A photovoltaic module detection method and system based on image recognition

Abstract

The application discloses a photovoltaic module detection method and system based on image recognition. The method comprises the following steps: acquiring image information of the same scene under at least two different light vibration modes; based on the image difference under different light vibration modes, separating and suppressing interference such as mirror reflection to obtain an image free of interference; performing local detail enhancement and non-random signal smoothing on the interference-removed image to highlight defects and reduce noise; and identifying defects based on the processed image and evaluating the reliability reflected by the degree of separation of defect features and interference information. The application separates reflection interference and enhances defect features under single equipment conditions by using the difference in light vibration modes, thereby solving the problem that reflection interference and real defects are difficult to distinguish under outdoor light.

Description

Technical Field

[0001] This application relates to the field of photovoltaic module testing technology, and in particular to a photovoltaic module testing method and system based on image recognition. Background Technology

[0002] During long-term outdoor operation, photovoltaic (PV) modules are affected by environmental factors such as ultraviolet radiation, temperature cycling, and wind and sand erosion. As a result, their surfaces gradually develop various defects, including microcracks, hot spots, snail trails, frame corrosion, and junction box aging. These defects not only reduce the photoelectric conversion efficiency of PV modules but can also, in severe cases, cause localized overheating or even fire hazards. Therefore, regular inspection of operating PV modules to detect and locate defects is a crucial aspect of PV power plant operation and maintenance. Automated inspection methods based on image recognition, due to their non-contact and high-efficiency advantages, are gradually replacing traditional manual visual inspections.

[0003] However, optical inspection of photovoltaic (PV) modules faces unique challenges. The surface of PV modules is typically covered with tempered glass panels, which produce strong specular reflections under natural light. Simultaneously, the module surface may be contaminated with interfering substances such as dust, bird droppings, oil, and water stains. These interfering substances may exhibit similar visual characteristics to real defects in ordinary visible light images. Traditional visible light image inspection methods, which only collect single brightness information, struggle to effectively distinguish between specular reflection interference and genuine surface defects. Furthermore, PV modules are usually installed outdoors, where the intensity and angle of natural sunlight vary with time and weather conditions. Images acquired at different times exhibit differences in brightness and contrast, further increasing the difficulty of defect identification.

[0004] To address the aforementioned issues, existing technologies primarily employ multispectral or multi-angle imaging schemes. These schemes acquire multiple images at different spectral bands or shooting angles, utilizing the differences between the images to enhance defect features or suppress interference. However, multispectral schemes require multiple filters and cameras for different bands, resulting in high hardware costs and system complexity. Multi-angle schemes rely on mechanical rotating platforms, leading to slow detection speeds and difficulty in meeting the rapid inspection needs of large-scale photovoltaic power plants. None of these solutions can effectively separate reflected interference from actual defects under conditions of a single acquisition device and a single shooting position.

[0005] Therefore, there is an urgent need for a method that can quickly and accurately detect surface defects of photovoltaic modules by utilizing the image differences of the same scene under different light vibration modes under outdoor natural lighting conditions, while preserving the true defect characteristics and effectively suppressing the effects of reflection interference and changes in ambient lighting. Summary of the Invention

[0006] To address the problems mentioned in the background art, this application provides a photovoltaic module inspection method and system based on image recognition. This method utilizes the differences in image information under different light vibration modes to achieve effective separation of reflection interference and reliable extraction of defect features under single acquisition equipment conditions.

[0007] In a first aspect, this application provides a photovoltaic module detection method based on image recognition, comprising the following steps: Acquire image information of the same scene under at least two different light vibration modes; Based on the image information under the different light vibration modes, interference information with specific light vibration modes is separated and suppressed to obtain an image without interference; The image after removing interference is subjected to local detail enhancement and non-random signal smoothing to highlight defect features and reduce electronic noise; Based on the image after local detail enhancement and non-random signal smoothing, defects on the surface of the photovoltaic module are identified, and the reliability of the defects is evaluated according to the distinguishability between defect features and interference information.

[0008] Optionally, the step of separating and suppressing interference information with specific light vibration modes based on image information under different light vibration modes to obtain an interference-free image includes: Acquire image information of the same scene under at least two different light vibration modes; Calculate the rate of change of light vibration patterns in a local region of the image; Determine whether the rate of change of light vibration patterns in a local area exceeds a preset threshold; The parameters for separating interference information are dynamically adjusted when the threshold is exceeded. By combining local texture and brightness gradient information to separate and suppress interference information, an image with interference removed is obtained.

[0009] Optionally, the step of performing local detail enhancement and non-random signal smoothing on the interference-removed image includes: Acquire brightness information, light vibration intensity information, and light vibration direction information of the image after interference removal; The similarity between two image patches at different locations is evaluated based on brightness information, light vibration intensity information, and light vibration direction information. The images are weighted and averaged based on the similarity of image patches to obtain images with clearer defect features.

[0010] Optionally, the step of evaluating the similarity between two image patches at different locations includes: Each image block is divided into local regions based on the light vibration intensity and light vibration direction of each pixel; Calculate the statistical distribution characteristics of each local region; Compare the statistical distribution characteristics of the corresponding local regions; The similarity between two image patches is calculated based on differences in statistical distribution characteristics and area proportions.

[0011] Optionally, the step of performing a weighted average based on image patch similarity includes: Identify a set of pixels whose light vibration patterns around the pixel to be processed are extremely similar; Initial weights are assigned based on brightness differences; The initial weights are adjusted based on minute differences in the intensity and direction of light vibrations; A weighted average is calculated using the adjusted weights.

[0012] Optionally, the step of identifying a set of pixels with extremely small differences in light vibration patterns includes: Deblurring is performed on a local area of ​​the image; Calculate the difference in light beam vibration intensity and the angle between the light beam vibration directions; Compare the difference with the first threshold, and the included angle with the second threshold; When all values ​​are less than the corresponding threshold, they are identified as pixels with minimal pattern differences.

[0013] Optionally, the step of adjusting the initial weights includes: Obtain the current ambient temperature information of the image acquisition device; Based on the ambient temperature, the correlation between polarization filter material deformation and light vibration mode information distortion is queried to obtain polarization information distortion compensation parameters; The initial weights are adjusted by taking into account the minute differences between pixels and the compensation parameters.

[0014] Optionally, the step of identifying a set of pixels with extremely small differences in light vibration patterns further includes: Deblurring is performed on a local area of ​​the image; Identify areas of localized light scattering and refraction caused by oil or water stains; Local correction is performed on the light vibration pattern information of a local area; Calculate the corrected difference in light vibration intensity and the included angle of direction; Pixels with minimal pattern differences are identified based on the comparison results with a threshold.

[0015] Optionally, the step of identifying oil or water stain areas includes: Acquire the intensity and direction distribution of light vibrations in a local area, as well as brightness texture features and edge gradient information; Compare the distribution characteristics with the preset range of oil or water stain characteristics; Compare texture features and edge gradients with the features of real dirt; The comparison results determine whether the area is an oil or water stain or a genuine soiled area.

[0016] Secondly, this application provides a photovoltaic module inspection system based on image recognition, the system comprising: The image information acquisition module is used to acquire image information of the same scene under at least two different light vibration modes; The interference information separation and suppression module is used to separate and suppress interference information with specific light vibration modes based on image information under different light vibration modes, so as to obtain an image without interference; An image processing module is used to perform local detail enhancement and non-random signal smoothing on the image after interference removal to highlight defect features and reduce electronic noise; The defect identification and evaluation module is used to identify defects on the surface of photovoltaic modules based on the image after local detail enhancement and non-random signal smoothing, and to evaluate the reliability of the defects based on the distinguishability between defect features and interference information.

[0017] Compared with related technologies, this application has at least the following technical effects: By leveraging image differences of the same scene under different light vibration modes, interference information such as specular reflection with specific light vibration modes can be separated. This effectively suppresses reflection interference while preserving true defect features, overcoming the limitation of traditional visible light image detection methods in distinguishing between reflection and defects. Through image patch similarity evaluation and weighted averaging, combining information from three dimensions—light vibration intensity, light vibration direction, and brightness—defect edge and texture features are enhanced while smoothing electronic noise. Furthermore, compensation for polarization distortion caused by ambient temperature and a correction mechanism for oil / water stain areas are introduced, improving the robustness and accuracy of defect detection in complex outdoor environments.

[0018] Details of one or more embodiments of this application are set forth in the following drawings and description to make other features, objects and advantages of this application more readily apparent. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating an image recognition-based photovoltaic module detection method according to an exemplary embodiment.

[0020] Figure 2 This is a flowchart illustrating step S2, which involves separating and suppressing interference information, according to an exemplary embodiment.

[0021] Figure 3 This is a flowchart illustrating the image enhancement and smoothing process in step S3 according to an exemplary embodiment.

[0022] Figure 4 This is a flowchart illustrating the image patch similarity evaluation in step S32 according to an exemplary embodiment.

[0023] Figure 5 This is a flowchart illustrating the weighted average processing in step S33 according to an exemplary embodiment.

[0024] Figure 6 This is a flowchart illustrating the pattern difference judgment and oil stain identification in step S332 according to an exemplary embodiment.

[0025] Figure 7 This is a block diagram illustrating an image recognition-based photovoltaic module inspection system according to an exemplary embodiment. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other. The application will now be described in detail with reference to the accompanying drawings and embodiments.

[0027] Example 1

[0028] Embodiment 1 of this application provides a photovoltaic module detection method based on image recognition. Figure 1 This is a flowchart illustrating an image recognition-based photovoltaic module detection method according to an exemplary embodiment. Figure 1 As shown, the method includes the following steps: S1, acquire image information of the same scene under at least two different light vibration modes; In this step, an image acquisition device equipped with a polarization imaging component simultaneously acquires multiple polarization images of the same photovoltaic module scene at the same shooting position and time, with at least two different polarization angles (e.g., 0°, 45°, 90°, and 135°). The polarization imaging component can be a focal plane polarization camera, which integrates an array of micro-polarizers in different directions on a single image sensor, enabling the simultaneous acquisition of images under multiple light vibration modes in a single exposure. The acquired image information includes three dimensions: brightness (total light intensity), light vibration intensity (degree of polarization), and light vibration direction (polarization angle) for each pixel.

[0029] S2, based on the image information under the different light vibration modes, separate and suppress interference information with specific light vibration modes to obtain an image without interference; In this step, the differences in light vibration modes contained in multiple images with different polarization angles obtained in step S1 are used to identify and separate interference information with specific light vibration mode characteristics. Since the specular reflection light from the photovoltaic module surface exhibits high linear polarization, its vibration direction is perpendicular to the incident plane, while the diffuse reflection light generated by defects on the module surface shows lower polarization or depolarization characteristics, there is a fundamental difference in their light vibration modes. By performing differential operations on images at different polarization angles and combining this with light vibration intensity threshold judgment, the specular reflection component is separated from the image and suppressed, while the diffuse reflection component is retained as effective image information after interference removal. For interference objects such as dust and water stains on the module surface that only change local brightness and texture but do not significantly alter the light vibration mode characteristics, the texture gradient and light vibration mode change rate of the local area are analyzed for auxiliary judgment to avoid erroneous removal.

[0030] S3, perform local detail enhancement and non-random signal smoothing on the image after removing interference to highlight defect features and reduce electronic noise; In this step, the image obtained in step S2 after removing reflection interference is enhanced and smoothed. Surface defects of photovoltaic modules (such as microcracks and hot spots) typically appear in images as subtle brightness variations and edge discontinuities, while image sensors generate thermal noise in high-temperature outdoor environments. By combining information from three dimensions—light vibration intensity, light vibration direction, and brightness—the local region where a pixel is located in the image is analyzed: a weighted average is performed within the set of neighboring pixels with similar light vibration patterns to the current pixel. Redundant information from pixels with similar patterns is used to smooth the non-random signal; that is, smoothing is only performed on pixels with the same physical source (such as effective reflections from the same surface region), avoiding detail blurring caused by crossing physical boundaries (such as defect edges). Simultaneously, the gradients of local brightness and light vibration intensity are enhanced to highlight the edge contrast of the defect area.

[0031] S4. Based on the image after local detail enhancement and non-random signal smoothing, identify defects on the surface of the photovoltaic module, and evaluate the reliability of the defects based on the distinguishability between defect features and interference information. In this step, defect identification and reliability assessment are performed on the image after enhancement and smoothing in step S3. Defect identification employs an edge detection and region segmentation-based method, extracting connected regions in the image with significant brightness gradients and light vibration patterns that differ from the surrounding normal regions as candidate defect regions. Feature vectors are extracted for each candidate defect region, including shape features (area, aspect ratio, roundness), brightness statistical features (mean, variance, contrast with background), and light vibration pattern features (mean and variance of polarization degree within the region, consistency of polarization angle). The feature vectors of the candidate defect regions are compared with the feature ranges of various typical defects (hidden cracks, hot spots, snail trails, junction box aging, etc.) in a pre-established defect feature library to determine the defect type. Simultaneously, based on the defect features and steps... The ability to distinguish between separated interference information features is used to assess the reliability of defect detection results: the higher the distinguishability, the higher the reliability.

[0032] In the technical solution of the above embodiments, image information under multiple different light vibration modes is acquired at the same shooting position and time. The essential difference between specular and diffuse light in their light vibration modes is utilized to effectively separate and suppress reflection interference. Non-random signal smoothing is performed by combining local region analysis with light vibration mode information to reduce noise while preserving defect edge details. Finally, the reliability of defect detection is evaluated based on the distinguishability between defect features and interference information. This method overcomes the limitation of traditional visible light detection methods, which struggle to distinguish between reflection interference and real defects under single acquisition equipment conditions, and achieves rapid and accurate detection of photovoltaic module defects under outdoor natural lighting conditions.

[0033] In one example, during a drone inspection of a photovoltaic power plant, the drone's polarization camera simultaneously captured images at four polarization directions (0°, 45°, 90°, and 135°) in a single flight. Under the oblique sunlight of early morning, microscopic scratches on the camera lens surface caused significant structured glare in the images. Simultaneously, to suppress glare, the camera automatically increased its gain, resulting in images filled with high-intensity electronic noise. Traditional visible light image detection methods are almost incapable of distinguishing between minute cracks and glare artifacts on the component surface under these conditions. Using the method described in this application, the system first acquires image data in four polarization directions using the polarization camera, providing a multidimensional optical basis for subsequent interference separation.

[0034] In one possible design, Figure 2 This is a flowchart illustrating step S2, which involves separating and suppressing interference information, according to an exemplary embodiment. (Refer to...) Figure 2 In step S2, the step of separating and suppressing interference information with specific light vibration modes based on image information under different light vibration modes to obtain an image with interference removed includes: S21, acquire image information of the same scene under at least two different light vibration modes; In this step, raw image data for multiple light vibration modes (e.g., four polarization directions: 0°, 45°, 90°, and 135°) are extracted from a single exposure of a split-plane polarization camera. Each image under a light vibration mode records the light intensity response of the photovoltaic module surface in the corresponding polarization direction, denoted as follows: , , and For each pixel location, the Stokes parameter is calculated using the following formula: Total light intensity Brightness is calculated using the following formula: polarization degree Calculate using the following formula: polarization angle Calculate using the following formula: in, The total light intensity parameter is... The difference in light intensity between the 0° and 90° polarization directions. The difference in light intensity between the 45° and 135° polarization directions, along with these three factors, constitute a complete mathematical description of the polarization optical characteristics of this pixel. Degree of polarization This reflects the proportion of the linearly polarized component of the light at that pixel to the total light intensity, with a value ranging from 0 to 1; the polarization angle... It reflects the vibration direction of the linearly polarized component, with a value range of 0° to 180°.

[0035] S22, calculate the rate of change of light vibration modes in the local region of the image; In this step, a local region of M×M (M is, for example, 5 to 7 pixels) is selected centered on each pixel, and the spatial gradient of the light vibration direction of each pixel within this local region is calculated. The rate of change of the light vibration mode is defined as the average gradient magnitude of the polarization angle of each pixel within the local region. The polarization angle gradient is calculated by performing a Sobel operator convolution on the polarization angle distribution map to obtain the gradient components in the horizontal and vertical directions, and then calculating the square root of the sum of the squares of the two as the gradient magnitude. Regions with high rates of change typically correspond to locations where the optical properties of the material surface change abruptly (such as defect edges), while regions with low rates of change correspond to normal areas with uniform surfaces or specular reflection areas.

[0036] S23, determine whether the rate of change of the light vibration mode in the local area exceeds a preset threshold; In this step, the rate of change of the local light vibration mode calculated in step S22 is compared with a preset rate of change threshold. The threshold is set based on a prior statistical analysis of the normal surface area and the known defect area of ​​the photovoltaic module: the polarization angle gradient of the normal surface area is usually distributed in a low range, while the polarization angle gradient of the defect edge area is significantly higher. When the average polarization angle gradient of the local area exceeds this statistical threshold, it is determined that there is a sudden change in the optical properties of the material in that area.

[0037] S24, when the rate of change of the light vibration mode in the local area exceeds the preset threshold, the parameters for separating the interference information are dynamically adjusted. In this step, the specular reflection separation parameters are dynamically adjusted based on the location and extent of the detected high-rate-of-change regions. The estimation of the specular reflection component is based on polarization information: specular reflection regions have high polarization. Let the specular reflection separation threshold for a pixel be... Calculate the estimated values ​​of the specular reflection components using the formulas respectively: when hour, when hour,

[0038] in, The specular reflection separation threshold is determined by statistical analysis of the polarization degree distribution in the normal surface area of ​​the photovoltaic module. This is an estimated value for the specular reflection component, representing the intensity of the light intensity identified as the specular reflection component in that pixel. When the pixel's polarization degree... Exceeding the threshold When the polarization degree is within a certain threshold, it indicates that the light from that pixel is dominated by specular reflection, and its brightness value is separated proportionally; when the polarization degree does not exceed the threshold, it indicates that the pixel is dominated by diffuse reflection, and the estimated value of the specular reflection component is 0. For regions with high change rates, the threshold is set... Increase the threshold to reduce the area identified as specular reflection and retain more polarization signal components; for uniform areas with low change rates, maintain the original threshold to remove specular reflection components normally.

[0039] S25, combining local texture and brightness gradient information, separates and suppresses interference information with specific light vibration patterns to obtain an image without interference; In this step, the specular reflection component is calculated and separated for each pixel in each ray vibration mode image, based on the dynamic separation parameters, local texture information, and brightness gradient information determined in step S24. For each pixel, the estimated proportion of the specular reflection component to the total light intensity is calculated according to its Stokes parameters, and this proportion is combined with the dynamic separation parameters to determine the suppression coefficient. The brightness value of each pixel in the original image is multiplied by the suppression coefficient to obtain the brightness value after removing specular reflection interference. For pixels with minimal differences in ray vibration modes (polarization degree and polarization angle are basically the same) but significant differences in brightness between different ray vibration mode images, they are identified as dust or water stain occlusion areas rather than specular reflection areas. Brightness compensation is performed on these pixels without reflection suppression, generating the final de-interference image.

[0040] In the technical solution of the above embodiments, the spatial change rate of light vibration modes is quantified by calculating the polarization angle gradient of the local region. A change rate threshold is used to distinguish between defect edge regions and uniform surface regions. The reflection separation threshold is dynamically lowered in defect edge regions to prevent over-suppression, while reflection suppression is strengthened in regions with large brightness gradients but small polarization angle gradients. Finally, the precise separation of reflection interference is achieved by combining Stokes parameter specular reflection estimation and dynamic separation coefficients. This method solves the problem that defect edge regions are easily misclassified as specular reflection regions under fixed separation parameters, ensuring the integrity of defect features.

[0041] In one example, during the aforementioned drone inspection scenario, when the polarization camera acquired images with a 0° polarization direction, glare artifacts caused by lens scratches covered key areas of the component surface. When acquiring images with a 90° polarization direction, the glare artifacts were significantly suppressed because the polarization direction was perpendicular to the glare. The system calculated the pixel-level differences between the two polarization direction images, identified pixels with differences exceeding a threshold as specular reflection interference areas, and replaced the brightness values ​​of these pixels with the corresponding pixel values ​​from the 90° polarization direction image. This effectively removed glare interference while preserving the true texture information of the component surface.

[0042] In one possible design, Figure 3 This is a flowchart illustrating the image enhancement and smoothing process in step S3 according to an exemplary embodiment. (Refer to...) Figure 3In step S3, the step of performing local detail enhancement and non-random signal smoothing processing on the image after removing interference to highlight defect features and reduce electronic noise includes: S31, acquire the brightness information, light vibration intensity information, and light vibration direction information of the image after interference removal; In this step, the brightness value of each pixel is extracted from the de-interference image generated in step S2, and the polarization degree (light vibration intensity) and polarization angle (light vibration direction) corresponding to each pixel are calculated from the multiple polarization images obtained in step S1. A three-dimensional feature vector (brightness, polarization degree, polarization angle) is constructed as the basic data for subsequent image block similarity evaluation and weighted average processing.

[0043] S32, evaluate the similarity between two image patches at different locations in the image after removing interference, wherein the similarity evaluation is based on the brightness information, the light vibration intensity information, and the light vibration direction information; In this step, two image patches at different locations are used as centers to extract their respective three-dimensional feature vector sets. The similarity between the two patches is comprehensively evaluated by comparing the feature differences in three dimensions: brightness distribution statistics, polarization degree distribution statistics, and polarization angle distribution statistics. The similarity is calculated using the following formula: in, For brightness distribution similarity, For the similarity of polarization degree distribution, For the similarity of polarization angle distribution, α, β, and γ are three-dimensional weighting coefficients that satisfy: Three-dimensional similarity assessments are more accurate than traditional methods that rely solely on brightness in distinguishing between pixels from the same physical surface and pixels that cross physical boundaries.

[0044] S33, Based on the image patch similarity, perform a weighted average on the image after removing interference to obtain an image with higher purity and clearer defect features; In this step, each pixel in the entire image is used as the center pixel to be processed. A search window is set around it, and the similarity between the image patch containing each pixel within the window and the image patch containing the center pixel is calculated. Using similarity as the basis for weighting, the brightness values ​​of all pixels within the window are weighted and averaged. The weighted average result is used as the new brightness value of the center pixel. Pixels with higher similarity receive greater weights, while pixels with extremely low similarity (possibly from different physical surfaces or defect areas) receive weights close to zero. In this way, only valid signals with the same physical origin as the current pixel are smoothly superimposed, electronic noise is suppressed due to statistical independence, and the edge and texture details of defect areas are preserved or even enhanced due to weight selectivity.

[0045] In the technical solution of the above embodiments, the similarity of image blocks is evaluated by using three dimensions of information, namely brightness, degree of polarization and polarization angle, which is more accurate than the traditional similarity evaluation that only relies on brightness to determine whether pixels belong to the same physical surface area. Based on the three-dimensional similarity, a selective weighted average is performed, which effectively smooths electronic noise while avoiding signal mixing across physical boundaries, ensuring that defect edges and texture details are not blurred.

[0046] In one example, in the aforementioned inspection scenario, the image after glare removal still exhibits electronic noise due to high gain. The system performs CLAHE local contrast enhancement on this image, dividing it into 16×16 pixel blocks for histogram equalization and limiting contrast amplification to prevent excessive noise amplification. Subsequently, a nonlocal mean denoising algorithm is applied, searching for the most similar image block in the entire image for each 5×5 image block and performing a weighted average. This effectively eliminates electronic noise while preserving the sharpness of defect edges.

[0047] In one possible design, Figure 4 This is a flowchart illustrating the image patch similarity evaluation in step S32 according to an exemplary embodiment. (Refer to...) Figure 4 In step S32, the step of evaluating the similarity between two image patches at different locations in the image after removing interference includes: S321, perform local region division of light vibration pattern for each image block in the image with the interference removed, wherein the local region division is based on the light vibration intensity and light vibration direction of the pixels inside the image block; In this step, a K×K image block is taken centered on each pixel. All pixels within the image block are clustered based on their polarization degree and polarization angle. The clustering distance considers both the relative difference in polarization degree and the cosine distance of the polarization angle, grouping pixels with similar polarization characteristics into the same local region. The clustering result divides the image block into several sub-regions, where pixels within each sub-region have essentially consistent physical surface characteristics (such as material, orientation, and roughness). This region division based on light vibration patterns reflects the physical structure of the material surface better than spatial division based solely on brightness.

[0048] S322, calculate the statistical distribution characteristics of brightness information, light vibration intensity information and light vibration direction information of each local region; In this step, for each sub-region defined in step S321, the mean brightness, variance brightness, histogram distribution of all pixels within it are calculated, as well as the mean and variance of polarization degree and polarization angle. These statistical distribution characteristics reflect the overall optical and surface geometric properties of the sub-region. The mean and variance brightness reflect the albedo and surface roughness of the sub-region, the mean and variance polarization degree reflect the uniformity of scattering within the sub-region, and the mean and variance polarization angle reflect the consistency of surface normals within the sub-region.

[0049] S323, compare the statistical distribution characteristics of corresponding local regions in the image blocks at the two different locations; In this step, sub-region matching is performed between two image patches at different locations to find the pair of sub-regions with the closest polarization characteristics between the two image patches. For each pair of matched sub-regions, the differences in brightness distribution (measured by histogram intersection distance or Bach distance), polarization degree distribution, and polarization angle distribution are calculated. The difference values ​​for each dimension are normalized to between 0 and 1.

[0050] S324, calculate the similarity between the two image blocks at different locations based on the differences in statistical distribution characteristics between the local regions and the area proportion of the local regions in the image block; In this step, the statistical distribution differences of each pair of matching sub-regions are weighted and summed. The weight is the area ratio of the sub-region in the image patch; the larger the area of ​​the sub-region, the greater its contribution to the overall similarity. The weighting coefficients of the three dimensions of difference can be adjusted according to the actual detection scenario: in scenarios sensitive to reflection interference, the weight of polarization degree and polarization angle differences is increased; in scenarios with high requirements for brightness contrast, the weight of brightness differences is increased. The final similarity value is between 0 and 1: 1 indicates that the two image patches come from the same physical surface region, and 0 indicates completely different surface regions.

[0051] In the technical solution of the above embodiments, image patch similarity is calculated by dividing sub-regions into sub-regions based on polarization characteristics, comparing the statistical distribution features of each sub-region, and then using area proportion as a weight. This method achieves fine differentiation of different physical surface regions, avoiding the bias caused by mixing regions of different materials in a globally unified similarity measurement.

[0052] In one possible design, Figure 5 This is a flowchart illustrating the weighted average processing in step S33 according to an exemplary embodiment. (Refer to...) Figure 5 In step S33, the step of performing a weighted average on the image after removing interference based on the image patch similarity includes: S331, Obtain the brightness information, light vibration intensity information, and light vibration direction information of the image after interference removal; In this step, the brightness value of each pixel in the interference-free image output in step S2, as well as the corresponding polarization degree and polarization angle data, are read as the basic input for weighted average processing.

[0053] S332, Identify a set of pixels in the local region surrounding each pixel to be processed in the image to be removed from interference that have a light vibration pattern that is very similar to that of the pixel to be processed; In this step, a search window (e.g., 21×21 pixels) is set around each pixel in the image to be processed. Each pixel within the window is traversed, and the difference in light vibration mode between that pixel and the central pixel to be processed is evaluated. The judgment criteria include: whether the difference in polarization degree between the two is less than a first threshold (reflecting the consistency of surface scattering characteristics) and whether the cosine distance of their polarization angles is less than a second threshold (reflecting the consistency of surface normal directions). Pixels that simultaneously meet both conditions are grouped into a set of pixels with minimal differences. This set represents effective signal pixels that share the same physical origin as the pixel to be processed.

[0054] S333, Assign initial weights based on the difference in brightness between the pixels in the pixel set and the pixel to be processed; In this step, for each pixel in the pixel set determined in step S332, the absolute difference between its brightness value and the brightness value of the pixel to be processed is calculated. The brightness difference is then processed using a Gaussian kernel function. The initial weights are mapped and calculated using the following formula: in, For surrounding pixels The absolute difference in brightness between the pixel and the pixel to be processed. The bandwidth parameter of the Gaussian kernel. This represents an exponential function with base e (the natural constant). (Brightness difference) Smaller pixels receive higher initial weights and brightness differences. Larger pixels receive lower initial weights. (Bandwidth parameter) The settings are based on the actual image noise level and the desired smoothness: the greater the noise... The larger the value, the stronger the smoothing effect.

[0055] S334, Based on the slight differences in light vibration intensity and light vibration direction between the pixels in the pixel set and the pixel to be processed, the initial weights are adjusted; In this step, the initial weights in step S333 are finely adjusted by utilizing the subtle differences in polarization degree and polarization angle between each pixel in the pixel set and the pixel to be processed. Even if two pixels should have been grouped into sets with minimal differences, there are still slight differences in their polarization characteristics. By introducing an adjustment coefficient proportional to the difference in polarization degree and polarization angle, the initial weights are fine-tuned. Pixels with closer polarization characteristics receive a slight increase in weight, while pixels with relatively large differences in polarization characteristics receive a slight decrease in weight. The adjusted weights better reflect whether two pixels truly belong to completely identical physical surfaces.

[0056] S335, using the adjusted weights, perform a weighted average of the brightness values ​​of the pixel to be processed and its surrounding pixels; In this step, the adjusted weights obtained in step S334 are used as weighting coefficients to calculate a weighted average of the brightness values ​​of all pixels in the pixel set. The weighted average brightness value is then used as the new brightness value for the pixel to be processed. This process is repeated to traverse all pixels in the image, resulting in an output image that has undergone non-random signal smoothing. This output image, while eliminating electronic noise, can preserve or even enhance the edge and texture features of the defective region.

[0057] In the technical solution of the above embodiments, a pixel similarity set is established by dual threshold screening based on polarization degree difference and cosine distance of polarization angle. A Gaussian kernel is used to map brightness differences as initial weights, and then the weights are finely adjusted based on minute differences in polarization intensity and direction. Finally, a weighted average is performed within the set of similar pixels. This method uses light vibration mode information as a constraint to guide the selection of the weighting range, achieving non-random smoothing only for pixels from the same physical source.

[0058] In one example, in the aforementioned inspection scenario, the system processes a pixel located near the edge of a hidden crack defect. Within a 21×21 pixel search window around the pixel, the system identifies 12 similar pixels whose polarization degree difference from the central pixel is less than a first threshold and whose polarization angle difference is less than a second threshold. These pixels form a non-random, smoothed, weighted calculation set. For pixels outside the set with significant polarization angle differences, the system resets their weights to zero, avoiding the mixing of brightness information across the defect edge, ensuring that the hidden crack edge remains clearly distinguishable after processing.

[0059] In one possible design, step S332, the step of identifying a set of pixels in a local region surrounding each pixel to be processed in the image to be removed, that have a light vibration pattern that is extremely similar to that of the pixel to be processed, includes: S3321, Obtain image information of the local area surrounding the pixel to be processed, and perform deblurring on the image information of the local area to obtain the deblurred image information; In this step, multi-dimensional image data, including brightness, polarization degree, and polarization angle maps, is extracted from a local region window centered on the pixel to be processed. Deblurring is then performed on this local region using Wiener filtering or blind deconvolution algorithms. Based on the estimated point spread function, the image is deconvolved to restore its sharpness and high-frequency details, reducing image blur caused by camera lens aberrations, slight defocusing, or atmospheric disturbances. This deblurring process improves the accuracy of subsequent polarization characteristic difference calculations.

[0060] S3322, Calculate the difference in light vibration intensity between the pixel to be processed and the surrounding pixels in the deblurred image information; In this step, the absolute difference between the polarization degree value of the pixel to be processed and the polarization degree value of each surrounding pixel within the window is calculated for the deblurred polarization degree map. The polarization degree value range is 0-1, and the difference directly reflects the degree of difference in the surface scattering characteristics of the two pixels.

[0061] S3323, Calculate the angle between the light vibration direction of the pixel to be processed and the surrounding pixels in the deblurred image information; In this step, the polarization angle of the pixel to be processed is calculated relative to the polarization angle of each surrounding pixel within the window, based on the deblurred polarization angle map. The polarization angle is a periodic variable ranging from 0° to 180°. Therefore, the angle is calculated using a cyclic difference method, taking the absolute value of the difference. If the difference is greater than 90°, it is subtracted from 180° to ensure the angle remains between 0° and 90°. A smaller angle indicates that the normal directions of the surfaces of the two pixels are closer.

[0062] S3324, compare the light vibration intensity difference with a first preset threshold, and compare the light vibration direction angle with a second preset threshold; In this step, the light vibration intensity difference calculated in step S3322 is compared with a first preset threshold: if the intensity difference is less than the first threshold, it indicates that the two pixels come from surface regions with similar scattering characteristics. The angle of light vibration direction calculated in step S3323 is compared with a second preset threshold: if the angle is less than the second threshold, it indicates that the two pixels come from surface regions with the same normal direction.

[0063] S3325, when the difference in light vibration intensity is less than the first preset threshold and the angle between the light vibration directions is less than the second preset threshold, the surrounding pixels are identified as pixels whose light vibration patterns are very similar to those of the pixel to be processed. In this step, the surrounding pixels are only included in the set of pixels with minimal differences if both comparisons in step S3324 are satisfied simultaneously. The two conditions constrain the set from two physical dimensions: surface scattering similarity and surface geometric orientation consistency. This ensures that the pixels in the set do indeed have the same physical origin as the pixels to be processed, and are not mistakenly included simply because of coincidental similarity in brightness.

[0064] In the technical solution of the above embodiments, the sharpness of the image is improved by deblurring preprocessing, and then the polarization intensity and direction difference are finely calculated. The pixel similarity judgment is constrained by two independent physical dimensions: intensity difference and direction angle. This ensures that the pixels included in the similarity set have true physical homology.

[0065] In one possible design, step S334, the step of adjusting the initial weights based on the minute differences in light vibration intensity and light vibration direction between the pixels in the pixel set and the pixel to be processed, includes: S3341, Obtain the current ambient temperature information of the image acquisition device; In this step, the current ambient temperature of the device is read by a temperature sensor mounted on the polarization camera housing. The core optical element of the polarization camera is a micro-polarizer array. Its substrate material (usually a polymer or liquid crystal material) undergoes slight thermal expansion or contraction when the temperature changes, causing a slight shift in the transmission axis direction of the micro-polarizer, thereby introducing polarization information distortion into the acquired polarization image.

[0066] S3342, Based on the current ambient temperature information, query the pre-established correspondence between polarization filter material deformation and light vibration mode information distortion to obtain polarization information distortion compensation parameters; In this step, the system internally stores a temperature-distortion correspondence table or function model pre-established through experimental calibration. This model describes the systematic offset of the polarization degree and polarization angle data acquired by pixels in each polarization direction of the polarization camera micro-polarizer array relative to the true values ​​under different ambient temperatures, specifically represented by a two-dimensional compensation parameter: This is the polarization compensation value. This is the polarization angle compensation value. Based on the current ambient temperature T obtained in step S3341, the corresponding compensation parameters are obtained by looking up a table or substituting them into the function model. The compensation parameters are applied to polarization difference correction as follows: Corrected polarization degree difference Calculate using the following formula: Correction of polarization angle difference Calculate using the following formula: in, and These are the original measured differences in polarization degree and polarization angle, respectively. and This is the compensation amount obtained by looking up a table based on the current temperature T.

[0067] S3343, Based on the slight differences in light vibration intensity and light vibration direction between the pixels in the pixel set and the pixels to be processed, and in conjunction with the polarization information distortion compensation parameters, the initial weights are adjusted; In this step, the polarization degree difference and polarization angle difference calculated in step S332 are first corrected by adding the compensation parameters from step S3342: Corrected polarization degree difference = Original polarization degree difference + Corrected polarization angle difference = Original polarization angle + The corrected difference more accurately reflects the true physical polarization difference between the two pixels, eliminating systematic measurement errors introduced by temperature-induced polarization filter deformation. Using the corrected difference guides the weight adjustment in step S334, ensuring that weight fine-tuning based on polarization difference remains accurate and effective even under extreme ambient temperature conditions such as high and low temperatures.

[0068] In the technical solution of the above embodiments, polarization information distortion compensation parameters are obtained by real-time acquisition of the device's ambient temperature and querying a pre-calibrated temperature-distortion correspondence. These compensation parameters are then used to correct the polarization degree difference and polarization angle difference between pixels before being applied to weight fine-tuning. This method solves the problem of systematic polarization measurement errors introduced by the micro-polarizer array of the polarization camera when the temperature changes, ensuring the accuracy and consistency of weight adjustment under different ambient temperature conditions.

[0069] In one possible design, Figure 6 This is a flowchart illustrating step S332, specifically the pattern difference determination and oil stain identification, according to an exemplary embodiment. (Refer to...) Figure 6 In step S332, the step of identifying a set of pixels in a local region surrounding each pixel to be processed in the image after interference removal that have a light vibration pattern that is extremely similar to that of the pixel to be processed includes: S332A, Deblurring is performed on the image information of the local area surrounding the pixel to be processed to obtain the deblurred image information; In this step, the brightness map, polarization degree map, and polarization angle map of the local area are deblurred, and the processing method is the same as in step S3321.

[0070] S332B, Identify the local light scattering and refraction areas caused by oil or water stains in the deblurred image information; In this step, local regions of the deblurred image are analyzed to detect oil or water stains. Oil and water stain areas have unique optical characteristics: their brightness distribution exhibits an irregular patchy pattern, and on the polarization degree map, the polarization degree of this region is significantly lower than that of the surrounding clean area (because the oil or water stain layer alters the surface scattering properties). On the polarization angle map, the polarization angle distribution of this region shows a randomized trend or irregular fluctuations (because the irregular surface of the oil or water stain causes multiple refractions and scattering of light). By comparing with preset oil or water stain optical feature templates, potentially existing oil or water stain areas in the local region are identified and marked.

[0071] S332C, based on the identification results of the local light scattering and refraction regions, performs local correction on the light vibration pattern information of the local regions in the deblurred image information; In this step, the polarization degree and polarization angle data of each pixel within the oil or water stain area identified in step S332B are locally corrected. Because the oil or water stain layer alters the vibration mode of the reflected light that should directly reach the camera, the light undergoes additional scattering and refraction as it passes through the oil or water stain layer, and its polarization characteristics no longer reflect the true information of the component surface. The correction strategy is to utilize the effective polarization degree and polarization angle data of the clean area surrounding the oil or water stain area, and fill and replace the polarization degree and polarization angle values ​​of the oil or water stain area using bilinear or bicubic interpolation. The corrected polarization information reflects the inferred true polarization characteristics of the component surface in that area.

[0072] S332D, calculate the difference in light vibration intensity and the angle between the light vibration direction of the pixel to be processed and the surrounding pixels in the locally corrected light vibration mode information; In this step, the difference is calculated using the polarization degree and polarization angle data corrected in step S332C (using the same method as steps S3322 and S3323), but the input data has eliminated the distortion of the light vibration mode in the oil or water stain areas.

[0073] S332E, when the difference in light vibration intensity is less than a first preset threshold and the angle between the light vibration directions is less than a second preset threshold, the surrounding pixels are identified as pixels whose light vibration patterns are very similar to those of the pixel to be processed. In this step, the same judgment logic as in step S3325 is used, but due to the use of corrected polarization data and more refined difference calculation, the judgment result excludes the interference of oil or water stain areas, improving the accuracy of pixel similarity sets in the presence of surface contamination.

[0074] In the technical solution of the above embodiments, the unique optical feature patterns of oil or water stains in local images are identified, and the polarization information distortion of the contaminated area is corrected by interpolation filling with the polarization data of the surrounding clean area. Difference judgment is then performed based on the corrected polarization data. This method solves the problem of polarization characteristic distortion caused by oil and water stains on the surface of photovoltaic modules, and avoids contaminated areas being incorrectly excluded from similar pixel sets or incorrectly included in real defect areas due to abnormal polarization information.

[0075] In one possible design, step S332B, the step of identifying the local light scattering and refraction areas caused by oil or water stains in the deblurred image information, includes: S332B1, Obtain the light vibration intensity distribution and light vibration direction distribution of local areas in the deblurred image information, as well as brightness texture features and edge gradient information; In this step, four types of features are extracted from the deblurred local region: statistical parameters of polarization degree distribution (mean and variance), statistical parameters of polarization angle distribution (mean, variance, and angular entropy), brightness texture features (four texture parameters based on the gray-level co-occurrence matrix: contrast, energy, homogeneity, and correlation), and edge gradient information (based on the mean and directional consistency of gradient magnitudes of the Sobel operator). These four types of features describe the optical and morphological properties of the local region from different dimensions.

[0076] S332B2, compare the light vibration intensity distribution and light vibration direction distribution of the local area with the preset light vibration mode characteristic range of oil or water stains to obtain a first comparison result; In this step, the polarization degree distribution statistical parameters and polarization angle distribution statistical parameters extracted in step S332B1 are compared with the pre-established polarization feature template for oil stains or water stains. This template records the statistical characteristic range of known oil stain and water stain regions in terms of polarization degree and polarization angle distribution. The polarization degree values ​​of oil stain or water stain regions are usually distributed in a lower range (because additional scattering destroys the original polarization signal), and the angular entropy of the polarization angle is significantly higher (because refraction and scattering randomize the polarization direction). The comparison results are output as a membership function: the closer the match to the polarization feature template of the oil stain or water stain, the higher the membership degree.

[0077] S332B3, compare the brightness texture features and edge gradient information of the local area with the preset brightness texture features and edge gradient information of the real dirt to obtain a second comparison result; In this step, the brightness texture features and edge gradient information extracted in step S332B1 are compared with the pre-established brightness texture and edge gradient feature templates of real contamination (such as hidden cracks and hot spots). Real contamination areas typically have clear edge gradients, edge distributions with high directional consistency, and specific texture patterns (such as the linear texture of hidden cracks and the blocky texture of hot spots). The comparison results are output as a membership function: the closer the feature template matches the real contamination, the higher the membership degree.

[0078] S332B4, Based on the first comparison result and the second comparison result, determine whether the local area is an oil stain or water stain area, or a real soiled area; In this step, a comprehensive judgment is made based on the outputs of the two membership functions in steps S332B2 and S332B3. When the membership degree of the oil stain or water stain in the first comparison result is higher than the preset first judgment threshold, and the membership degree of the actual contamination in the second comparison result is lower than the preset second judgment threshold, the local area is judged to be an oil stain or water stain area, meaning that the visual anomaly in this area is caused by surface contaminants, not a real defect in the component itself. Conversely, when the first membership degree is low and the second membership degree is high, it is judged to be a real contamination area. When both membership degrees are in the middle range, the area is marked as an uncertain area, which may be a mixture of slight oil stains and real defects, or it may be other types of surface anomalies.

[0079] In the technical solution of the above embodiments, dual feature comparison of polarization characteristic distribution and brightness texture edge gradient is used. Membership functions are used to quantify the degree of matching with oil or water stain templates and real contamination templates, respectively. Based on a comprehensive judgment of the two membership degrees, surface contaminants and real defects are distinguished. This method solves the problem that oil stains and water stains on the surface of photovoltaic modules are easily confused with real defects in ordinary visible light images.

[0080] In summary, the photovoltaic module detection method based on image recognition provided in Embodiment 1 of this application acquires image information under multiple different light vibration modes at the same shooting position, separates specular reflection interference from the physical level by utilizing the difference in polarization characteristics, performs non-random signal smoothing by combining local region analysis of polarization information and three-dimensional similarity evaluation, eliminates noise while preserving defect edges, and introduces environmental temperature compensation and oil and water stain recognition mechanisms to improve robustness in complex outdoor environments. Ultimately, it achieves rapid and accurate detection and reliability assessment of photovoltaic module surface defects under the condition of a single acquisition device.

[0081] Example 2 Embodiment 2 of this application provides a photovoltaic module detection system based on image recognition. Figure 7 This is a block diagram illustrating an image recognition-based photovoltaic module inspection system according to an exemplary embodiment. Figure 7 As shown, the system includes: Image information acquisition module 01 is used to acquire image information of the same scene under at least two different light vibration modes.

[0082] The interference information separation and suppression module 02 is used to separate and suppress interference information with specific light vibration modes based on image information under different light vibration modes, so as to obtain an image without interference.

[0083] Image processing module 03 is used to perform local detail enhancement and non-random signal smoothing on the image after removing interference, so as to highlight defect features and reduce electronic noise.

[0084] The defect identification and evaluation module 04 is used to identify defects on the surface of the photovoltaic module based on the image after local detail enhancement and non-random signal smoothing, and to evaluate the reliability of the defects based on the distinguishability between defect features and interference information.

[0085] In summary, the photovoltaic module inspection system based on image recognition provided in Embodiment 2 of this application effectively separates specular reflection interference under the same equipment conditions through the cooperation of the image information acquisition module and the interference information separation and suppression module. The defect features are enhanced by three-dimensional similarity weighted smoothing of the image processing module, and finally the defect identification and evaluation module completes the location, classification and reliability assessment of the defects.

[0086] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0087] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A photovoltaic module detection method based on image recognition, characterized in that, Includes the following steps: Acquire image information of the same scene under at least two different light vibration modes; Based on the image information under the different light vibration modes, interference information with specific light vibration modes is separated and suppressed to obtain an image without interference; The image after removing interference is subjected to local detail enhancement and non-random signal smoothing to highlight defect features and reduce electronic noise; Based on the image after local detail enhancement and non-random signal smoothing, defects on the surface of the photovoltaic module are identified, and the reliability of the defects is evaluated according to the distinguishability between defect features and interference information.

2. The method according to claim 1, characterized in that, The step of separating and suppressing interference information with specific light vibration modes based on image information under different light vibration modes to obtain an interference-free image includes: Acquire image information of the same scene under at least two different light vibration modes; Calculate the rate of change of light vibration modes in the local region of the image; Determine whether the rate of change of the light vibration mode in the local area exceeds a preset threshold; When the rate of change of the light vibration mode in the local area exceeds the preset threshold, the parameters for separating the interference information are dynamically adjusted. By combining local texture and brightness gradient information, interference information with specific light vibration patterns is separated and suppressed to obtain an image with interference removed.

3. The method according to claim 1, characterized in that, The step of performing local detail enhancement and non-random signal smoothing on the image after removing interference to highlight defect features and reduce electronic noise includes: Obtain the brightness information, light vibration intensity information, and light vibration direction information of the image after interference removal; The similarity between two image patches at different locations in the image after interference removal is evaluated, wherein the similarity evaluation is based on the brightness information, the light vibration intensity information, and the light vibration direction information; Based on the image patch similarity, the image after removing interference is weighted and averaged to obtain an image with higher purity and clearer defect features.

4. The method according to claim 3, characterized in that, The step of evaluating the similarity between two image patches at different locations in the noise-removed image includes: For each image block in the image with the interference removed, a local region segmentation of the light vibration pattern is performed, wherein the local region segmentation is based on the light vibration intensity and light vibration direction of the pixels within the image block; Calculate the statistical distribution characteristics of brightness, light vibration intensity, and light vibration direction information for each local region; Compare the statistical distribution characteristics of corresponding local regions in the two image blocks at different locations; The similarity between the two image blocks at different locations is calculated based on the differences in statistical distribution characteristics between the local regions and the area proportion of the local regions in the image block.

5. The method according to claim 3, characterized in that, The step of performing a weighted average on the image after removing interference based on the image patch similarity to obtain an image with higher purity and clearer defect features includes: Obtain the brightness information, light vibration intensity information, and light vibration direction information of the image after interference removal; Identify a set of pixels in a local region surrounding each pixel to be processed in the image to be removed from interference that have a light vibration pattern that is very similar to that of the pixel to be processed. Initial weights are assigned based on the difference in brightness between the pixels in the pixel set and the pixel to be processed; The initial weights are adjusted based on the slight differences in light vibration intensity and direction between the pixels in the pixel set and the pixels to be processed. Using the adjusted weights, a weighted average is calculated on the brightness values ​​of the pixel to be processed and its surrounding pixels.

6. The method according to claim 5, characterized in that, The step of identifying a set of pixels in a local region surrounding each pixel to be processed in the image after interference removal that have a light vibration pattern that is extremely similar to that of the pixel to be processed includes: The image information of the local area surrounding the pixel to be processed is obtained, and the image information of the local area is deblurred to obtain the deblurred image information; Calculate the difference in light vibration intensity between the pixel to be processed and the surrounding pixels in the deblurred image information; Calculate the angle between the light vibration direction of the pixel to be processed and the surrounding pixels in the deblurred image information; The difference in light vibration intensity is compared with a first preset threshold, and the angle of light vibration direction is compared with a second preset threshold. When the difference in light vibration intensity is less than the first preset threshold and the angle between the light vibration directions is less than the second preset threshold, the surrounding pixels are identified as pixels whose light vibration patterns are very similar to those of the pixel to be processed.

7. The method according to claim 5, characterized in that, The step of adjusting the initial weights based on the minute differences in light vibration intensity and direction between the pixels in the pixel set and the pixel to be processed includes: Obtain the current ambient temperature information of the image acquisition device; Based on the current ambient temperature information, the pre-established correspondence between polarization filter material deformation and light vibration mode information distortion is queried to obtain polarization information distortion compensation parameters; Based on the slight differences in light vibration intensity and direction between the pixels in the pixel set and the pixel to be processed, the initial weights are adjusted in conjunction with the polarization information distortion compensation parameters.

8. The method according to claim 5, characterized in that, The step of identifying a set of pixels in a local region surrounding each pixel to be processed in the image after interference removal that have a light vibration pattern that is extremely similar to that of the pixel to be processed includes: The image information of the local area surrounding the pixel to be processed is obtained, and the image information of the local area is deblurred to obtain the deblurred image information; Identify the local light scattering and refraction areas caused by oil or water stains in the deblurred image information; Based on the identification results of the local light scattering and refraction regions, the light vibration pattern information of the local regions in the deblurred image information is locally corrected. Calculate the difference in light vibration intensity and the angle between the light vibration direction of the pixel to be processed and the surrounding pixels in the locally corrected light vibration mode information; When the difference in light vibration intensity is less than a first preset threshold and the angle between the light vibration directions is less than a second preset threshold, the surrounding pixels are identified as pixels whose light vibration patterns are very similar to those of the pixel to be processed.

9. The method according to claim 8, characterized in that, The step of identifying local light scattering and refraction areas caused by oil or water stains in the deblurred image information includes: The light vibration intensity distribution and light vibration direction distribution of local regions in the deblurred image information, as well as brightness texture features and edge gradient information, are obtained. The light vibration intensity distribution and light vibration direction distribution of the local area are compared with the preset light vibration mode characteristic range of oil or water stains to obtain a first comparison result; The brightness texture features and edge gradient information of the local area are compared with the preset brightness texture features and edge gradient information of the actual dirt to obtain a second comparison result; Based on the first comparison result and the second comparison result, it is determined whether the local area is an oil or water stain area or a real soiled area.

10. A photovoltaic module inspection system based on image recognition, characterized in that, The system includes: The image information acquisition module is used to acquire image information of the same scene under at least two different light vibration modes; The interference information separation and suppression module is used to separate and suppress interference information with specific light vibration modes based on image information under different light vibration modes, so as to obtain an image without interference; An image processing module is used to perform local detail enhancement and non-random signal smoothing on the image after interference removal to highlight defect features and reduce electronic noise; The defect identification and evaluation module is used to identify defects on the surface of photovoltaic modules based on the image after local detail enhancement and non-random signal smoothing, and to evaluate the reliability of the defects based on the distinguishability between defect features and interference information.

Images