On-line detection method for surface defects of EVA adhesive film based on machine vision

Abstract

The invention discloses an on-line detection method for surface defects of an EVA (Ethylene Vinyl Acetate) adhesive film based on machine vision, which belongs to the technical field of industrial automatic quality detection and comprises the following steps: acquiring multiband image data of the surface of the EVA adhesive film acquired by a preset multi-angle self-adaptive imaging system; performing background difference processing on the multiband image data to generate a difference image; performing adaptive contrast enhancement processing on the difference image to obtain an enhanced image; extracting texture feature data and morphological feature data from the enhanced image to generate a multi-feature data set; performing dynamic weighted fusion processing on the multi-feature data set to generate a defect probability graph; the defect probability graph is input into the layered detection network for processing, a defect detection result is generated, the technical scheme that multi-angle self-adaptive imaging, multi-feature dynamic fusion and the layered detection network are combined is adopted, the defect imaging quality and feature significance can be improved in a self-adaptive mode, and high-speed accurate detection and dynamic optimization of the surface defects of the EVA adhesive film are achieved.

Description

Technical Field

[0001] This invention relates to the field of industrial automation quality inspection technology, and in particular to an online detection method for surface defects of EVA film based on machine vision. Background Technology

[0002] EVA film, or ethylene-vinyl acetate copolymer film, is widely used in photovoltaic module encapsulation and other fields due to its excellent adhesion, durability, and optical properties. It serves as a key material for protecting solar cells and ensuring their long-term stable operation. The surface quality of the EVA film directly affects the power generation efficiency and lifespan of photovoltaic modules; therefore, high-speed and accurate online detection of surface defects during its production process is crucial. Machine vision inspection technology, as a mainstream automation method to replace manual visual inspection, has been widely applied in this field.

[0003] Existing methods for detecting surface defects in EVA films typically employ a fixed machine vision system. This system generally includes a fixed industrial light source, an industrial camera, and a pre-set image processing algorithm. The workflow usually involves illuminating the moving film surface with the light source, capturing images with the camera, and then using conventional image processing algorithms such as filtering, edge detection, and thresholding to identify and segment suspected defect areas. Finally, the defects are classified based on simple geometric or grayscale features such as area and grayscale.

[0004] However, existing technical solutions have significant technical shortcomings in practical applications. First, the EVA film surface has a certain degree of transparency and reflectivity. Using fixed illumination methods makes it difficult to achieve clear imaging of various defects with different optical characteristics, such as bubbles, scratches, and foreign objects. This often results in some defects being missed due to low contrast, or false alarms caused by reflection. Second, traditional image processing algorithms are mostly global operations, which are poorly adaptable to uneven background texture and brightness variations caused by fluctuations in the production process. Finally, relying on single or a few simple features for defect discrimination makes it difficult to effectively distinguish between defects with complex shapes or similar features and interference. Furthermore, once the system parameters are set, they are difficult to adaptively adjust with changes in production conditions, leading to unstable detection performance. Summary of the Invention

[0005] To address the aforementioned issues, this invention provides an online detection method for EVA film surface defects based on machine vision. It employs a technical solution combining multi-angle adaptive imaging, multi-feature dynamic fusion, and a hierarchical detection network, which can adaptively improve the defect imaging quality and feature saliency, achieving high-speed, accurate detection and dynamic optimization of EVA film surface defects.

[0006] The above objectives can be achieved through the following approach: A machine vision-based online defect detection method for EVA film surfaces includes: acquiring multi-band image data of the EVA film surface through a preset multi-angle adaptive imaging system; performing background subtraction processing on the multi-band image data to generate a difference image; performing adaptive contrast enhancement processing on the difference image to obtain an enhanced image; extracting texture feature data and morphological feature data from the enhanced image to generate a multi-feature dataset; performing dynamic weighted fusion processing on the multi-feature dataset to generate a defect probability map; and inputting the defect probability map into a hierarchical detection network for processing to generate a defect detection result.

[0007] Optionally, acquiring multi-band image data of the EVA film surface collected by a preset multi-angle adaptive imaging system includes: adjusting the intensity and angle of a preset annular multi-band LED light source to generate a multi-band light field covering the EVA film surface; filtering the reflected light generated after the multi-band light field interacts with the EVA film surface using a polarizing filter to generate original image data after suppressing reflection; and acquiring the original image data after suppressing reflection as the multi-band image data.

[0008] Optionally, generating the difference image includes: acquiring a defect-free template image of a standard EVA film based on preset historical data; calculating the pixel-level difference between the multi-band image data and the defect-free template image to generate a set of difference values; and generating the difference image based on the set of difference values, wherein high-value regions in the difference image are identified as potential defect regions.

[0009] Optionally, generating the enhanced image includes: analyzing the local grayscale distribution of the difference image and generating local grayscale statistics; calculating a contrast parameter based on the local grayscale statistics; and applying the contrast parameter to enhance the edge and texture details of the difference image to generate the enhanced image.

[0010] Optionally, generating the multi-feature dataset includes: applying a local binary mode algorithm to the enhanced image to generate texture feature data; performing multi-scale morphological operations on the enhanced image to generate morphological feature data; and fusing the texture feature data and morphological feature data to generate the multi-feature dataset.

[0011] Optionally, generating the defect probability map includes: obtaining weight parameters set for different defect types; applying the weight parameters to perform weighted calculations on the texture feature data and morphological feature data in the multi-feature dataset to generate a defect score; and generating the defect probability map based on the defect score, wherein the pixel values ​​of the defect probability map represent the confidence level of the existence of the defect.

[0012] Optionally, the hierarchical detection network includes: in the first layer, performing high-speed screening on the defect probability map to quickly locate and generate suspected defect areas; in the second layer, performing refined classification analysis on the multi-feature dataset corresponding to the suspected defect areas to generate refined inspection results; and in the third layer, adaptively updating the classifier model in the hierarchical detection network based on the refined inspection results and preset manual feedback samples to output the defect detection results.

[0013] Optionally, the adaptive update of the classifier model in the hierarchical detection network based on the fine detection results and preset manual feedback samples includes: collecting an incremental training dataset consisting of false positives and false negatives; optimizing the weight parameters in the dynamic weighted fusion processing of the multi-feature dataset using the incremental training dataset; and fine-tuning the classifier model using the incremental training dataset.

[0014] Optionally, the method further includes: generating a light source control command based on the defect detection result; and sending the light source control command to the multi-angle adaptive imaging system to adjust the light source parameters of the annular multi-band LED light source.

[0015] Based on the same inventive concept, this invention also provides an online detection system for surface defects of EVA film based on machine vision, comprising: The multi-band acquisition module is used to acquire multi-band image data of the EVA film surface through a preset multi-angle adaptive imaging system; The background difference module is used to perform background difference processing on the multi-band image data to generate a difference image; The enhancement processing module is used to perform adaptive contrast enhancement processing on the difference image to obtain an enhanced image; A multi-feature extraction module is used to extract texture feature data and morphological feature data from the enhanced image to generate a multi-feature dataset. The dynamic fusion module is used to perform dynamic weighted fusion processing on the multi-feature dataset to generate a defect probability map. The layered detection module is used to input the defect probability map into the layered detection network for processing and generate defect detection results.

[0016] Compared with the prior art, the present invention has the following advantages: 1. By constructing a closed-loop adaptive imaging system, the parameters of the multi-angle, multi-band light source at the front end can be dynamically adjusted based on the back-end detection results. This active illumination strategy ensures that the original image with the highest information entropy can always be obtained for different types of defects, maximizing the visual contrast between defects and the background from the source, and significantly improving the system's ability to capture complex and variable defects and its adaptability to environmental changes.

[0017] 2. A refined processing workflow from preprocessing to feature engineering is proposed. Background subtraction effectively suppresses interference from the inherent texture of the substrate, and combined with adaptive contrast enhancement technology, it can accurately amplify weak defect signals without introducing excessive noise. Subsequently, complementary features of texture and morphology are extracted in parallel and dynamically weighted and fused to construct a comprehensive and profound description of the defects, significantly improving the accuracy of subsequent identification.

[0018] 3. A highly efficient hierarchical detection network structure was designed, achieving an effective balance between detection speed and accuracy. By rapidly filtering the defect probability map in the first layer to quickly identify potential targets and concentrating computational resources on a few suspicious areas, and then retrieving the complete original multi-feature dataset for fine-grained classification of these areas in the second layer, this coarse-to-fine strategy ensures that the system meets the high real-time requirements of industrial online production while guaranteeing a low false alarm rate and a low false negative rate in the final detection results.

[0019] 4. An adaptive update mechanism based on human feedback is integrated, enabling the entire detection system to continuously learn and self-optimize. By collecting and utilizing false positive and false negative samples, the system can dynamically fine-tune the classifier model and optimize the weight parameters of the front-end feature fusion, allowing it to proactively adapt to fluctuations in production conditions such as changes in raw material batches. This ensures long-term stability and reliability of detection performance and reduces manual maintenance costs.

[0020] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art 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.

[0022] Figure 1 This is a schematic diagram of the online detection method for EVA film surface defects based on machine vision, according to an embodiment of the present invention.

[0023] Figure 2 This is a schematic diagram comparing the multi-band imaging and background difference effects of EVA film according to an embodiment of the present invention.

[0024] Figure 3 This is a schematic diagram of the effect curve of multi-feature dynamic fusion and defect probability map generation in an embodiment of the present invention.

[0025] Figure 4 This is a schematic diagram of the structure of the online detection system for EVA film surface defects based on machine vision, according to an embodiment of the present invention. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] Reference Figure 1 One embodiment of the present invention proposes an online detection method for EVA film surface defects based on machine vision. It adopts a technical solution that combines multi-angle adaptive imaging, multi-feature dynamic fusion and hierarchical detection network, which can adaptively improve the defect imaging quality and feature saliency, and realize high-speed and accurate detection and dynamic optimization of EVA film surface defects.

[0028] The method described in this embodiment specifically includes: Acquire multi-band image data of the EVA film surface through a preset multi-angle adaptive imaging system; The multi-band image data is subjected to background subtraction processing to generate a difference image; An enhanced image is obtained by performing adaptive contrast enhancement processing on the difference image; Texture feature data and morphological feature data are extracted from the enhanced image to generate a multi-feature dataset; The multi-feature dataset is dynamically weighted and fused to generate a defect probability map; The defect probability map is input into a hierarchical detection network for processing to generate defect detection results.

[0029] Specifically, a complete technology chain is constructed, from high-quality image acquisition to intelligent hierarchical decision-making. A proactively adjustable multi-angle, multi-band adaptive imaging system optimizes lighting conditions for the surface characteristics of EVA film, acquiring raw image data that highlights defect features. Background subtraction technology compares the dynamically acquired images with a standard defect-free template, effectively separating potential abnormal regions from complex background textures to form difference images. To further enhance these weak abnormal signals, an adaptive contrast enhancement algorithm is used to sharpen local details in the difference images, resulting in enhanced images with clearer features. Two complementary feature dimensions—texture and morphology—are extracted in parallel from the enhanced images and fused into a multi-feature dataset, achieving a comprehensive quantitative description of defects. Multi-dimensional features are intelligently fused using dynamic weighting to generate an intuitive defect probability map, where each pixel value represents the confidence level of a defect at that point. The defect probability map is input into a hierarchical detection network, which analyzes the data using a coarse-to-fine strategy, first quickly filtering high-probability regions, then refining these regions to obtain the final detection result. This significantly improves the accuracy, robustness, and intelligence of online defect detection on EVA film surfaces. Through front-end adaptive imaging technology, a high signal-to-noise ratio for image data is guaranteed from the source, effectively overcoming the problems of missed and false detections caused by single lighting conditions in traditional detection methods. The combined preprocessing workflow of background subtraction and adaptive enhancement greatly enhances the distinction between defects and background, making even tiny and low-contrast defects readily apparent. Multi-feature fusion and defect probability map generation transform the complex pattern recognition problem into a more cohesive and discriminative decision-making basis, improving the ability to identify various complex defects. Finally, the application of a hierarchical detection network cleverly balances detection speed and accuracy, meeting the high-speed requirements of industrial production lines while enabling in-depth and detailed analysis of suspicious targets. This fundamentally improves the overall performance of the detection system, enabling stable and reliable automated quality control in variable production environments.

[0030] Optionally, acquiring multi-band image data of the EVA film surface obtained through a preset multi-angle adaptive imaging system includes: Adjust the intensity and angle of the preset ring-shaped multi-band LED light source to generate a multi-band light field covering the surface of the EVA film; The reflected light generated after the interaction of multi-band light fields with the surface of EVA film is filtered by a polarizing filter to generate the original image data after suppressing reflection. The original image data after suppressing reflection is acquired and used as the multi-band image data.

[0031] Specifically, a preset ring-shaped multi-band LED light source is activated and adjusted. This light source consists of LED units with multiple bands, including red, green, and blue bands of visible light, as well as possible non-visible light bands, arranged in a ring array. Adjusting its intensity and angle changes the illumination angle of the light source relative to the EVA film surface, generating a uniformly covered and controllable multi-band light field on the EVA film surface. This light field interacts with the moving EVA film surface, producing reflected light containing surface and subsurface information. Due to the characteristics of the EVA film material, a polarizing filter is placed in the optical path of the imaging sensor to address the specular reflections in the reflected light that often interfere with defect features, i.e., high-gloss reflections. When reflected light carrying complex information passes through this polarizing filter, high-gloss reflections in specific polarization directions are effectively filtered out, while diffuse reflections carrying defect information can pass through relatively well. After filtering, the original image data after suppressing reflections is finally formed. High-speed industrial cameras and other acquisition devices acquire this original image data after suppressing reflections, generating multi-band image data.

[0032] Optionally, generating the difference image includes: Obtain a defect-free template image of a standard EVA film based on preset historical data; Calculate the pixel-level differences between the multi-band image data and the defect-free template image to generate a set of difference values; The difference image is generated based on the set of difference values, wherein high-value regions in the difference image are identified as potential defect regions.

[0033] Specifically, under the same multi-angle self-contained imaging system conditions as online inspection, a large number of images of standard EVA film samples confirmed to be free of defects are acquired. These images constitute the preset historical data. By performing pixel-level averaging on these images, the effects of random noise and minor material fluctuations can be eliminated, ultimately obtaining a defect-free template image that represents the ideal surface state of the EVA film. In the online inspection process, when a new frame of multi-band image data is acquired, pixel-level subtraction is performed. For ease of calculation, the multi-band image data and the defect-free template image can first be converted into a single-channel grayscale image. This calculation process can be expressed by the following formula: Where (i,j) represents the pixel coordinates in the image. I(i,j) is the pixel grayscale value of the currently acquired multi-band image data at the corresponding coordinate point. T(i,j) is the pixel grayscale value of the pre-generated defect-free template image at the same coordinate point. D(i,j) is the difference value obtained by calculating the absolute value of the difference between the two grayscale values. After performing this operation on all pixels in the image, the set of all D(i,j) values ​​constitutes the difference value set. A new image is generated based on the difference value set, namely the difference image. In this difference image, the defect-free areas in the original image that are similar to the template image have a difference value D(i,j) close to zero, and therefore appear as dark areas in the image; while the defective areas, due to their different optical properties from the standard template, have significantly larger difference values, thus forming bright high-value areas in the image. These high-value areas are clearly identified as potential defective areas. For the specific effects of multi-band imaging and background subtraction, see [link to documentation]. Figure 2 After differentiation, potential defects such as bubbles and scratches are clearly presented in high-value areas, effectively removing background texture interference.

[0034] Optionally, generating the enhanced image includes: Analyze the local grayscale distribution of the difference image to generate local grayscale statistics; Calculate the contrast parameter based on local grayscale statistics; The contrast parameter is applied to enhance the edge and texture details of the difference image to generate the enhanced image.

[0035] Specifically, a neighborhood window of a preset size is defined centered on each pixel in the difference image. The grayscale distribution information of all pixels within this window is calculated and extracted, thereby generating local grayscale statistics for the central pixel. The two most crucial statistics are the local grayscale mean and the local grayscale standard deviation. For the local grayscale value and the local grayscale standard deviation: in, This is a local grayscale value. Let p(i,j) be the local grayscale standard deviation, p(i,j) be the grayscale value of a pixel within the window, and N² be the total number of pixels within the window. The local grayscale mean reflects the average brightness of the region, while the local grayscale standard deviation quantifies the degree of grayscale variation in that region, i.e., local contrast. Based on the local grayscale statistics, a unique contrast parameter is dynamically generated for each pixel. This parameter is typically represented as a gain coefficient, and its calculation method aims to inversely adjust the local contrast, i.e., apply a stronger gain to low-contrast areas and a weaker gain to high-contrast areas. The calculation method is as follows: Here, G(i,j) is the contrast parameter calculated for the pixel at coordinate (i,j). M is the global average gray value of the entire difference image, serving as a stable reference. σ(i,j) is the local gray standard deviation calculated within the neighborhood of this pixel. ε is a very small positive number, its function being to prevent mathematical errors such as a zero denominator when the local area is completely uniform, i.e., the standard deviation is zero. The resulting contrast parameter G(i,j) is adaptive; in areas with inconspicuous defects and flat backgrounds, σ(i,j) is smaller, and the value of G(i,j) is larger; conversely, in edge or noisy areas, σ(i,j) is larger, and the value of G(i,j) is smaller. The contrast parameter, tailored to each pixel, is applied to the original difference image to generate the final enhanced image. The enhanced pixel value E(i,j) can be determined jointly by the original pixel value D(i,j) and the contrast parameter G(i,j), thereby achieving fine enhancement of image edges and texture details.

[0036] Optionally, generating the multi-feature dataset includes: The enhanced image is processed using the Local Binary Pattern Algorithm to generate texture feature data; Perform multi-scale morphological operations on the enhanced image to generate morphological feature data; The multi-feature dataset is generated by fusing texture feature data and morphological feature data.

[0037] Specifically, to obtain texture feature data characterizing the microstructure and local spatial relationships of defect regions, the input enhanced image is processed using a Local Binary Pattern (LBMM) algorithm. The LBMM algorithm selects a circular or square neighborhood centered on each pixel in the image and compares the gray value of the center pixel with the gray values ​​of other pixels within the neighborhood. If the neighborhood pixel value is greater than or equal to the center pixel value, the corresponding position is recorded as 1; otherwise, it is recorded as 0, forming a binary sequence. The decimal value represented by the binary sequence is the LBMM value of the center pixel. This operation is performed on all pixels of the entire enhanced image to generate a LBMM map. A statistical histogram of the LBMM map is calculated. The statistical histogram quantifies the frequency of different micro-texture patterns, and the statistical histogram vector constitutes the texture feature data. To capture information such as the macroscopic contour, size, and shape of defects, multi-scale morphological operations are performed on the same enhanced image to generate morphological feature data. Multi-scale morphological operations employ a series of structuring elements of increasing size, such as 3x3, 5x5, and 7x7 pixel square or circular operators. By applying these structuring elements of different scales to basic morphological operations such as erosion, dilation, opening, and closing, defect regions of different sizes are effectively detected and segmented. For example, multi-scale top-hat transform can extract bright spot-like defects of different sizes, while multi-scale morphological gradients can delineate defect edges of varying thicknesses. After the operations are completed, the resulting image is analyzed to calculate a series of geometric parameters such as region area, perimeter, density, and elongation, forming a parameter set. This parameter set ultimately constitutes the morphological feature data. By concatenating the histogram vector representing texture with the parameter set representing morphology, a higher-dimensional feature vector is formed, which is the multi-feature dataset.

[0038] Optionally, generating the defect probability map includes: Obtain the weight parameters set for different defect types; The texture feature data and morphological feature data in the multi-feature dataset are weighted using the weight parameters to generate a defect score; The defect probability map is generated based on the defect score, wherein the pixel values ​​of the defect probability map represent the confidence level of the presence of a defect.

[0039] Specifically, using a pre-set training dataset containing known defect samples of various types, each component of each feature in the multi-feature dataset—namely, texture and morphological features—is learned to correspond to its defect category label. This determines the contribution of each feature in identifying a specific defect, and the contribution is quantified as a weight parameter. The weight parameters are then applied to the multi-feature dataset of the current frame. For each pixel in the image or its neighborhood, a weighted calculation is performed to generate a comprehensive defect score. This calculation process can be represented as: Where S(i,j) is the defect score calculated at image coordinates (i,j). This represents the value of the k-th feature component in the multi-feature dataset at that coordinate point. This is the weight parameter corresponding to the k-th feature. Multiple feature values ​​describing the texture and morphology of a point are fused into a single scalar score using a linear weighting method. The higher the score, the more the feature combination exhibited by that point conforms to a typical defect pattern. The defect score is normalized using the Sigmoid activation function, mapping it to the interval between 0 and 1 to generate a defect probability map. The pixel value P(i,j) of each pixel in this map represents the confidence level of a defect at that location; the closer the value is to 1, the higher the probability that there is a defect at that location. For the logic of multi-feature fusion and probability map generation, see [link to documentation]. Figure 3 By dynamically assigning higher weights to morphological features for scratch defects, the final probability map clearly presents the confidence distribution of defects, providing accurate input for layered detection.

[0040] Optionally, the hierarchical detection network includes: In the first layer, the defect probability map is subjected to high-speed screening to quickly locate and generate suspected defect areas; In the second layer, a refined classification analysis is performed on the multi-feature dataset corresponding to the suspected defective areas to generate detailed inspection results; In the third layer, the classifier model in the hierarchical detection network is adaptively updated based on the detailed inspection results and preset manual feedback samples, and the defect detection results are output.

[0041] Specifically, in the first layer, the network performs high-speed filtering on the input defect probability map. A threshold segmentation algorithm is used, setting a probability threshold and identifying all pixels in the defect probability map with values ​​higher than this threshold as candidate points. Through image processing techniques such as connected component analysis, spatially adjacent candidate points are aggregated into independent regions, forming suspected defect areas generated after rapid localization. In the second layer, the high-dimensional feature vectors of the suspected defect areas, without weighted fusion, are retrieved from the original multi-feature dataset. A lightweight neural network is used as a classifier model to perform deep analysis of the features, accurately classifying the suspected defect areas, such as distinguishing between bubbles, scratches, and foreign objects, and eliminating some false defects misclassified in the first layer, generating a detailed inspection result containing information such as defect type, location, and size. In the third layer, the detailed inspection result is connected to preset human feedback samples from external input. These human feedback samples come from the review of inspection results by production line quality inspectors, especially for correcting false positives and false negatives. Upon receiving the review of the inspection results, a preset incremental learning module is activated. The incremental learning module uses these new, correctly labeled samples to fine-tune the parameters of the classifier model in the second layer. Simultaneously, this feedback is used to optimize the weight parameters in the defect probability map generation step.

[0042] Optionally, the adaptive update of the classifier model in the hierarchical detection network based on the detailed detection results and preset manual feedback samples includes: Collect an incremental training dataset consisting of false positives and false negatives; The weight parameters in the dynamic weighted fusion processing of the multi-feature dataset are optimized using the incremental training dataset. The classifier model was fine-tuned using an incremental training dataset.

[0043] Specifically, the refined inspection results output from the second layer of the hierarchical detection network are compared with the review results from human quality inspectors on the production line. If a normal area is misclassified as a defect, or a real defect is missed, the human quality inspector will mark and correct it. The manually corrected samples are automatically captured and stored, along with their corresponding multi-feature datasets and correct category labels, gradually accumulating into a continuously growing incremental training dataset. After obtaining the incremental training dataset, a dual-path optimization process is initiated. The first path optimizes the weight parameters in the dynamic weighted fusion processing of the multi-feature dataset. Using erroneous samples from the incremental training dataset, the effectiveness of the existing weight parameters is re-evaluated. For example, for a missed sample, indicating that its feature combination failed to produce a sufficiently high defect score, the system will adjust the weight parameters of the relevant features through an optimization algorithm to increase the defect score of that type of sample. Conversely, for a falsely detected sample, the weight of the corresponding feature combination will be reduced. This optimization process aims to make the generated defect probability map more accurately reflect the distribution of real defects. The second path fine-tunes the classifier model in the second layer of the hierarchical detection network using the same incremental training dataset. Based on the existing model parameters, small iterative updates are performed using new incremental data. This process adjusts the decision boundary of the classifier model, enabling it to better classify newly discovered difficult samples correctly.

[0044] Optionally, the method further includes: Generate light source control instructions based on the defect detection results; The light source control command is sent to the multi-angle adaptive imaging system to provide feedback and adjust the light source parameters of the ring-shaped multi-band LED light source.

[0045] Specifically, the defect detection results are analyzed in depth to assess the quality of the current detection. For example, if a specific type of defect is frequently detected, but its confidence level in the defect probability map is generally low, it indicates that the current light source parameters are not optimal for imaging this type of defect. Based on information such as the defect type, morphology, and detection confidence level, a set of light source control instructions is generated to optimize the imaging effect. These instructions are quantified, executable parameter sets. For example, the instructions may specifically specify adjusting the illumination angle of a specific band of the LED array in the ring-shaped multi-band LED light source to a grazing angle and increasing its luminous intensity by a certain percentage to enhance the ability to capture minor surface scratches. The light source control instructions are sent in real time to the preset hardware controller of the multi-angle adaptive imaging system. After receiving and parsing the instructions, the hardware controller immediately drives the corresponding hardware execution unit to precisely adjust the light source parameters of the ring-shaped multi-band LED light source, changing the luminous brightness of the LEDs, adjusting the on / off state of each LED light source group, and changing the illumination angle of the light source through precision transmission mechanisms such as stepper motors.

[0046] Based on the same inventive concept, such as Figure 4 As shown, the present invention also provides an online detection system for surface defects of EVA film based on machine vision, the system including... The multi-band acquisition module is used to acquire multi-band image data of the EVA film surface through a preset multi-angle adaptive imaging system; The background difference module is used to perform background difference processing on the multi-band image data to generate a difference image; The enhancement processing module is used to perform adaptive contrast enhancement processing on the difference image to obtain an enhanced image; A multi-feature extraction module is used to extract texture feature data and morphological feature data from the enhanced image to generate a multi-feature dataset. The dynamic fusion module is used to perform dynamic weighted fusion processing on the multi-feature dataset to generate a defect probability map. The layered detection module is used to input the defect probability map into the layered detection network for processing and generate defect detection results.

[0047] Example 1: To verify the feasibility of this invention in practice, it was applied to the online quality inspection process of EVA film in a photovoltaic module manufacturing company. The EVA film on this company's production line is transported at high speeds, and its surface is prone to specular reflection. Traditional inspection methods struggle to effectively detect minute defects such as bubbles, scratches, and foreign matter, thus affecting product yield. This invention aims to achieve high-speed, high-precision online identification of these defects through a complete machine vision inspection process.

[0048] In this embodiment, images of the film surface are acquired using a multi-angle adaptive imaging system. This system includes a ring light source composed of red, green, and blue LEDs and a high-speed industrial camera with a polarizing filter. The acquired multi-band image data then enters the back-end processing flow, sequentially undergoing background subtraction, adaptive contrast enhancement, multi-feature extraction and fusion, and defect probability map generation. Finally, a three-layer hierarchical detection network outputs the final information, including defect type, location, and size.

[0049] To verify the effectiveness of the present invention, a three-month continuous test was conducted on an actual production line, and the detection performance of the method of the present invention and the traditional fixed light source visual inspection method under different production conditions was recorded and compared.

[0050] During the image acquisition phase, when a tiny bubble with a diameter of approximately 0.5 mm was detected, its contrast was extremely low under normal lighting. The light source control module in this invention automatically generates light source control commands based on historical detection results showing low confidence levels for this type of defect. This commands adjust the blue LED in the ring light source to a 30-degree grazing angle and enhance its brightness. Under these illumination conditions, a distinct highlight contour is produced at the bubble's edge, while the polarizing filter effectively suppresses specular reflections from the film background, resulting in an approximately 45% improvement in the signal-to-noise ratio of the bubble feature in the acquired multi-band image data.

[0051] In the background difference processing, 1000 manually verified defect-free EVA film images were used beforehand. Pixel averaging was then applied to generate a standard defect-free template image T(i,j). During online inspection, when an image I(i,j) containing foreign matter contamination was acquired, pixel-level difference operations D(i,j) = |I(i,j) - T(i,j)| were performed. In the generated difference image, the pixel values ​​of the defect-free background texture area were close to 0, appearing as dark areas, while the foreign matter contamination area formed a bright area with a brightness value exceeding the threshold of 150 due to grayscale differences, thus being effectively separated.

[0052] For a fine scratch with low contrast in a difference image, its local grayscale standard deviation σ(i,j) is only 5.2. The adaptive contrast enhancement module calculates that the contrast parameter G(i,j) of this region is relatively high based on the global average grayscale value M=80, and performs local enhancement on this region. After enhancement, the edge details of the scratch become sharper, and the grayscale difference with the background increases from the original 15 to 60, providing high-quality input for subsequent feature extraction.

[0053] In the feature extraction stage, the enhanced scratch region is processed in parallel. Directional linear texture features are extracted using a local binary pattern algorithm; simultaneously, morphological features with a length of 15mm, an area of ​​3mm², and an aspect ratio of 25:1 are calculated through multi-scale morphological operations. These two sets of data are fused into a high-dimensional, multi-feature dataset that comprehensively describes the characteristics of the scratch.

[0054] In the generation of the defect probability map, preset weight parameters are retrieved. These parameters were determined during offline training to have a high weight in the morphological feature "aspect ratio" when identifying scratches. Through dynamic weighted fusion calculation of the multi-feature dataset, the defect score of the scratch area is much higher than that of other areas. After normalization, a linear highlight band with a confidence level of 0.95 is formed in the defect probability map.

[0055] Finally, the defect probability map was fed into a hierarchical detection network. The first layer performed high-speed screening with a probability threshold of 0.7, identifying three suspected defect areas, including the aforementioned scratch, within 5ms. The second layer retrieved the original multi-feature datasets corresponding to these three areas and performed refined analysis using a support vector machine classifier, accurately identifying the area as a "scratch" defect within 15ms and excluding two other false defects caused by tiny wrinkles in the adhesive film. The third layer's adaptive update mechanism was also validated. For example, in the first week of testing, the false negative rate for a novel "glue dot" defect was high. Operators labeled 50 false negative samples through a manual feedback interface. These samples were automatically collected to form an incremental training dataset, fine-tuning the weight parameters and classifier model. After one week, the detection rate for the "glue dot" defect increased from 65% to 97%.

[0056] Through the above process, this invention achieves accurate detection of surface defects in EVA films. Data comparison shows that the method of this invention significantly outperforms traditional methods in the detection rate of various types of defects, especially in the identification of minute and low-contrast defects, while effectively controlling the false alarm rate. The adaptive update mechanism and light source feedback control further ensure the long-term stability and robustness of the system in the face of changes in the production environment.

[0057] Table 1 Comparison of EVA film surface defect detection performance

[0058] Table 2. Impact of Adaptive Update Mechanism on "Glue Dot" Defect Detection Performance

[0059] Table 3. Impact of Adaptive Light Source Control on Scratch Detection Confidence Level

[0060] Tables 1 to 3 above record the actual application data of the present invention on the EVA film production line, and show in detail its performance in terms of detection accuracy, adaptive learning capability and intelligent control.

[0061] As can be clearly seen from Table 1, the method of the present invention significantly improves the detection rate of various defects compared with the traditional method. In particular, for defects such as tiny bubbles, scratches and glue dots that are difficult to identify by the traditional method, the detection rate is increased by 20 to 50 percentage points, and the overall false alarm rate is controlled below 1.5%, which proves the great advantages of the present invention in detection accuracy and reliability.

[0062] The data in Table 2 validates the effectiveness of the third-layer adaptive update mechanism. After fine-tuning the model by introducing human feedback samples, the detection rate of novel "glue dot" defects increased significantly from the initial 65.0% to 97.2%, indicating that the invention has the ability to continuously learn and self-improve, and can quickly adapt to new situations on the production line.

[0063] Table 3 illustrates the effects of the light source feedback closed-loop control. In adaptive light source mode, by optimizing the lighting conditions, the average detection confidence of scratch defects increased from 0.76 to 0.94, and the proportion of samples with low confidence dropped sharply from 62% to 3%. This ensures the quality of image acquisition from the source and greatly enhances the stability and robustness of detection.

[0064] It should be noted that the electrical connections between the various units described above do not necessarily represent direct or indirect connections. Any indirect connection method can be applied to the embodiments of the present invention as long as it achieves the purpose of the present invention. The above descriptions are merely exemplary embodiments of the present invention and should not be construed as limiting the scope of the present invention.

[0065] All equivalent changes and modifications made in accordance with the teachings of this invention are still within the scope of this invention. Those skilled in the art will readily conceive of other embodiments of this invention upon considering the specification and the disclosure of practical truth. This application is intended to cover any variations, uses, or adaptations of this invention that follow the general principles of this invention and include common knowledge or conventional techniques in the art not described herein.

Claims

1. A machine vision-based online detection method for surface defects in EVA film, characterized in that, include: Acquire multi-band image data of the EVA film surface through a preset multi-angle adaptive imaging system; The multi-band image data is subjected to background subtraction processing to generate a difference image; An enhanced image is obtained by performing adaptive contrast enhancement processing on the difference image; Texture feature data and morphological feature data are extracted from the enhanced image to generate a multi-feature dataset; The multi-feature dataset is dynamically weighted and fused to generate a defect probability map; The defect probability map is input into a hierarchical detection network for processing to generate defect detection results.

2. The online detection method for EVA film surface defects based on machine vision according to claim 1, characterized in that, The acquisition of multi-band image data of the EVA film surface collected by a preset multi-angle adaptive imaging system includes: Adjust the intensity and angle of the preset ring-shaped multi-band LED light source to generate a multi-band light field covering the surface of the EVA film; The reflected light generated after the interaction of multi-band light fields with the surface of EVA film is filtered by a polarizing filter to generate the original image data after suppressing reflection. The original image data after suppressing reflection is acquired and used as the multi-band image data.

3. The online detection method for EVA film surface defects based on machine vision according to claim 1, characterized in that, The generated difference image includes: Obtain a defect-free template image of a standard EVA film based on preset historical data; Calculate the pixel-level differences between the multi-band image data and the defect-free template image to generate a set of difference values; The difference image is generated based on the set of difference values, wherein high-value regions in the difference image are identified as potential defect regions.

4. The online detection method for EVA film surface defects based on machine vision according to claim 1, characterized in that, The generated enhanced image includes: Analyze the local grayscale distribution of the difference image to generate local grayscale statistics; Calculate the contrast parameter based on local grayscale statistics; The contrast parameter is applied to enhance the edge and texture details of the difference image to generate the enhanced image.

5. The online detection method for EVA film surface defects based on machine vision according to claim 1, characterized in that, The generated multi-feature dataset includes: The enhanced image is processed using the Local Binary Pattern Algorithm to generate texture feature data; Perform multi-scale morphological operations on the enhanced image to generate morphological feature data; The multi-feature dataset is generated by fusing texture feature data and morphological feature data.

6. The online detection method for EVA film surface defects based on machine vision according to claim 1, characterized in that, The generation of the defect probability map includes: Obtain the weight parameters set for different defect types; The texture feature data and morphological feature data in the multi-feature dataset are weighted using the weight parameters to generate a defect score; The defect probability map is generated based on the defect score, wherein the pixel values ​​of the defect probability map represent the confidence level of the presence of a defect.

7. The online detection method for EVA film surface defects based on machine vision according to claim 1, characterized in that, The hierarchical detection network includes: In the first layer, the defect probability map is subjected to high-speed screening to quickly locate and generate suspected defect areas; In the second layer, a refined classification analysis is performed on the multi-feature dataset corresponding to the suspected defective areas to generate detailed inspection results; In the third layer, the classifier model in the hierarchical detection network is adaptively updated based on the detailed inspection results and preset manual feedback samples, and the defect detection results are output.

8. The online detection method for EVA film surface defects based on machine vision according to claim 7, characterized in that, The adaptive updating of the classifier model in the hierarchical detection network based on the detailed detection results and preset manual feedback samples includes: Collect an incremental training dataset consisting of false positives and false negatives; The weight parameters in the dynamic weighted fusion processing of the multi-feature dataset are optimized using the incremental training dataset. The classifier model was fine-tuned using an incremental training dataset.

9. The online detection method for EVA film surface defects based on machine vision according to claim 2, characterized in that, The method further includes: Generate light source control instructions based on the defect detection results; The light source control command is sent to the multi-angle adaptive imaging system to provide feedback and adjust the light source parameters of the ring-shaped multi-band LED light source.

10. A machine vision-based online detection system for surface defects in EVA film, applied to the machine vision-based online detection method for surface defects in EVA film as described in any one of claims 1-9, characterized in that, The system includes: The multi-band acquisition module is used to acquire multi-band image data of the EVA film surface through a preset multi-angle adaptive imaging system; The background difference module is used to perform background difference processing on the multi-band image data to generate a difference image; The enhancement processing module is used to perform adaptive contrast enhancement processing on the difference image to obtain an enhanced image; A multi-feature extraction module is used to extract texture feature data and morphological feature data from the enhanced image to generate a multi-feature dataset. The dynamic fusion module is used to perform dynamic weighted fusion processing on the multi-feature dataset to generate a defect probability map. The layered detection module is used to input the defect probability map into the layered detection network for processing and generate defect detection results.

Images