Method and system for detecting surface defects of photovoltaic cell pieces

Abstract

The application discloses a photovoltaic cell surface defect detection method and system, comprising the following steps: acquiring a cell surface visible light image, an internal electrical defect EL image and a lattice defect PL image, and pre-processing the collected original image; adopting an adaptive background suppression algorithm to perform background stripping on the pre-processed image, and retaining irregular, low-contrast abnormal feature areas in the image. The photovoltaic cell surface defect detection method and system of the application embodiment synchronously collect multi-modal images, combine adaptive background suppression to strip background interference such as silicon wafer texturing and grid line structure, and then enhance and enlarge the contrast of weak defects and the substrate through multi-scale feature enhancement, thereby solving the problem that weak defects are easily covered by the background, the missing detection and misjudgment rate is high in the prior art, improving the detection accuracy of various defects, especially small low-contrast defects, and guaranteeing the quality of photovoltaic cells.

Description

Technical Field

[0001] This application relates to the technical field of photovoltaic cells, and in particular to a method and system for detecting surface defects in photovoltaic cells. Background Technology

[0002] Photovoltaic cells are prone to various surface and subsurface defects during processes such as texturing, coating, screen printing, and sintering, including scratches, dirt, edge chipping, grid breakage, microcracks, and abnormal surface paste. These defects can reduce the photoelectric conversion efficiency of the cells, cause safety hazards such as local leakage, hot spots, and module burnout, and also lead to power degradation and reduced yield of the finished product. Therefore, it is necessary to carry out full-process inspection of surface defects during the production process, promptly remove defective cells, ensure the output performance and long-term operational reliability of the cells, and improve the product yield and consistency of finished product quality.

[0003] Currently, existing technologies generally use AOI visible light vision or EL / PL imaging vision equipment to complete the defect detection of photovoltaic cells. However, during the detection process, due to the small pixel ratio of minute defects such as microcracks, tiny dirt spots, fine grid breaks, and shallow scratches, and the small difference in grayscale contrast with the silicon wafer substrate, traditional image processing algorithms such as threshold segmentation, edge extraction, and morphological processing are easily interfered with by background information such as uneven illumination of the wafer, inherent texture of the silicon wafer, and grid structure. The minute defect features are easily covered by the substrate background, which leads to frequent missed detections and false detections during the detection process, affecting the accuracy of the detection results. Summary of the Invention

[0004] This application aims to at least partially address one of the technical problems in the related art.

[0005] Therefore, one objective of this application is to provide a method and system for detecting surface defects in photovoltaic cells. By simultaneously acquiring multimodal images, combining adaptive background suppression to remove background interference such as texturing textures and grid structures from the silicon wafer, and then enhancing and amplifying the contrast between weak defects and the substrate through multi-scale feature enhancement, this method solves the problem in the prior art where weak defects are easily covered by the background and the rate of missed detection and false detection is high. It improves the detection accuracy of various defects, especially small low-contrast defects, and ensures the quality of photovoltaic cells leaving the factory.

[0006] To achieve the above objectives, the first aspect of this application proposes a method for detecting surface defects in photovoltaic cells, comprising the following steps: Multimodal image synchronous acquisition: acquire visible light images of the surface appearance of the solar cell, EL images of internal electrical defects, and PL images of lattice defects, and preprocess the acquired raw images; Complex background feature suppression processing: An adaptive background suppression algorithm is used to strip the background from the preprocessed image, while retaining irregular and low-contrast abnormal feature regions in the image; Multi-scale weak defect feature enhancement: Construct a multi-scale feature enhancement algorithm to enhance the features of small pixel areas and low gray-level difference areas in the image, amplify the gray-level contrast between the defect and the substrate, and highlight the edge and regional features of weak defects. Initial defect identification: The features of visible light images, EL images, and PL images that have undergone background suppression and feature enhancement are fused across modes to initially locate the defect areas on the surface of the solar cell; Defect verification: Establish defect judgment constraint rules, perform secondary screening of suspected defect areas, eliminate false defects and false detection targets, and at the same time, verify and check small areas with weak features and suspected missed detections. Defect grading and labeling: Label the actual defect areas after verification, record the defect information, and classify the cells into good products, slightly defective products, and severely defective products.

[0007] In addition, the photovoltaic cell surface defect detection method and system proposed in this application may also have the following additional technical features: In one embodiment of this application, in the multimodal image synchronous acquisition step, an integrated detection device that integrates AOI visible light imaging, EL electroluminescence imaging, and PL photoluminescence imaging is used to acquire images.

[0008] In one embodiment of this application, the adaptive background suppression algorithm includes the following steps in the complex background feature suppression processing step: Background texture feature segmentation: Extract three basic features from the solar cell image: grayscale mean, texture gradient, and edge frequency. Divide the target regions into three categories: regular texture regions of silicon wafer texturing, regular structure regions of grid lines, and suspected defect and abnormal regions. Preliminarily distinguish the background from potential defect regions. Background interference weight adaptive calculation: For areas with regular texturing texture and areas with regular grid line structure, the background interference weight of each pixel is dynamically calculated in combination with the illumination fluctuation characteristics of mass production images of battery cells. Background weakening processing: Gradual grayscale weakening and texture smoothing processing are applied to the uniform texturing texture and continuous gate lines of the silicon wafer. While preserving the basic outline of the background, high-frequency texture noise is eliminated, and small, low-contrast irregular defect feature areas are completely preserved to achieve background interference removal. Defect feature boundary repair and correction: Corrects the blurring of weak defect edges and feature distortion during the background weakening process, enhances the boundary difference between low-contrast defect areas and normal substrate, and obtains an optimized image with a clean background and complete defect features.

[0009] In one embodiment of this application, in the multi-scale weak defect feature enhancement step, a layered feature extraction method is used to extract weak defects, including microcracks, tiny dirt spots, fine grid breaks, and shallow scratches.

[0010] In one embodiment of this application, the defect types in the initial defect identification step include scratches, dirt spots, broken grids, microcracks, and printing anomalies.

[0011] In one embodiment of this application, in the defect verification step, the defect determination constraint rules include grayscale feature constraint rules, morphological topology constraint rules, pixel size constraint rules, position distribution constraint rules and edge gradient constraint rules, and false defect detection targets include silicon wafer texture, illumination fluctuation and gate line edges.

[0012] A photovoltaic cell surface defect detection system includes an image acquisition module, a background feature suppression module, a defect enhancement module, a defect recognition module, and a verification module, wherein... The image acquisition module is used to acquire visible light images of the surface appearance of the solar cell, EL images of internal electrical defects, and PL images of lattice defects, and to preprocess the images. The background feature suppression module is used to strip the background from the preprocessed image to obtain an optimized image with a clean background and complete defect features. The defect enhancement module is used to amplify the grayscale contrast between the defect and the substrate, highlighting the edge and regional features of weak defects. The defect identification module is used to initially locate the defect area on the surface of the battery cell, and cooperates with the verification module to perform secondary screening of suspected defect areas, eliminate false defect detection targets, and output the real defects.

[0013] Compared with the prior art, the technical solution provided in this application has the following beneficial effects: I. This application uses an adaptive background suppression algorithm to remove complex background noise such as silicon wafer texturing texture and dense gate lines. Combined with multi-scale feature enhancement technology, it significantly improves the feature recognition of small targets and low-contrast defects such as microcracks, tiny dirt spots, fine gate breaks, and shallow scratches. It solves the drawback of traditional threshold, edge, and morphological algorithms where defect features are easily covered by the substrate background, reduces the false negative rate and false positive rate of subtle defects, and improves the defect detection accuracy.

[0014] Second, this application adopts a multi-modal image synchronous acquisition mechanism to integrate AOI visible light, EL, PL multi-dimensional defect information, while taking into account surface appearance defects as well as internal lattice and electrical latent defects. The detection dimensions are more comprehensive, avoiding the problem of missed detection caused by incomplete single-modal detection information. At the same time, through the defect verification mechanism, false defects caused by background interference are effectively eliminated, further improving the detection accuracy. It can be adapted to the high-speed detection scenario of mass production line.

[0015] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0016] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a flowchart of the photovoltaic cell surface defect detection method of this application; Figure 2 This is a flowchart of the adaptive background suppression algorithm in this application; Figure 3 This is a schematic diagram of the structure of the photovoltaic cell surface defect detection system of this application.

[0017] As shown in the figure: 1. Image acquisition module; 2. Background feature suppression module; 3. Defect enhancement module; 4. Defect recognition module; 5. Verification module. Detailed Implementation

[0018] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.

[0019] The following describes the photovoltaic cell surface defect detection method and system according to embodiments of this application with reference to the accompanying drawings.

[0020] like Figure 1 , Figure 2 As shown in the figure, a method for detecting surface defects in photovoltaic cells according to an embodiment of this application includes: S1, Multimodal Image Synchronous Acquisition: Acquire visible light images of the surface appearance of the solar cell, EL images of internal electrical defects, and PL images of lattice defects, and preprocess the acquired raw images.

[0021] Specifically, images are acquired using an integrated detection device that combines AOI visible light imaging, EL electroluminescence imaging, and PL photoluminescence imaging.

[0022] Then, an adaptive grayscale equalization algorithm is used to compensate for the illumination of the image, eliminating noise interference such as uneven illumination, local reflection, and color difference on the surface of the battery cell under production line conditions. Then, a Gaussian filtering algorithm is used to remove salt-and-pepper noise and Gaussian noise during the image acquisition process, while preserving the detailed features of the defect edges, avoiding the problem of blurring of subtle defect features caused by traditional filtering methods, thus completing the basic image optimization.

[0023] It should be noted that the AOI visible light imaging uses an industrial CMOS camera with a resolution of 2048×2048 pixels and a ring-shaped coaxial white light source. The EL electroluminescence imaging uses an 850nm near-infrared camera with a constant current pulsed power supply probe and an injection current of 0.8–1.2A. The PL photoluminescence imaging uses a 940nm laser excitation source and a short-wave infrared camera. All three use a common reference guide rail and have a uniform pixel resolution of 5μm / pixel.

[0024] Furthermore, adaptive gray-scale equalization algorithm and Gaussian filtering algorithm are existing technologies and are conventional techniques in this field, so their specific calculation processes will not be elaborated upon.

[0025] Understandably, by simultaneously acquiring three types of image information—appearance, electrical properties, and lattice—it is possible to fully cover both visible defects on the surface of the solar cell and hidden internal defects.

[0026] By preprocessing the images, problems such as unstable lighting and image noise interference in mass production scenarios can be solved. The overall brightness difference of the battery cells and local reflection defects can be effectively corrected, and invalid noise generated by image acquisition can be eliminated. While reducing noise, the edge details of weak defects such as microcracks and shallow scratches can be preserved to the maximum extent. The preprocessing process can avoid damaging the effective defect features and provide a high-quality image foundation for subsequent feature extraction and defect recognition.

[0027] S2, Complex Background Feature Suppression Processing: An adaptive background suppression algorithm is used to strip the background from the preprocessed image, preserving irregular and low-contrast abnormal feature regions in the image.

[0028] The adaptive background suppression algorithm includes the following steps: D1. Background texture feature segmentation: Extract three basic features from the solar cell image: grayscale mean, texture gradient, and edge frequency. Divide the image into three target regions: regular texture region of silicon wafer texturing, regular structure region of grid line, and suspected defect and abnormal region. Preliminarily distinguish the background from potential defect regions.

[0029] Among them, the average gray level is: the average gray level G of the 16×16 pixel region window; Texture gradient: The Sobel operator calculates the mean magnitude of the gradient in the x / y directions, Grad; Edge frequency: The percentage R of pixels within a window whose gradient is greater than the threshold T0; Region determination rules: ① Texturing area: G fluctuation < 8, Grad < 15, R < 0.12; ② Regular grid area: Long continuous area, fixed width 20-40 pixels, edge frequency R > 0.3; ③ Suspected defect area: Gray scale abrupt change, gradient fluctuation > 25, no fixed continuous shape.

[0030] D2. Adaptive Calculation of Background Interference Weights: For regular textured areas and regular grid line structure areas, the background interference weights of each pixel are dynamically calculated based on the illumination fluctuation characteristics of mass production images of battery cells. High suppression weights are assigned to base textures and grid line structures with high repetition and strong regularity, while zero suppression weights are assigned to suspected defect areas with gray-scale abrupt changes and irregularities.

[0031] The calculation formula is: weight W(x,y)=1−exp(−Grad(x,y) / 30); For the texturing / grid area, W is set to 0.7–0.95 (high suppression), and for the suspected defect area, W=0.

[0032] D3. Background weakening processing: The uniform texturing texture and continuous gate lines of the silicon wafer are gradually weakened in grayscale and the texture is smoothed. While preserving the basic outline of the background, high-frequency texture noise is eliminated. Small, low-contrast irregular defect feature areas are completely preserved to achieve background interference removal.

[0033] It should be noted that the quantization parameters for background weakening and boundary repair are as follows: the base texture grayscale progressive decay coefficient is 0.65, the number of raster smoothing iterations is 3, and the defect boundary gradient enhancement coefficient is 1.8 to correct edge blur distortion.

[0034] D4. Defect Feature Boundary Repair and Correction: Corrects the blurring of weak defect edges and feature distortion during the background weakening process, enhances the boundary difference between low-contrast defect areas and normal substrates, and obtains an optimized image with a clean background and complete defect features.

[0035] Understandably, this step can accurately distinguish between normal substrate texture, grid line structure and defect abnormal features of solar cells, significantly suppress repetitive and highly interfering background information in the image, effectively remove the occlusion interference of complex background on subtle defects, and highlight low-contrast small abnormal areas. This solves the problem that traditional algorithms are easily affected by silicon wafer texture and grid line interference, and defect features are submerged by the background, thus reducing the probability of subsequent false detections and false negatives.

[0036] S3, Multi-scale weak defect feature enhancement: Construct a multi-scale feature enhancement algorithm to enhance the features of small pixel areas and low gray-level difference areas in the image, amplify the gray-level contrast between the defect and the substrate, and highlight the edge and regional features of weak defects, including micro-cracks, small dirt spots, fine grid breaks and shallow scratches.

[0037] Specifically, a multi-scale feature enhancement algorithm is used to enhance the features of small pixel regions and low gray-scale difference regions in the image, amplify the gray-scale contrast between the defect and the substrate, highlight the edge and regional features of weak defects, and make up for the problem that traditional algorithms are not good at extracting features of small targets and low-contrast defects.

[0038] The multi-scale feature enhancement algorithm includes the following steps: Multi-scale layered construction: Three progressive scale branches are set up, corresponding to micro-dot defects, narrow line defects, and small area defects, respectively. Among them, small scale (1-5 pixels): gain coefficient 2.8, for micro-dirt and micro-crack breaks; medium scale (5-20 pixels): gain coefficient 2.0, for shallow scratches and fine grid breaks; large scale (>20 pixels): gain coefficient 1.2, for sheet-like printing color difference defects.

[0039] Image low-frequency background and high-frequency feature separation: The clean image after background suppression is decomposed in the frequency domain (Fourier transform to separate high and low frequencies, with a low frequency cutoff frequency of 12Hz to retain the smooth substrate; all high frequency components (>12Hz) are retained for defect enhancement), separating the low-frequency smooth background component and the high-frequency detail component of the image, retaining effective high-frequency features such as defect edges and texture abrupt changes, and suppressing residual uniform background noise.

[0040] Defect saliency calculation: Calculate the local gray-level variance and gradient response value, dynamically determine the saliency of defects in the region, automatically match high enhancement gain for weak defect regions with low variance and low contrast, and match low gain or zero gain for normal and stable base regions.

[0041] Among them, local window grayscale variance > 12 is judged as a high-response defect region, and high enhancement gain is automatically matched; variance < 5 is a normal base, and the gain is set to 0.3.

[0042] Multi-scale feature weighted fusion reconstruction: Adaptive weighted fusion is performed on the feature maps enhanced at three scales. The fusion weights are dynamically allocated according to the defect response intensity of each branch. Effective defect features at each scale are retained first, and various subtle defects such as point, line and sheet are complementaryly covered. Finally, an enhanced feature image with high defect contrast, clear edges and pure background is reconstructed.

[0043] Understandably, this step, which involves targeted feature enhancement for subtle defects with small pixel proportions and low grayscale differences, can effectively improve the feature recognition of weak targets such as tiny dirt spots, microcracks, and shallow scratches, widen the grayscale difference between the defect area and the normal substrate, solve the problem that traditional algorithms cannot effectively extract the features of weak defects, and make the difficult-to-identify subtle defects clearly distinguishable, providing core feature support for high-precision defect identification.

[0044] S4, Initial Defect Identification: The visible light image, EL image, and PL image features, after background suppression and feature enhancement, are fused across modes to initially locate the defect area on the surface of the solar cell. Among them, the defect types include scratches, dirt spots, broken grids, microcracks, and printing anomalies.

[0045] Specifically, the visible light image, EL image, and PL image features that have undergone background suppression and feature enhancement are fused across modalities to integrate multi-dimensional feature information of surface appearance defects, internal electrical defects, and lattice defects. The fused feature map is then traversed and identified using a lightweight deep learning detection model (model type: U-Net lightweight semantic segmentation network) that has been trained, thus initially locating various defect regions on the surface and subsurface of the solar cell.

[0046] It should be noted that the lightweight multimodal U-Net network has the following specific parameters and structure: Network structure: A symmetrical encoding and decoding structure with 4 encoding layers and 4 decoding layers is adopted, and deep redundant convolutional branches are eliminated; the basic number of channels is set to 32, 64, 128, and 256, increasing layer by layer, and high-dimensional redundant channels are removed.

[0047] Lightweight improvement methods: replace standard convolution with depthwise separable convolution, replace max pooling with embedded attention pooling downsampling, and reduce redundant parameters in the top fully connected layer of traditional U-Net to reduce the number of model parameters and computational cost.

[0048] Multimodal feature fusion layer: A three-channel modal fusion layer is set at the encoding input end to perform channel-wise weighted fusion of the three types of image features: visible light, EL, and PL. Cross-modal feature complementarity is achieved through 1×1 convolution dimensionality reduction, and multi-dimensional defect information is preserved.

[0049] Training dataset: Composed of mass-produced real samples, including low-contrast micro-defect samples such as micro-cracks, shallow scratches, tiny dirt spots, fine grid breaks, and printing defects, covering multiple batches, multiple lighting conditions, and multiple textures.

[0050] Loss function: A composite loss function combining Dice loss and cross-entropy loss is used to solve the problem of imbalance between positive and negative samples caused by the extremely small proportion of pixels with minor defects.

[0051] Model accuracy metrics: After training, the model achieves an Intersection over Union (IoU) of ≥92% for segmenting minor defects and a pixel-level detection accuracy of ≥99%, meeting the requirements for high-speed and high-precision mass production inspection. It performs traversal recognition of the fused feature map to initially locate various defect regions on the surface and subsurface of the solar cell, distinguishing defect types including scratches, dirt spots, broken grids, microcracks, and printing anomalies. Understandably, by fusing complementary features from multimodal images, taking into account multidimensional defect information such as appearance, electrical properties, and lattice structure, the limitations of single-modal detection can be overcome. Through the feature recognition capabilities of deep learning models, various defect morphologies can be adaptively adapted, enabling rapid and comprehensive preliminary localization and classification of all types of defects, thus ensuring the comprehensiveness and efficiency of detection.

[0052] S5, Defect Detection: Establish defect judgment constraint rules, perform secondary screening of suspected defect areas, eliminate false defect detection targets, and at the same time, verify and check small areas with weak features and suspected missed detections. False defect detection targets include silicon wafer texture, light fluctuation and gate line edges.

[0053] Specifically, precise verification is carried out on the initial identification results. Defect judgment constraint rules are established by combining defect morphological features, grayscale threshold range, regional pixel size, and position distribution patterns. The defect judgment constraint rules include grayscale feature constraint rules, morphological topology constraint rules, pixel size constraint rules, position distribution constraint rules, and edge gradient constraint rules.

[0054] Among them, the grayscale feature constraint rule is as follows: the average grayscale difference of the 3×3 neighborhood is used as the judgment standard, and the effective defect grayscale difference threshold is set to ≥18; the grayscale difference <18 and the grayscale of the area is continuously and gradually changed as a pseudo defect of lighting and base color difference, and the grayscale abrupt change and exceeding the threshold is judged as a real defect.

[0055] Morphological topology constraint rules: Dot-like dirt points satisfy a pixel area of ​​3–20 pixels and a roundness of ≥0.7; scratches and hidden cracks satisfy an aspect ratio of ≥5 and continuous linear topology; broken gates satisfy a local missing length of ≥3 pixels; jagged edges of gates and texture protrusions with a roundness of <0.3 and no continuous extension characteristics are uniformly judged as pseudo-defects.

[0056] Pixel size constraint rule: The effective defect pixel range is set to 3–80 pixels; single-point random noise with a pixel size <3 and large-area dark areas with illumination >80 pixels are directly judged as invalid pseudo-defect areas, and only real and subtle defect targets within the range are retained.

[0057] Location distribution constraint rules: The standard threshold is strictly implemented in the central area of ​​the cell, the gray scale threshold of the grid line area is increased by 20%, and the size threshold of the edge area is increased by 30%; regular abnormal responses caused by edge reflective spots and grid line periodic textures are judged as false defects, and irregular abnormalities are judged as real defects.

[0058] Edge gradient constraint rules: Set the effective defect gradient amplitude to ≥25 and the gradient change rate to abrupt change; if the base texture gradient amplitude is <10 and the gradient change is continuous and gentle, it belongs to normal base fluctuation and is judged as a false defect and removed.

[0059] Meanwhile, false defects caused by silicon wafer texture, lighting fluctuations, and gate edge interference are eliminated, and small areas with weak features or suspected missed detections are re-verified to improve detection accuracy.

[0060] Understandably, filtering out false defects caused by background interference through multiple constraint rules effectively reduces false detection problems caused by texture, lighting, and raster lines. At the same time, special verification of suspected areas with weak defects effectively avoids missing minor defects. The dual verification mechanism improves the accuracy of defect detection.

[0061] S6, Defect Classification and Labeling: Label the actual defect areas after verification, record defect information, and classify the cells into good products, slightly defective products, and severely defective products.

[0062] Specifically, the actual defective areas after verification are accurately marked, and key information such as defect location, size, type, and severity are recorded. Defective cells are classified according to industry cell grading standards, distinguishing between good products, slightly defective products, and severely defective products.

[0063] Simultaneously, image data, defect data, and classification results from each inspection are archived and stored to form an inspection data ledger, providing data support for production process optimization and product quality traceability.

[0064] In addition, the thresholds for judging good products, slightly defective products, and severely defective products can be preset according to the quality standards of the target production line to adapt to the quality control needs of different scenarios.

[0065] like Figure 3 As shown in the figure, a photovoltaic cell surface defect detection system according to an embodiment of this application includes an image acquisition module, a background feature suppression module, a defect enhancement module, a defect identification module, and a verification module. The image acquisition module is used to acquire visible light images of the cell surface appearance, EL images of internal electrical defects, and PL images of lattice defects, and to preprocess the images.

[0066] The background feature suppression module is used to strip the background from the preprocessed image to obtain an optimized image with a clean background and complete defect features.

[0067] The defect enhancement module is used to amplify the grayscale contrast between defects and the substrate, highlighting the edges and regional features of weak defects.

[0068] The defect identification module is used to initially locate defect areas on the surface of the battery cell, and works with the verification module to perform secondary screening of suspected defect areas, eliminate false defects and output the real defects.

[0069] Understandably, during system operation, the image acquisition module synchronously acquires multimodal images to complete preprocessing and optimization, eliminating uneven illumination and noise interference in the original images. Subsequently, the background feature suppression module adaptively removes background interference such as silicon wafer textures and gate lines, fully preserving potential defect features. The defect enhancement module performs targeted enhancement for weak defects of different sizes, amplifying the contrast between the defect and the substrate. Finally, the defect identification module first completes the initial defect localization, and then the verification module eliminates false defects and verifies suspected missed areas through multiple constraint rules, ultimately outputting the true defect results, balancing the comprehensiveness and accuracy of detection throughout the entire process.

[0070] In summary, the photovoltaic cell surface defect detection method and system of this application significantly improves the feature recognition of small targets and low-contrast defects such as microcracks, tiny dirt spots, fine grid breaks, and shallow scratches by using an adaptive background suppression algorithm to remove complex background noise such as silicon wafer texturing texture and dense grid lines, combined with multi-scale feature enhancement technology. This solves the drawback of traditional threshold, edge, and morphological algorithms where defect features are easily covered by the substrate background, reduces the false negative rate and false positive rate of minor defects, and improves the defect detection accuracy.

[0071] Furthermore, by adopting a multi-modal image synchronous acquisition mechanism, it integrates multi-dimensional defect information from AOI visible light, EL, and PL, while also taking into account surface appearance defects as well as internal lattice and electrical latent defects. This provides a more comprehensive detection dimension and avoids the problem of missed detection caused by incomplete information from a single modality. At the same time, through a defect verification mechanism, it effectively eliminates false defects caused by background interference, further improving the detection accuracy and making it suitable for high-speed inspection scenarios in mass production lines.

[0072] In the description of this specification, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0073] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0074] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A method for detecting surface defects in photovoltaic cells, characterized in that, Includes the following steps: Multimodal image synchronous acquisition: acquire visible light images of the surface appearance of the solar cell, EL images of internal electrical defects, and PL images of lattice defects, and preprocess the acquired raw images; Complex background feature suppression processing: An adaptive background suppression algorithm is used to strip the background from the preprocessed image, while retaining irregular and low-contrast abnormal feature regions in the image; Multi-scale weak defect feature enhancement: Construct a multi-scale feature enhancement algorithm to enhance the features of small pixel areas and low gray-level difference areas in the image, amplify the gray-level contrast between the defect and the substrate, and highlight the edge and regional features of weak defects. Initial defect identification: The features of visible light images, EL images, and PL images that have undergone background suppression and feature enhancement are fused across modes to initially locate the defect areas on the surface of the solar cell; Defect verification: Establish defect judgment constraint rules, perform secondary screening of suspected defect areas, eliminate false defects and false detection targets, and at the same time, verify and check small areas with weak features and suspected missed detections. Defect grading and labeling: Label the actual defect areas after verification, record the defect information, and classify the cells into good products, slightly defective products, and severely defective products.

2. The method for detecting surface defects in photovoltaic cells according to claim 1, characterized in that, In the multimodal image synchronous acquisition step, images are acquired using an integrated detection device that combines AOI visible light imaging, EL electroluminescence imaging, and PL photoluminescence imaging.

3. The method for detecting surface defects in photovoltaic cells according to claim 1, characterized in that, In the complex background feature suppression processing step, the adaptive background suppression algorithm includes the following steps: Background texture feature segmentation: Extract three basic features from the solar cell image: grayscale mean, texture gradient, and edge frequency. Divide the target regions into three categories: regular texture regions of silicon wafer texturing, regular structure regions of grid lines, and suspected defect and abnormal regions. Preliminarily distinguish the background from potential defect regions. Background interference weight adaptive calculation: For areas with regular texturing texture and areas with regular grid line structure, the background interference weight of each pixel is dynamically calculated in combination with the illumination fluctuation characteristics of mass production images of battery cells. Background weakening processing: Gradual grayscale weakening and texture smoothing processing are applied to the uniform texturing texture and continuous gate lines of the silicon wafer. While preserving the basic outline of the background, high-frequency texture noise is eliminated, and small, low-contrast irregular defect feature areas are completely preserved to achieve background interference removal. Defect feature boundary repair and correction: Corrects the blurring of weak defect edges and feature distortion during the background weakening process, enhances the boundary difference between low-contrast defect areas and normal substrate, and obtains an optimized image with a clean background and complete defect features.

4. The method for detecting surface defects in photovoltaic cells according to claim 1, characterized in that, In the multi-scale weak defect feature enhancement step, a layered feature extraction method is used to extract weak defects, including microcracks, tiny dirt spots, fine grid breaks, and shallow scratches.

5. The method for detecting surface defects in photovoltaic cells according to claim 1, characterized in that, In the initial defect identification step, the defect types include scratches, dirt spots, broken grids, microcracks, and printing anomalies.

6. The method for detecting surface defects in photovoltaic cells according to claim 1, characterized in that, In the defect verification step, the defect determination constraint rules include grayscale feature constraint rules, morphological topology constraint rules, pixel size constraint rules, positional distribution constraint rules, and edge gradient constraint rules. False defects and false positives include silicon wafer texture, illumination fluctuations, and gate line edges.

7. A photovoltaic cell surface defect detection system, applied to the photovoltaic cell surface defect detection method as described in any one of claims 1-6, characterized in that, It includes an image acquisition module, a background feature suppression module, a defect enhancement module, a defect recognition module, and a verification module, among which, The image acquisition module is used to acquire visible light images of the surface appearance of the solar cell, EL images of internal electrical defects, and PL images of lattice defects, and to preprocess the images. The background feature suppression module is used to strip the background from the preprocessed image to obtain an optimized image with a clean background and complete defect features. The defect enhancement module is used to amplify the grayscale contrast between the defect and the substrate, highlighting the edge and regional features of weak defects. The defect identification module is used to initially locate the defect area on the surface of the battery cell, and cooperates with the verification module to perform secondary screening of suspected defect areas, eliminate false defect detection targets, and output the real defects.

Images