the results of each binary classifier are finally combined to determine the class of the sample. A two-stage classification of the sample region is performed based on the distance scalar. First, the region is divided into normal regions and abnormal regions. Then, the abnormal regions are further subdivided into bubble-type defects and wrinkle-type defects. Two thresholds are set for classification judgment. In the first stage, the threshold is used to divide the sample region into two major categories: normal and abnormal. Regions with a distance scalar exceeding the threshold are considered normal regions, and regions with a distance scalar below the threshold are judged as abnormal regions. In the second stage, the identified anomalous regions are further subdivided. Using a specially trained bubble-wrinkle binary classifier or a secondary threshold based on a distance scalar, the anomalous regions are further classified into bubble-type defects or wrinkle-type defects. This two-stage classification strategy effectively improves classification accuracy, especially for classification problems with a clear hierarchical structure.

[0044] For each identified defective region, a confidence score is calculated, and a threshold is set to filter out low-confidence regions, retaining only high-reliability identification results. The confidence score is calculated based on the distance between the sample and the decision boundary. The larger the absolute value of the distance, the higher the confidence; a distance close to zero indicates that the sample is near the decision boundary, and the reliability of the classification result is low. Typically, the distance value is mapped to a range of 0 to 1 as the confidence score, and then a threshold (e.g., 0.75) is set for filtering, retaining only identification results with a confidence level higher than the threshold. This filtering mechanism effectively reduces the risk of false positives and improves the reliability of the detection results.

[0045] The high-reliability identification results are mapped back to the original image according to their original coordinates, generating color-coded region markers. During this process of mapping interface anomaly regions, the previously established region lookup table is used to establish a correspondence between the classification results and the original regions. Normal regions are typically marked in green, bubble-type defect regions in red, and wrinkle-type defect regions in yellow. Each region is colored according to its original coordinates and overlaid on the original image to form an intuitive mapping result of interface anomaly regions, facilitating rapid location and resolution of problems by process engineers.

[0046] For example, in the OLED display film bonding quality inspection of a smartwatch, after obtaining a 32-dimensional bonding uniformity feature vector from the preprocessing, standardization was first performed. The mean of the corresponding feature in the training sample set was subtracted from each feature, and then divided by its standard deviation to achieve feature uniformity. Taking the roughness feature as an example, the original value was 3.82, with a corresponding mean of 1.2 and a standard deviation of 0.8; after standardization, the value was 3.28. All standardized features formed a normalized feature vector. Based on a training matrix of 1000 samples, including 300 normal samples, 400 bubble-type defect samples, and 300 wrinkle-type defect samples, a support vector machine with a radial basis function kernel was used for training, with a penalty coefficient set to 10 and a kernel parameter set to 0.01. Parameters were adjusted through 5-fold cross-validation, and the final model achieved a classification accuracy of 94% on the validation set. The sample to be detected was input into the trained classifier, and the distance to the decision boundary was calculated to be -2.3; a negative value indicates that the sample is on the abnormal category side. In the two-stage classification, the first threshold was set to 0, and a distance less than 0 was considered abnormal; the second threshold was set to -1.5, and a distance less than -1.5 was considered a bubble-type defect. Therefore, this sample was classified as a bubble-type defect. The confidence score was 0.85, which is higher than the threshold requirement of 0.75, so this identification result was retained. Finally, the defective area was marked in red on the display imaging map, and several other detected defective areas were also marked with corresponding colors, forming a complete interface abnormality area mapping map. This visually shows the location and type of defects that occurred during the film lamination process, providing a precise basis for adjusting the production line process.

[0047] In one specific embodiment, the process of executing step S105 may specifically include the following steps: (1) Perform region segmentation on the defect marker region in the interface abnormal region mapping result to separate the bubble defect region and the wrinkle defect region. (2) The contour tracking algorithm is used to extract the bubble boundary line segments of the separated bubble defect area and calculate the bubble area and shape features; (3) Establish a depth mapping matrix based on the gray gradient information of the bubble region, estimate the depth of each bubble region, and generate a bubble height model; (4) Apply the skeleton extraction algorithm to the fold defect area to obtain the fold center line and width distribution, and calculate the direction and curvature of the fold; (5) Based on the morphological characteristics and location information of the folds, the stress distribution at the interface is calculated using a thin plate model, and the stress value matrix of the folded region is constructed. (6) Integrate the position and volume data of all bubbles into a three-dimensional spatial distribution model to form the interlayer bubble distribution; (7) Summarize the fold stress value matrix data, express the stress concentration area and propagation path in the form of heat map, and generate fold stress field map.

[0048] Specifically, the defect-marked regions in the interface anomaly mapping results are segmented to separate bubble defect regions and wrinkle defect regions. First, the classification result label map generated by the support vector classifier is read, and different types of defect regions are distinguished by label values. Bubble defect regions are typically marked in red or with a value of 1, while wrinkle defect regions are marked in yellow or with a value of 2. Connectivity analysis is used to group similar labels to ensure that each independent defect region has a unique identifier. This process establishes bubble defect sets and wrinkle defect sets, recording their respective locations, areas, and boundary information, laying the foundation for subsequent processing. For the separated bubble defect regions, a contour tracking algorithm is used to extract bubble boundary segments. When calculating the bubble area and shape features, a boundary tracking algorithm is used to process each bubble region individually. This algorithm starts from any point on the edge of the bubble region and searches for adjacent boundary points in a clockwise or counterclockwise direction until it returns to the starting point, forming a complete closed boundary line. After obtaining the boundary, calculate shape feature parameters such as bubble area (total number of pixels inside the boundary), perimeter (number of boundary points), roundness (measures how close the bubble shape is to a circle), and ratio of major axis to minor axis (reflects the flatness of the bubble). These parameters can effectively distinguish between normal micro bubbles and abnormally large bubbles.

[0049] A depth mapping matrix is ​​established based on the gray-level gradient information of the bubble region. Depth estimation is performed for each bubble region, and the generation of the bubble height model utilizes the optical property changes caused by bubbles at the film interface. First, the gray-level distribution of the bubble region is extracted from the original enhanced image. The bubble center typically exhibits brighter or darker characteristics, forming a gradient with the surrounding area. By analyzing this gradient distribution, a correspondence between gray-level values ​​and depth is established, converting the two-dimensional gray-level image into a three-dimensional depth map. For each bubble, the depth change is estimated from the edge to the center based on the gray-level change trend, forming a three-dimensional contour model of the bubble. This gray-level-based depth estimation method fully utilizes the advantages of non-contact optical detection, obtaining the three-dimensional structural information of the bubble without damaging the sample. A skeleton extraction algorithm is applied to the wrinkled defect region to obtain the wrinkle centerline and width distribution. When calculating the direction and curvature of the wrinkles, a morphological thinning operation is used to gradually "strip" away the edge pixels of the wrinkled region, preserving the centerline structure. The thinning process continuously deletes boundary points while maintaining regional connectivity, ultimately obtaining a skeleton line with a width of one pixel, i.e., the centerline of the wrinkle. After obtaining the centerline, the distance from each centerline point to the nearest boundary point is calculated to form a fold width distribution map. Further analysis of the centerline's directional changes reveals the main orientation (direction angle) and curvature changes of the folds. These parameters reflect the directionality and severity of fold formation, providing crucial input for stress analysis.

[0050] Based on the morphological characteristics and location information of the folds, a thin-plate model is used to calculate the stress distribution at the interface. When constructing the stress value matrix of the folded region, the thin film-conductive layer interface is regarded as a thin-plate structure undergoing stress deformation. Based on the acquired geometric features of the folds (centerline position, width distribution, curvature, etc.), the stress state at each point is estimated using thin-plate deformation theory. The strain in the folded region is closely related to the centerline curvature and fold height, and the stress distribution at each point of the film is calculated accordingly. The stress calculation considers physical parameters such as the material's elastic modulus and Poisson's ratio, generating a stress value matrix covering the entire folded region, which quantitatively represents the stress state at each point of the interface. All bubble position and volume data are integrated into a three-dimensional spatial distribution model. When forming the interlayer bubble distribution, a unified three-dimensional coordinate system is first established, with the horizontal and vertical axes corresponding to the film plane position and the vertical axis representing the depth direction. The center position, maximum depth, volume, and other parameters of each detected bubble are mapped to this coordinate system to form point cloud data. For densely distributed bubble groups, cluster analysis is performed to identify bubble aggregation regions. The final generated 3D model intuitively displays the spatial distribution characteristics of bubbles in the thin film interface, including key information such as bubble density, size distribution, and depth distribution.

[0051] By compiling the stress value matrix data of folds and representing stress concentration areas and propagation paths in the form of a heatmap, the calculated stress value matrix is ​​converted into a color-coded heatmap during the generation of the fold stress field map. A red-yellow-green-blue spectrum is typically used, with red representing high-stress areas and blue representing low-stress areas. The color gradient visually displays the stress distribution, highlighting stress concentration points and stress propagation paths. Overlaying the fold center line and isostress lines on the heatmap further enhances the visualization, enabling process engineers to quickly locate key problem areas.

[0052] For example, in the film bonding inspection of a flexible display screen, a support vector classifier identified three bubble defects and two wrinkle defects. First, these defect areas were separated to establish independent defect sets. For the largest bubble area, a contour tracking algorithm was applied, extracting 154 boundary points. The bubble area was calculated to be 1862 pixels, its perimeter 154 pixels, and its circularity 0.84, indicating a relatively regular bubble shape. Analysis of the bubble area's grayscale distribution revealed a grayscale value of 216 in the central region and an average grayscale value of 142 in the edge region. A grayscale-depth mapping relationship was established, estimating the maximum bubble depth to be 15 micrometers, and the overall structure to be hemispherical. For the most significant wrinkle area, a skeleton extraction algorithm was applied, yielding a centerline 326 pixels long. The average width measured along the centerline was 23 pixels, with a maximum width of 48 pixels, located in the middle of the wrinkle. Analysis of the centerline's direction showed a main directional angle of 37 degrees, roughly parallel to the display screen edge, suggesting the wrinkles may originate from edge stress transmission. Based on the geometric characteristics of the folds, the stress distribution was calculated. The maximum stress value appeared at the widest part of the fold, reaching 3.2 MPa, more than four times the normal interface stress. Integrating all bubble data into the 3D model revealed that the bubbles were mainly concentrated in the upper right corner of the display area, exhibiting a significant clustering effect, with a maximum bubble depth of 22 micrometers. The fold stress heatmap clearly showed a high-stress band extending inward from the edge, a finding that provided important reference for optimizing the bonding process. Based on these analysis results, the bonding pressure distribution and temperature parameters were adjusted, significantly reducing the bubble and fold defect rate in subsequent production.

[0053] In one specific embodiment, the process of executing step S106 may specifically include the following steps: (1) Extract bubble characteristic indicators from the interlayer bubble distribution, including bubble density value, bubble size distribution value, maximum bubble diameter value and total bubble volume value; (2) Extract fold characteristic indicators from the fold stress field diagram, including the total fold length, average fold width, peak fold stress, and fold direction consistency. (3) Input the extracted bubble feature index and wrinkle feature index into the interface integrity evaluation model to establish the mapping relationship between features and integrity score; (4) Assign weights to bubble factors and wrinkle factors in the interface integrity evaluation model, and calculate the weighted defect score according to the severity of each indicator; (5) Compare the weighted defect score with the standard defect level threshold to determine the defect severity level of the current sample; (6) Based on the severity level of defects and the weighted defect score, the interface integrity evaluation model is used to convert the scores into a standardized fit interface integrity score.

[0054] Specifically, in extracting bubble characteristic indicators from interlayer bubble distribution, including bubble density, bubble size distribution, maximum bubble diameter, and total bubble volume, the bubble data in the three-dimensional spatial distribution model is first read. The bubble density is calculated by counting the number of bubbles per unit area, expressed as bubbles per square centimeter, reflecting the concentration of bubble distribution. The bubble size distribution is calculated by grouping all detected bubbles by diameter and statistically analyzing the proportion of bubbles in each size range, typically categorized into four levels: micro-small (<30μm), small (30-100μm), medium (100-300μm), and large (>300μm), forming a size distribution histogram. The maximum bubble diameter is the value of the largest bubble among all bubbles, directly reflecting the scale of the most severe bubble defect. The total bubble volume is an estimated volume by summing all bubbles, reflecting the overall scale of the bubble defect; the volume calculation is based on the previously established depth mapping matrix. These four indicators together constitute a comprehensive quantitative description of bubble defects, providing a data foundation for subsequent defect assessment. Extracting fold characteristic indicators from the fold stress field map, including the total fold length, average fold width, peak fold stress, and fold direction consistency, begins with analyzing the fold stress field heatmap and fold centerline data. The total fold length is the sum of all detected fold centerline lengths, reflecting the overall scale of the fold defect. The average fold width is obtained by calculating the average width of each fold measured along its centerline, characterizing the average severity of the folds. The peak fold stress is the maximum stress value detected in the stress field map, representing the intensity of the most severe stress concentration location. The fold direction consistency is calculated by analyzing the direction angles of each fold centerline; specifically, it's determined by statistically analyzing the proportion of fold length in the main direction (the direction angle interval with the highest frequency) to the total length, reflecting the directional characteristics of fold formation. High direction consistency usually indicates a systematic fitting problem. These four indicators quantify the characteristics of fold defects from different perspectives, providing crucial evidence for fitting quality assessment.

[0055] The extracted bubble and wrinkle feature indicators are input into the interface integrity evaluation model. A multi-feature fusion evaluation strategy is employed to establish the mapping relationship between features and integrity scores. The interface integrity evaluation model is a hybrid model based on expert knowledge and machine learning, standardizing each feature indicator and mapping it to the defect score space. Standardization allows for comparison and combination of feature indicators with different dimensions and numerical ranges. For bubble features, bubble density, size distribution, maximum diameter, and total volume are typically compared with predetermined standard thresholds to calculate a normalized defect score. Similarly, for wrinkle features, the total wrinkle length, average width, peak stress, and directional consistency are compared with standard thresholds to obtain a normalized score. This standardization establishes a quantitative relationship between the original feature indicators and the severity of defects, laying the foundation for subsequent weighted calculations. In the interface integrity evaluation model, weights are assigned to bubble and wrinkle factors. During the calculation of weighted defect scores based on the severity of each indicator, weight coefficients are set according to the impact of various defects on product performance in different application scenarios. The allocation of weights for bubble and wrinkle factors follows actual product requirements. For example, for displays requiring high optical transparency, bubble defects have a greater impact, so the bubble factor weight will be higher; for flexible displays requiring high mechanical bending performance, wrinkle defects have a greater impact, so the wrinkle factor weight will be higher. In practice, the bubble defect score is typically multiplied by the bubble factor weight, the wrinkle defect score is multiplied by the wrinkle factor weight, and then summed to obtain the weighted defect score. This weighting mechanism ensures that the evaluation results accurately reflect the actual impact of different types of defects on product quality.

[0056] In determining the severity level of a sample's defects by comparing weighted defect scores with standard defect level thresholds, a multi-level defect severity classification standard is established. Typically, defect severity is divided into four levels: minor (Level A), moderate (Level B), severe (Level C), and extremely severe (Level D), each corresponding to a score range. For example, a weighted defect score of 0-10 corresponds to Level A, 10-25 to Level B, 25-40 to Level C, and greater than 40 to Level D. By comparing the calculated weighted defect score with these thresholds, the sample's defect severity level is determined. This grading mechanism makes quality assessment results more intuitive and facilitates process control and quality management.

[0057] Based on the severity level of defects and weighted defect scores, a composite scoring function is used to calculate the standardized fit interface integrity score during the conversion process from the interface integrity evaluation model. The standardized fit interface integrity score is typically based on a 100-point scale, with 100 points representing perfect fit and 0 points representing complete failure. The scoring formula is as follows:

[0058] in, (Interface Integrity Score) represents the score for the integrity of the interface. This represents the full score, usually 100. (Chi Global) is the global adjustment coefficient that controls the rate of score decrease; (DefectScore) is the weighted defect score; This is a penalty factor based on the defect level, which increases as the defect level increases. The numerical representation of the severity level of the defect is as follows: A is 1, B is 2, C is 3, and D is 4. For reference, it is usually set to 2; This is a grade sensitivity parameter that controls the score difference between different grades; This represents the fault tolerance coefficient for the current product type. This is the maximum fault tolerance factor; This is a tolerance index that controls the non-linearity of the tolerance effect. Using this formula, the final interface integrity score is generated by comprehensively considering defect scores, severity levels, and product-specific factors.

[0059] The above describes the thin-film bonding defect detection method based on image processing in the embodiments of this application. The following describes the thin-film bonding defect detection system based on image processing in the embodiments of this application. Please refer to [link to relevant documentation]. Figure 2 One embodiment of the thin-film bonding defect detection system based on image processing in this application includes: The acquisition module is used to perform differential optical acquisition of the thin film-conductive layer interface through a polarization-enhanced imaging device, and to preprocess the original image through interface reflection enhancement filtering to obtain an interface-enhanced preprocessed image. The analysis module is used to extract the edges of the bonding area based on the interface enhancement preprocessed image using a two-layer gradient edge detection algorithm, and obtain a micro-defect sensitive area marker map; The extraction module is used to extract features from the segmented region based on the micro-defect sensitive area marker map using a multi-scale morphology feature extractor, and obtain the interface smoothness index and the fitting uniformity feature vector. The recognition module is used to input the uniformity feature vector of the fit into the support vector classifier to identify bubbles and wrinkles, and obtain the mapping result of abnormal areas of the interface. The reconstruction module is used to perform three-dimensional reconstruction of microbubbles and microfolds based on the mapping results of abnormal interface areas and the contour extraction algorithm, so as to obtain the distribution of interlayer bubbles and the stress field map of folds. The evaluation module is used to evaluate the bonding quality of the thin film-conductive layer based on the interlayer bubble distribution and wrinkle stress field diagram using an interface integrity evaluation model, and obtain a bonding interface integrity score.

[0060] Through the synergistic cooperation of the aforementioned components, differential optical acquisition using a polarization-enhanced imaging device combined with interface reflection enhancement filtering preprocessing effectively improves the optical contrast and visibility of micro-defects at the interface, solving the problem of traditional imaging's difficulty in capturing minute interface anomalies. It achieves a detection rate as high as 95% for defects at the 30μm level. The dual-layer gradient edge detection algorithm organically integrates the Canny and Sobel operators and uses adaptive threshold segmentation to effectively extract edge information, significantly improving the detection sensitivity for irregular edges. This allows for precise labeling of minute defect areas, with edge localization accuracy improved to the sub-pixel level. The multi-scale morphological feature extractor performs feature analysis on the segmented region, integrating geometric morphological parameters, gray-level statistical features, and wavelet decomposition features to construct a comprehensive defect feature description system. The interface smoothness index and fitting uniformity feature vector can quantitatively characterize the severity of defects, with feature extraction reaching 32 dimensions, providing a rich feature space for accurate classification. The support vector classifier, based on artificial intelligence technology, achieves accurate identification of bubble-wrinkle types. The classifier, trained on a large number of labeled samples, employs a nonlinear kernel function to map a multidimensional feature space and automatically optimizes the decision boundary, achieving a classification accuracy of over 94%. This demonstrates the crucial role of machine learning algorithms in defect identification. Furthermore, the maximum margin characteristic unique to the support vector machine model makes the classification results more robust, effectively addressing noise interference and feature fluctuations in the production environment. The contour extraction algorithm, combined with the depth mapping matrix, enables the 3D reconstruction of microbubbles and microwrinkles, transforming 2D image information into 3D spatial structures and providing key parameters such as defect volume and depth, offering significant guidance for stress analysis and quality prediction. The interface integrity evaluation model comprehensively considers various defect features and indicators, establishing a standardized scoring mechanism that transforms complex defect information into an intuitive integrity score. This model integrates expert knowledge and data-driven intelligent evaluation methods, not only objectively evaluating the bonding quality of the current sample but also tracking process change trends, providing precise basis for parameter adjustments during production. This achieves closed-loop control of defect detection and process optimization, ultimately significantly improving the yield and consistency of thin-film bonded products.

[0061] Reference Figure 3 This invention also provides a computer device, which can be a server, and its internal structure can be as follows: Figure 3As shown, the computer device includes a processor, memory, display screen, input device, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database stores the data corresponding to this embodiment. The network interface is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements the above-described method.

[0062] Those skilled in the art will understand that Figure 3 The structures shown are merely block diagrams of some structures related to the present invention and do not constitute a limitation on the computer devices on which the present invention is applied.

[0063] An embodiment of the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method. It is understood that the computer-readable storage medium in this embodiment can be a volatile readable storage medium or a non-volatile readable storage medium.

[0064] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the present invention and embodiments can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual-rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM, etc.

[0065] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0066] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0067] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for detecting film bonding defects based on image processing, characterized in that, The image processing-based method for detecting film bonding defects includes: Differential optical acquisition of the thin film-conductive layer interface is performed using a polarization-enhanced imaging device, and the original image is preprocessed by interface reflection enhancement filtering to obtain an interface-enhanced preprocessed image. Based on the interface enhancement preprocessed image, the edge of the bonding area is extracted by a two-layer gradient edge detection algorithm to obtain a micro-defect sensitive area marker map; Based on the micro-defect sensitive area marker map, the segmented region is used to extract features through a multi-scale morphology feature extractor to obtain the interface smoothness index and the fitting uniformity feature vector. The uniformity feature vector of the fit is input into the support vector classifier to identify bubbles and wrinkles, and the mapping result of the abnormal area of ​​the interface is obtained. Based on the mapping results of the interface anomaly region, the microbubbles and microfolds are reconstructed in three dimensions using a contour extraction algorithm to obtain the interlayer bubble distribution and fold stress field map. The bonding quality of the thin film-conductive layer is evaluated using an interface integrity evaluation model based on the interlayer bubble distribution and wrinkle stress field diagram, and the bonding interface integrity score is obtained.

2. The thin-film bonding defect detection method based on image processing according to claim 1, characterized in that, The process involves differential optical acquisition of the thin-film-conductive layer interface using a polarization-enhanced imaging device, followed by preprocessing the original image using interface reflection enhancement filtering to obtain an interface-enhanced preprocessed image, including: By illuminating the thin film-conductive layer interface with a polarized light source set at a polarization angle of 45°, the change in the intensity of the reflected light is captured by a line scan camera to obtain the original image of the interface reflection; Based on the original image of the interface reflection, random noise is eliminated by Gaussian filtering algorithm to generate a noise-reduced image; Adaptive histogram equalization is applied to the denoised image to extract interface reflection contrast features and form a contrast-enhanced image. The contrast-enhanced image is fed into the interface reflection enhancement filter to enhance the brightness change characteristics at the interface between the thin film and the conductive layer, thereby generating an interface layer prominence map. Geometric correction and compensation are performed on the interface hierarchy highlighting diagram to eliminate lens distortion and shooting angle deviation, forming a geometrically accurate interface diagram. The geometrically accurate interface map is dynamically expanded using a nonlinear contrast adjustment function to obtain the interface enhancement preprocessed image.

3. The thin-film bonding defect detection method based on image processing according to claim 1, characterized in that, The step of extracting edges from the bonding area using a dual-layer gradient edge detection algorithm based on the interface enhancement preprocessed image to obtain a micro-defect sensitive area marker map includes: The Canny operator and the Sobel operator were applied to the interface enhancement preprocessed image to perform primary edge extraction, resulting in two sets of edge images; The two sets of edge images are fused using a logical AND operation to generate preliminary edge detection results; An adaptive threshold segmentation algorithm is used to apply the preliminary edge detection results, and a dynamic threshold is set to control the edge detection sensitivity to obtain edge candidate regions; A two-layer gradient calculation is applied to the edge candidate region to extract gradient direction and gradient magnitude features, forming a gradient feature map; The gradient feature map is subjected to region growing processing to connect discontinuous edges and construct a complete edge contour. The complete edge contour is divided into watershed segments and marked to define the film bonding area, conductive layer area, and interface area, thus obtaining the marking map of the micro-defect sensitive area.

4. The thin-film bonding defect detection method based on image processing according to claim 1, characterized in that, The process involves extracting features from the segmented regions using a multi-scale topography feature extractor based on the micro-defect sensitive area marker map, resulting in an interface smoothness index and a bonding uniformity feature vector, including: The micro-defect sensitive area marking map is numbered and indexed, a unique identifier is assigned to each area, and an area retrieval table is established; The boundary point set of each segmented region is read sequentially from the region retrieval table, and the geometric parameters of the region are calculated, including area, perimeter, equivalent radius and shape factor. For each segmented region, its gray-level histogram is extracted, and first-order and second-order statistics are calculated to form a gray-level distribution descriptor. A curvature calculation algorithm is applied to the thin film-conductive layer interface region. The curvature value of each point on the interface line is calculated by fitting a quadratic polynomial, and the curvature value sequence on the interface line is integrated to obtain the interface smoothness index. Wavelet decomposition was applied to each segmented region to extract low-frequency and high-frequency components at different scales in sequence, and a multi-scale spectral feature library was constructed. The interface region between the thin film and the conductive layer is divided into a fixed number of equal-length segments. The local smoothness index of each segment is calculated, and the overall bonding uniformity index is synthesized by weighted averaging. All extracted morphological, statistical, and multi-scale features are integrated, and feature dimensionality reduction and normalization are performed to generate a fixed-dimensional fitting uniformity feature vector.

5. The thin-film bonding defect detection method based on image processing according to claim 1, characterized in that, The step of inputting the uniformity feature vector into a support vector classifier for bubble-wrinkle recognition to obtain the interface anomaly region mapping result includes: The uniformity feature vector of the fit is standardized by unifying its dimensions and normalizing its values ​​to generate a normalized feature vector. A sample training matrix is ​​established based on normalized feature vectors, and the parameters of the support vector classifier are optimized by gradient descent to construct the decision boundary of the feature space. Input the fitting uniformity feature vector of the sample to be detected into the trained support vector classifier and calculate the distance scalar between the sample and the classification boundary; The sample region is classified in two stages based on the distance scalar size. First, the region is divided into normal region and abnormal region. Then, the abnormal region is further subdivided into bubble-type defects and wrinkle-type defects. Calculate a confidence score for each identified defective region and set a threshold to filter out low-confidence regions, retaining high-reliability identification results; The high-reliability identification results are mapped back to the original image according to their original coordinate positions, generating region markers with different color codes to form the interface abnormal region mapping results.

6. The thin-film bonding defect detection method based on image processing according to claim 1, characterized in that, The mapping results based on the interface anomaly region are used to perform three-dimensional reconstruction of microbubbles and microfolds through a contour extraction algorithm, resulting in interlayer bubble distribution and wrinkle stress field maps, including: The defect marker region in the interface abnormal region mapping result is segmented to separate the bubble defect region and the wrinkle defect region. The contour tracking algorithm is used to extract the bubble boundary line segments of the separated bubble defect area and calculate the bubble area and shape features. A depth mapping matrix is ​​established based on the gray-level gradient information of the bubble region, and the depth of each bubble region is estimated to generate a bubble height model. A skeleton extraction algorithm is applied to the fold defect area to obtain the fold centerline and width distribution, and to calculate the fold direction and curvature. Based on the morphological characteristics and location information of the folds, a thin plate model is used to calculate the stress distribution at the interface and construct the stress value matrix of the folded region. The location and volume data of all bubbles are integrated into a three-dimensional spatial distribution model to form the interlayer bubble distribution; The stress value matrix data of the folds is summarized, and the stress concentration areas and propagation paths are expressed in the form of a heat map to generate the stress field map of the folds.

7. The thin-film bonding defect detection method based on image processing according to claim 1, characterized in that, The bonding quality of the thin film-conductive layer is evaluated using an interface integrity evaluation model based on the interlayer bubble distribution and wrinkle stress field diagram, resulting in a bonding interface integrity score, including: Bubble characteristic indicators are extracted from the interlayer bubble distribution, including bubble density value, bubble size distribution value, maximum bubble diameter value, and total bubble volume value; Extract fold characteristic indicators from the fold stress field diagram, including the total fold length, average fold width, peak fold stress, and fold direction consistency. The extracted bubble feature index and wrinkle feature index are input into the interface integrity evaluation model to establish a mapping relationship between features and integrity scores. In the interface integrity evaluation model, bubble factor weights and wrinkle factor weights are assigned, and a weighted defect score is calculated based on the severity of each indicator. The weighted defect score is compared with the standard defect level threshold to determine the severity level of the defect in the current sample. Based on the severity level of the defect and the weighted defect score, the interface integrity evaluation model is used to convert the score into a standardized fit interface integrity score.

8. A thin-film bonding defect detection system based on image processing, used to implement the thin-film bonding defect detection method based on image processing as described in any one of claims 1-7, characterized in that, The image processing-based thin-film bonding defect detection system includes: The acquisition module is used to perform differential optical acquisition of the thin film-conductive layer interface through a polarization-enhanced imaging device, and to preprocess the original image through interface reflection enhancement filtering to obtain an interface-enhanced preprocessed image. The analysis module is used to extract the edges of the bonding area based on the interface enhancement preprocessed image using a two-layer gradient edge detection algorithm, and obtain a micro-defect sensitive area marker map; The extraction module is used to extract features from the segmented region based on the micro-defect sensitive area marker map using a multi-scale morphology feature extractor, and obtain the interface smoothness index and the fitting uniformity feature vector. The recognition module is used to input the uniformity feature vector of the fit into the support vector classifier to identify bubbles and wrinkles, and obtain the mapping result of abnormal areas of the interface. The reconstruction module is used to perform three-dimensional reconstruction of microbubbles and microfolds based on the mapping results of abnormal interface areas and the contour extraction algorithm, so as to obtain the distribution of interlayer bubbles and the stress field map of folds. The evaluation module is used to evaluate the bonding quality of the thin film-conductive layer based on the interlayer bubble distribution and wrinkle stress field diagram using an interface integrity evaluation model, and obtain a bonding interface integrity score.

9. A computer device, characterized in that, The method includes a memory and a processor, the memory storing a computer program that can run on the processor, characterized in that the processor executes the computer program to implement the image processing-based thin film bonding defect detection method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, the computer program causing a processor, when executed by a processor, to perform the image processing-based thin-film bonding defect detection method as described in any one of claims 1 to 7.

Images