Photovoltaic panel visual inspection adaptive optimization method for harsh marine environment

Abstract

The application discloses a photovoltaic panel visual detection adaptive optimization method for offshore harsh environment and a related device, wherein the method comprises the following steps: collecting original visible light images, preset waveband near-infrared images, linear polarization degree images and polarization angle images of a target photovoltaic array, and acquiring light intensity, haze concentration index and relative humidity in real time; inputting a specular reflection area semantic segmentation mask, a fog-enhanced image and a wet reflection probability map into a multi-condition generative adversarial network, converting the original visible light image from a current harsh environment domain to a pseudo-standard light image domain to obtain a final enhanced image; inputting the final enhanced image and the specular reflection area semantic segmentation mask into an optimal detection model pipeline for defect detection, and outputting a preliminary defect detection result and identifying a reliability suspicious area. The application automatically selects an optimal detection model pipeline according to a real-time environment state vector, and improves the precision of offshore harsh environment photovoltaic panel visual detection.

Description

Technical Field

[0001] This invention relates to the fields of photovoltaic monitoring and computer vision technology, specifically to an adaptive optimization method and apparatus for visual inspection of photovoltaic panels in harsh marine environments, as well as a computing device. Background Technology

[0002] Offshore photovoltaic (PV) power plants are typically built in aquatic environments such as oceans and lakes, offering advantages such as not occupying land and excellent cooling. However, factors such as high humidity, high salt spray corrosion, strong specular reflection from the sea surface, variable weather (e.g., fog, haze, low light), and equipment swaying caused by wind and waves severely interfere with the reliability and accuracy of PV panel defect detection. For example, PV panels may develop various defects during operation, such as hot spots, microcracks, snail trails, corrosion, and dirt. If these defects are not detected and addressed promptly, they can lead to decreased power generation efficiency, safety hazards, and even fires.

[0003] Traditional manual inspection methods are inefficient and susceptible to subjective factors, making drone inspections equipped with visible light cameras the mainstream solution for onshore photovoltaic power plant inspections. However, in harsh marine environments, the strong specular reflection (glare) from the sea surface and the damp photovoltaic panel surface obscures the true texture and defect features, leading to false positives and false negatives. Fog, haze, and moisture at sea reduce image contrast, distort colors, and blur details. Furthermore, salt spray adheres to the photovoltaic panel surface, forming an uneven coating, while moisture causes dynamic changes in surface reflectivity, interfering with visual perception.

[0004] To address the aforementioned issues, this invention proposes an adaptive optimization method for visual inspection of photovoltaic panels in harsh marine environments, aiming to improve the accuracy of visual inspection of marine photovoltaic panels. Summary of the Invention

[0005] In view of the above problems, the present invention provides an adaptive optimization method and apparatus for visual inspection of photovoltaic panels in harsh marine environments, as well as a computing device.

[0006] According to one aspect of the present invention, an adaptive optimization method for visual inspection of photovoltaic panels in harsh marine environments is provided, comprising: The system synchronously acquires raw visible light images and preset band near-infrared images of the target photovoltaic array through a multi-sensor module installed on the inspection equipment; obtains linear polarization degree images and polarization angle images through a polarization imaging unit; and obtains light intensity, haze concentration index and relative humidity in real time through a micro weather station. The system then performs spatial registration and timestamp alignment of the raw visible light images, preset band near-infrared images, linear polarization degree images and polarization angle images to form an environmental state perception data set. The environmental state perception data set, light intensity, haze concentration index and relative humidity are input into the environmental meta-feature encoder, and the environmental state vector is output. The environmental state vector includes the quantized specular reflection intensity, atmospheric scattering coefficient, overall illuminance level and environmental corrosion factor of the current scene. The optimal detection model pipeline matching the current environmental conditions is dynamically selected based on the environmental state vector, and the preprocessing channel is activated based on the threshold values ​​of each component in the environmental state vector. The detection model pipeline includes a YOLO-World-R model optimized for high-reflection scenarios, a YOLO-World-H model optimized for low-illuminance haze scenarios, and a YOLO-World-C model optimized for high-humidity salt spray corrosion scenarios. When the specular reflection intensity component exceeds a preset threshold, the linear polarization degree image, polarization angle image, preset band near-infrared image, and original visible light image are input into the reflectance component analysis network, outputting a semantic segmentation mask for the specular reflection region. When the atmospheric scattering coefficient component exceeds a preset threshold, the original visible light image and haze concentration index are defogging to obtain a defogging enhanced image. The relative humidity is combined with the overall illuminance level component in the environmental state vector to generate a humid reflectivity probability map. The semantic segmentation mask of the specular reflection region, the dehazing and enhanced image and the wet reflection probability map are input into the multi-condition generative adversarial network to transform the original visible light image from the current harsh environment domain to the pseudo-standard illumination image domain to obtain the final enhanced image. The final enhanced image and the semantic segmentation mask of the specular reflection region are input into the optimal detection model pipeline for defect detection, and the output is a preliminary defect detection result with localization and classification information; the uncertainty of the preliminary defect detection result is evaluated by uncertainty quantification, and regions with questionable reliability are identified.

[0007] In one alternative embodiment, the inspection device is an autonomous cruise drone or a surface robot; the multi-sensor module consists of a global shutter RGB visible light camera, two near-infrared cameras equipped with different optical bandpass filters, and a polarization imaging unit; the wavelength distributions of the two near-infrared cameras are 850±10nm and 1200±20nm.

[0008] In one alternative, the polarization imaging unit employs a CMOS image sensor with an on-chip polarization filter array, wherein each superpixel unit of the on-chip polarization filter array consists of subpixels covered by micro-polarization filters in four directions: 0°, 45°, 90°, and 135°.

[0009] In one alternative approach, acquiring the linear polarization degree image and polarization angle image via the polarization imaging unit further includes: The original polarized Bayer image output by the polarization imaging unit is reconstructed into four independent polarization direction intensity images through interpolation of adjacent pixels; The first three Stokes parameters are calculated based on the four polarization direction intensity images, and the linear polarization degree image and polarization angle image are obtained by solving the first three Stokes parameters.

[0010] In one alternative approach, the YOLO-World-R model, the YOLO-World-H model, and the YOLO-World-C model are all based on the YOLO-World-v2 framework. The YOLO-World-v2 framework includes an image encoder, a text encoder, and a reparameterizable visual prediction head. The image encoder uses a CSPDarknet network to extract multi-scale visual features from the input image. The text encoder uses a pre-trained CLIP text encoder to encode the category names of photovoltaic panel defects into text feature vectors. The visual prediction head is used to locate and classify the fused visual features.

[0011] In one alternative approach, combining relative humidity with the overall illuminance level component in the environmental state vector to generate a damp reflectivity probability map further includes: Based on relative humidity, overall horizontal illuminance component, photovoltaic panel installation tilt angle and current solar altitude angle, the base reflectivity map is calculated using the Fresnel reflection model. The basic reflectivity map, linear polarization image, and environmental corrosion factor in the environmental state vector are input into a convolutional neural network, which outputs a wet reflectivity probability map in the range of [0,1].

[0012] In one alternative approach, performing an uncertainty quantification assessment on the preliminary defect detection results to identify areas of questionable reliability further includes: The preliminary results of statistical defect detection include the frequency of defect boxes, the variance of bounding box vertex coordinates, and the entropy value of the category prediction probability distribution. Calculate the local contrast of the defect box in the original visible light image, the polarization consistency in the linear polarization image, and the proportion of the overlap area with the semantic segmentation mask of the specular reflection region and the probability map of wet reflection. When the inspection equipment performs multi-angle or continuous frame observations on the same photovoltaic panel module, calculate the positional drift of the defect in the same spatial location in the detection results of 3 consecutive frames. Defect boxes with entropy values, overlap area ratios, and positional drift exceeding the first threshold are marked as highly suspicious regions; defect boxes between the first and second thresholds are marked as moderately suspicious regions; and those below the second threshold are considered reliable detection regions.

[0013] In an alternative approach, the method further includes: When any photovoltaic panel module is marked as a highly suspicious area or a medium-suspicious defect area, the flight altitude, shooting angle and sensor exposure parameters of the inspection equipment are controlled to re-acquire image data of that area; During the resampling process, the 1200±20nm near-infrared band image is prioritized to penetrate the salt spray surface; the two detection results are jointly updated with confidence, and if the area is considered a reliable detection area, a manual review warning is triggered.

[0014] According to another aspect of the present invention, an adaptive optimization device for visual inspection of photovoltaic panels for harsh marine environments is provided, comprising: The environmental perception data acquisition module is used to simultaneously acquire raw visible light images and preset band near-infrared images of the target photovoltaic array through a multi-sensor module installed on the inspection equipment; acquire linear polarization degree images and polarization angle images through a polarization imaging unit; and acquire light intensity, haze concentration index and relative humidity in real time through a micro weather station; and spatially register and timestamp align the raw visible light images, preset band near-infrared images, linear polarization degree images and polarization angle images to form an environmental state perception data set. The environmental meta-feature quantization module is used to input the environmental state perception data set, light intensity, haze concentration index and relative humidity into the environmental meta-feature encoder, and output an environmental state vector, wherein the environmental state vector includes quantified specular reflection intensity, atmospheric scattering coefficient, overall illuminance level and environmental corrosion factor of the current scene. An adaptive detection pipeline module is used to dynamically select the optimal detection model pipeline that matches the current environmental conditions based on the environmental state vector, and to activate the preprocessing channel based on the threshold values ​​of each component in the environmental state vector. The detection model pipeline includes a YOLO-World-R model optimized for high-reflection scenarios, a YOLO-World-H model optimized for low-illuminance haze scenarios, and a YOLO-World-C model optimized for high-humidity salt spray corrosion scenarios. When the specular reflection intensity component exceeds a preset threshold, the linear polarization degree image, polarization angle image, preset band near-infrared image, and original visible light image are input into the reflectance component analysis network to output a semantic segmentation mask for the specular reflection region. When the atmospheric scattering coefficient component exceeds a preset threshold, the original visible light image and haze concentration index are defogging to obtain a defogging enhanced image. The relative humidity is combined with the overall illuminance level component in the environmental state vector to generate a humid reflectivity probability map. The image domain transformation and enhancement module is used to input the semantic segmentation mask of the specular reflection region, the dehazing and enhancement image and the wet reflection probability map into the multi-condition generative adversarial network to transform the original visible light image from the current harsh environment domain to the pseudo-standard illumination image domain to obtain the final enhanced image; The defect detection and evaluation module is used to input the final enhanced image and the semantic segmentation mask of the specular reflection region into the optimal detection model pipeline for defect detection, and output preliminary defect detection results with localization and classification information; it performs uncertainty quantification evaluation on the preliminary defect detection results and identifies areas with questionable reliability.

[0015] According to another aspect of the present invention, a computing device is provided, comprising: a processor, a memory, a communication interface, and a communication bus, wherein the processor, the memory, and the communication interface communicate with each other via the communication bus; The memory is used to store at least one executable instruction, which causes the processor to perform the operation corresponding to the above-described adaptive optimization method for visual inspection of photovoltaic panels in harsh marine environments.

[0016] According to the solution provided by the present invention, a multi-sensor module installed on the inspection equipment synchronously acquires the original visible light image and the preset band near-infrared image of the target photovoltaic array; a polarization imaging unit acquires the linear polarization degree image and the polarization angle image; and a micro weather station acquires the light intensity, haze concentration index, and relative humidity in real time. The original visible light image, the preset band near-infrared image, the linear polarization degree image, and the polarization angle image are spatially registered and time-stamped to form an environmental state perception data set. The environmental state perception data set, light intensity, haze concentration index, and relative humidity are input to an environmental meta-feature encoder, which outputs an environmental state vector. The environmental state vector includes quantized specular reflection intensity, atmospheric scattering coefficient, overall illuminance level, and environmental corrosion factor of the current scene. The optimal detection model pipeline matching the current environmental conditions is dynamically selected based on the environmental state vector, and the preprocessing channel is activated based on the threshold of each component in the environmental state vector. The detection model pipeline includes a YOLO-World-R model optimized for high-reflection scenes and a YOLO-World-R model optimized for low-illuminance haze scenes. The OLO-World-H model and the YOLO-World-C model optimized for high-humidity salt spray corrosion scenarios are used. When the specular reflection intensity component exceeds a preset threshold, the linear polarization degree image, polarization angle image, preset band near-infrared image, and original visible light image are input into the reflection component analysis network to output a semantic segmentation mask for the specular reflection region. When the atmospheric scattering coefficient component exceeds a preset threshold, the original visible light image and haze concentration index are dehazed to obtain a dehazed enhanced image. The relative humidity is combined with the overall illuminance level component in the environmental state vector to generate a wet reflectivity probability map. The semantic segmentation mask for the specular reflection region, the dehazed enhanced image, and the wet reflectivity probability map are input into a multi-condition generative adversarial network to transform the original visible light image from the current harsh environment domain to the pseudo-standard illumination image domain to obtain the final enhanced image. The final enhanced image and the semantic segmentation mask for the specular reflection region are input into the optimal detection model pipeline for defect detection, and the preliminary defect detection results with localization and classification information are output. The uncertainty quantification assessment of the preliminary defect detection results is performed to identify areas with questionable reliability. This invention improves the accuracy of visual inspection of photovoltaic panels in harsh marine environments by automatically selecting the optimal detection model pipeline through real-time environmental state vectors.

[0017] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and in order to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0018] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 This invention illustrates a flowchart of an adaptive optimization method for visual inspection of photovoltaic panels in harsh marine environments, according to an embodiment of the present invention. Figure 1 ; Figure 2 A schematic diagram of the inspection equipment according to an embodiment of the present invention is shown; Figure 3 A schematic diagram illustrating the probability of wet reflection according to an embodiment of the present invention is shown; Figure 4 This invention illustrates a flowchart of an adaptive optimization method for visual inspection of photovoltaic panels in harsh marine environments, according to an embodiment of the present invention. Figure 2 ; Figure 5 A schematic diagram of the frame of the adaptive optimization device for visual inspection of photovoltaic panels for harsh marine environments, according to an embodiment of the present invention, is shown. Figure 6 A schematic diagram of the structure of a computing device according to an embodiment of the present invention is shown. Detailed Implementation

[0019] Exemplary embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the invention are shown in the drawings, it should be understood that the invention may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this invention will be thorough and complete, and will fully convey the scope of the invention to those skilled in the art.

[0020] Figure 1 , Figure 2 The flowcharts of the adaptive optimization method for visual inspection of photovoltaic panels in harsh marine environments according to embodiments of the present invention are shown respectively. Figure 1 , Two Specifically, such as Figure 1 , Figure 2 As shown, it includes the following steps: Step S101: The original visible light image and the preset band near-infrared image of the target photovoltaic array are simultaneously acquired by the multi-sensor module set on the inspection equipment; the linear polarization degree image and the polarization angle image are acquired by the polarization imaging unit; and the light intensity, haze concentration index and relative humidity are acquired in real time by the micro weather station; the original visible light image, the preset band near-infrared image, the linear polarization degree image and the polarization angle image are spatially registered and time-stamped to form an environmental state perception data set.

[0021] In this embodiment, the polarization imaging unit directly quantifies the polarization state of light waves by acquiring images of linear polarization degree and polarization angle, enabling a physical distinction between surface-reflected light (partially polarized) and object-reflected light (unpolarized / weakly polarized). Near-infrared bands (such as 850nm and 1200nm) can penetrate fog and salt spray, and polarization imaging can suppress strong specular reflections from water surfaces or photovoltaic panel surfaces. Combining these two technologies can significantly enhance target visibility in low-visibility, high-reflectivity scenarios.

[0022] In one alternative embodiment, the inspection device is an autonomous cruise drone or a surface robot; the multi-sensor module consists of a global shutter RGB visible light camera, two near-infrared cameras equipped with different optical bandpass filters, and a polarization imaging unit; the wavelength distributions of the two near-infrared cameras are 850±10nm and 1200±20nm.

[0023] In this embodiment, the UAV can conduct multi-angle inspections of large offshore photovoltaic arrays, including high-altitude overhead views and close-up side views, while the surface robot can perform close-up observations of the underwater support structure and surface connection points of the floating photovoltaic system. When the UAV is flying at high speed or the surface robot is moving, the global shutter ensures that all pixels of the image are exposed simultaneously, avoiding image distortion and stretching (jelly effect) caused by platform movement. This is particularly suitable for shooting high-contrast scenes in strong light environments (such as the boundary between the photovoltaic panel and the sea surface), avoiding brightness inconsistencies caused by exposure time differences. The 850nm band is located in the near-infrared short-wave region, where water molecule absorption is relatively weak, providing good penetration of thin fog and light water vapor while maintaining high sensor sensitivity and image resolution. The 1200nm band is located in the near-infrared long-wave region, exhibiting significantly better penetration of salt spray particles and moist water films than short-wave near-infrared and visible light, and can partially penetrate the salt crystal layer and moist film on the photovoltaic panel surface. Dual-band image comparison analysis can distinguish between surface deposits (salt stains, dust) and inherent defects (cracks, hot spots). For example, surface salt stains may appear faint or disappear in a 1200nm image, while real cracks will be visible in both wavelengths.

[0024] In one alternative, the polarization imaging unit employs a CMOS image sensor with an on-chip polarization filter array, wherein each superpixel unit of the on-chip polarization filter array consists of subpixels covered by micro-polarization filters in four directions: 0°, 45°, 90°, and 135°.

[0025] In this embodiment, a static filter array is used, eliminating the need for mechanical rotating parts, thus improving system stability and preventing corrosion of moving parts in the high humidity and salt spray environment at sea. Figure 2As shown, images in four polarization directions are acquired simultaneously in a single exposure, avoiding inconsistencies caused by time-division acquisition due to scene changes (such as cloud movement or water surface ripples). Intensity measurements in four independent directions can completely resolve the first three Stokes parameters, thereby accurately calculating the degree of linear polarization and polarization angle. Furthermore, four-point sampling provides redundant information compared to three-point sampling, improving the signal-to-noise ratio of polarization parameter calculation and enabling reliable polarization images to be obtained even in low-light or high-noise environments.

[0026] In one alternative approach, acquiring the linear polarization degree image and polarization angle image via the polarization imaging unit further includes: The original polarized Bayer image output by the polarization imaging unit is reconstructed into four independent polarization direction intensity images through interpolation of adjacent pixels; The first three Stokes parameters are calculated based on the four polarization direction intensity images, and the linear polarization degree image and polarization angle image are obtained by solving the first three Stokes parameters.

[0027] In this embodiment, adjacent pixel interpolation is used to reconstruct four independent polarization direction intensity images from the original polarization image in Bayer format. Each polarization direction can obtain an image close to the sensor's nominal resolution, avoiding a significant decrease in spatial resolution caused by the polarization filter array. Adjacent pixel interpolation itself has a low-pass filtering effect, which can suppress high-frequency random noise and compensate for non-uniformity caused by differences in the transmittance of micro-polarization filters and inconsistent sensor responses.

[0028] Step S102: Input the environmental state perception data set, light intensity, haze concentration index and relative humidity into the environmental meta-feature encoder, and output the environmental state vector, wherein the environmental state vector includes the quantized specular reflection intensity of the current scene, atmospheric scattering coefficient, overall illuminance level and environmental corrosion factor.

[0029] In this embodiment, by using images and meteorological data, it is possible to distinguish whether the air humidity is high or the surface is actually damp, and whether the dampness causes severe reflection.

[0030] Step S103: Dynamically select the optimal detection model pipeline that matches the current environmental conditions based on the environmental state vector, and start the preprocessing channel according to the threshold of each component in the environmental state vector; wherein, the detection model pipeline includes the YOLO-World-R model optimized for high reflectivity scenarios, the YOLO-World-H model optimized for low-illuminance haze scenarios, and the YOLO-World-C model optimized for high-humidity salt spray corrosion scenarios; when the specular reflection intensity component exceeds the preset threshold, input the linear polarization degree image, polarization angle image, preset band near-infrared image and the original visible light image into the reflectivity component analysis network, and output the semantic segmentation mask of the specular reflection region; when the atmospheric scattering coefficient component exceeds the preset threshold, perform defogging processing on the original visible light image and haze concentration index to obtain a defogging enhanced image; combine the relative humidity with the overall illuminance level component in the environmental state vector to generate a humid reflectivity probability map.

[0031] In this embodiment, the YOLO-World-R model optimizes its specular reflection suppression capability by incorporating a large amount of high-reflection scene data during training, enabling it to learn to distinguish between real defects and reflection artifacts. The YOLO-World-H model enhances its feature extraction capabilities in low-light and hazy environments, exhibiting better defect recognition performance for blurred and low-contrast images. The YOLO-World-C model is trained for salt spray corrosion and wet surface scenarios, enabling it to identify marine-specific defects such as corrosion spots and salt crystals. When the specular reflection intensity exceeds a threshold, it is activated, using polarization and near-infrared information to accurately segment the specular reflection area. Targeted defogging enhancement is performed based on the atmospheric scattering coefficient threshold and haze concentration index to restore details obscured by haze. Furthermore, potential reflective areas are predicted based on humidity and illuminance information.

[0032] In one alternative approach, the YOLO-World-R model, the YOLO-World-H model, and the YOLO-World-C model are all based on the YOLO-World-v2 framework. The YOLO-World-v2 framework includes an image encoder, a text encoder, and a reparameterizable visual prediction head. The image encoder uses a CSPDarknet network to extract multi-scale visual features from the input image. The text encoder uses a pre-trained CLIP text encoder to encode the category names of photovoltaic panel defects into text feature vectors. The visual prediction head is used to locate and classify the fused visual features.

[0033] In this embodiment, the CLIP text encoder encodes the names of photovoltaic panel defect categories (such as "hot spots," "hidden cracks," and "corrosion") into text feature vectors, enabling the model to detect open-vocabulary targets. Even if a certain type of defect sample has not been seen during the training phase, it can still be effectively identified as long as its description is similar to an existing concept in the CLIP space, significantly improving the generalization detection capability of unknown or rare defects in complex marine scenarios. The image encoder adopts the CSPDarknet network structure, which effectively preserves shallow detail information and fuses deep semantic features through cross-stage local connections. It has high sensitivity to small defects on the surface of photovoltaic panels (such as fine cracks and local delamination), while also ensuring robustness in detecting large-scale abnormal areas (such as large-area stains and component damage). A reparameterizable visual prediction head is used, employing a multi-branch structure to enhance expressive power during the training phase and merging it into a single-path convolution during the inference phase. This improves training performance and reduces computational overhead during deployment, making it suitable for resource-constrained platforms such as marine inspection drones or surface robots. The YOLO-World-R / H / C models share a unified YOLO-World-v2 framework, with customization only for different harsh environments (high reflectivity, low-light haze, high-humidity salt spray) in the training data and optimization objectives. This ensures that the interface logic does not need to be refactored when dynamically switching detection model pipelines. The image encoder uses CSPDarknet-L or CSPDarknet-X as the backbone network. For the characteristics of the marine environment, an attention mechanism module can be added to CSPDarknet to make the network pay more attention to areas susceptible to environmental influences. The text encoder (CLIP) uses the text encoder portion of CLIP-ViT-B / 16 or CLIP-ViT-L / 14. The visual prediction head includes classification and regression branches, predicting class confidence and bounding box coordinates, respectively.

[0034] In one alternative approach, combining relative humidity with the overall illuminance level component in the environmental state vector to generate a damp reflectivity probability map further includes: Based on relative humidity, overall horizontal illuminance component, photovoltaic panel installation tilt angle and current solar altitude angle, the base reflectivity map is calculated using the Fresnel reflection model. The basic reflectivity map, linear polarization image, and environmental corrosion factor in the environmental state vector are input into a convolutional neural network, which outputs a wet reflectivity probability map in the range of [0,1].

[0035] In this embodiment, the Fresnel reflection model is used to calculate the basic reflectivity map, which can more accurately simulate and predict the reflection of photovoltaic panel surfaces in humid environments, improving the accuracy of detection results, especially in complex and variable marine environments. Considering the influence of different environmental factors (such as relative humidity and light intensity) on the reflectivity characteristics of photovoltaic panel surfaces, a humid reflectivity probability map for specific environmental conditions can be dynamically adjusted and generated. The basic reflectivity map, linear polarization image, and environmental corrosion factors are input into a convolutional neural network, taking into account not only changes in optical properties but also the effects of factors such as material aging. For example, in the early morning, due to sea fog, the relative humidity is high, and the sun has just risen, resulting in a small incident angle of light, which may cause significant humid reflectivity on the photovoltaic panel surface. Based on real-time relative humidity data provided by the micro-weather station, the overall illuminance level measured by the light sensor, combined with the specific installation tilt angle of the photovoltaic panel and the current solar altitude angle, a preliminary basic reflectivity map is calculated using the Fresnel reflection model. Further detailed analysis of the linear polarization image acquired by the multi-sensor module (which can distinguish between specular and diffuse reflection components) can improve the accuracy of identifying humid reflectivity areas. The above information is then input into a convolutional neural network to output a detailed probability map of moisture reflectivity, identifying areas that are easily misjudged as defects due to moisture reflectivity, thereby improving the overall accuracy of visual inspection. For example... Figure 3 As shown, the probability value of reflectivity in damp areas is close to 1 ([0.8,1]), corresponding to obvious water film or salt spray coverage on the photovoltaic panel surface. Medium-probability reflectivity areas are characterized by a moderate probability value of reflectivity in damp areas ([0.5,0.8]), possibly due to localized dampness or slight reflection. Low-probability reflectivity areas are characterized by a low probability value of reflectivity in damp areas ([0,0.5]), with reflection mainly caused by the photovoltaic panel material or minor defects. Low-probability reflectivity areas have a probability value of reflectivity in damp areas close to 0, corresponding to dry or non-reflective photovoltaic panel areas.

[0036] Step S104: Input the semantic segmentation mask of the specular reflection region, the dehazing enhanced image, and the wet reflection probability map into the multi-condition generative adversarial network to transform the original visible light image from the current harsh environment domain to the pseudo-standard illumination image domain to obtain the final enhanced image.

[0037] In this embodiment, the semantic segmentation mask of the specular reflection region, the dehazing enhanced image, and the wet reflectivity probability map are used as conditional inputs. The multi-condition generative adversarial network can identify and suppress imaging degradation caused by complex factors such as high humidity, strong light, and haze at sea, restoring the appearance of photovoltaic panels under near-standard lighting conditions, thereby significantly improving the accuracy of defect detection. Mapping the original image from the "harsh environment domain" to the "pseudo-standard lighting image domain" allows the defect detection model to be directly applied to complex maritime scenarios without retraining, significantly reducing deployment costs and improving generalization ability. For example, in a drone inspection under strong midday sunlight, the photovoltaic panel surface produces large areas of specular highlights due to water vapor condensation and the glass material, while the presence of thin fog in the distance reduces the overall image contrast. Directly using the original visible light image for defect detection, the YOLO model may misclassify the highlight areas as "hot spots" or "burnt areas," or miss hidden cracks under low contrast. In this application, the reflectivity analysis network uses polarization information to segment the highlight areas, generates a specular reflection mask, and enhances the image by combining it with the haze index. The wet reflectivity probability map indicates that battery cells near the edges and with smaller tilt angles are more prone to water accumulation and reflectivity. A multi-condition GAN integrates this information to generate a pseudo-standard image "taken under cloudy, windless, dry, and clean conditions" (highlights are suppressed, fog is removed, and water film reflections are corrected). The YOLO-World-R model accurately identifies a tiny microcrack in this enhanced image without misreporting normal highlight areas as defects, significantly improving detection reliability.

[0038] Step S105: Input the final enhanced image and the semantic segmentation mask of the specular reflection region into the optimal detection model pipeline for defect detection, and output the preliminary defect detection results with localization and classification information; perform uncertainty quantification evaluation on the preliminary defect detection results and identify areas with questionable reliability.

[0039] In this embodiment, for example, a certain offshore photovoltaic power station encountered strong afternoon sunlight and a high humidity environment on the sea surface, resulting in a large area of ​​specular reflection on the surface of the photovoltaic panels, while slight salt spray deposition occurred in the edge areas. Polarization imaging identified the central region of the panel as a high specular reflection area and generated a corresponding semantic segmentation mask. Multi-condition GAN converted the original blurred, overexposed visible light image into a pseudo-standard image, preserving the texture details of real defects (such as a transverse hidden crack). During detection, the YOLO-World-R model reduced the confidence of suspected "bright spots" in the central high-reflection area based on the mask information, while giving high-confidence predictions for clear hidden cracks in the edge low-reflection area. The uncertainty assessment module identified the hidden crack box as having low category entropy, stable position across multiple frames, small overlap area with the reflection mask, and consistent polarization response, thus determining it as a reliable detection. Another "circular bright spot" located in the high-reflection area was detected by the model, but its overlap rate with the reflection mask reached 85%, its category entropy was high, and its polarization consistency was poor, so it was marked as a highly suspicious area. A drone was scheduled to re-encode the highly suspicious area at a low angle and use the 1200nm near-infrared band to penetrate the salt spray. The secondary detection confirmed that it was a water reflection, thus eliminating false alarms.

[0040] In one alternative approach, performing an uncertainty quantification assessment on the preliminary defect detection results to identify areas of questionable reliability further includes: The preliminary results of statistical defect detection include the frequency of defect boxes, the variance of bounding box vertex coordinates, and the entropy value of the category prediction probability distribution. Calculate the local contrast of the defect box in the original visible light image, the polarization consistency in the linear polarization image, and the proportion of the overlap area with the semantic segmentation mask of the specular reflection region and the probability map of wet reflection. When the inspection equipment performs multi-angle or continuous frame observations on the same photovoltaic panel module, calculate the positional drift of the defect in the same spatial location in the detection results of 3 consecutive frames. Defect boxes with entropy values, overlap area ratios, and positional drift exceeding the first threshold are marked as highly suspicious regions; defect boxes between the first and second thresholds are marked as moderately suspicious regions; and those below the second threshold are considered reliable detection regions.

[0041] In this embodiment, surface reflections, salt spray deposits, and wet film reflections in the marine environment are often misidentified as defects. By calculating the overlap ratio between the defect frame and the specular reflection mask and the probability map of wet reflection, misjudgments caused by environmental interference can be accurately identified. The spatial stability of the detection results is judged by the positional drift of three consecutive frames (the real defect position should be relatively fixed, while false defects affected by light flicker or jitter show large drift), thereby achieving dynamic filtering. The results are divided into high / medium / low suspicion levels, providing a priority basis for manual review and improving the efficiency of operation and maintenance and the rationality of resource allocation. The evaluation indicators are all combined with interference factors unique to the marine environment, such as strong reflection, high humidity, and salt spray, and are highly scenario-specific. For example, a marine photovoltaic array was inspected by a patrol drone under the conditions of morning fog and gentle waves. Three anomalies were initially detected. Defect A is located in the center of the panel, with a stable bounding box, a positional drift of less than 2 pixels across three frames, a category entropy of 0.1, 5% overlap with the reflective mask, high local contrast, and consistent polarization angles. Defect B is near the edge, appearing only in one frame, with an entropy of 0.85, 70% overlap with the wet reflection probability map, and dispersed polarization angles. Defect C appears in all three frames, but its center point drifts by 15 pixels, 60% overlap with the specular reflection mask, and low local contrast. The uncertainty assessment results are as follows: Defect A: All indicators are below the second threshold → Reliable detection area (confirmed as a real hidden crack); Defect B: High entropy + high overlap → Exceeds the first threshold → Highly suspicious area (judged as a morning dew reflection artifact); Defect C: Large drift + high overlap → Between the two thresholds → Medium suspicious area (possibly a minor stain, but affected by wave reflection). The drone was dispatched to take low-angle re-images of areas B and C and to activate the 1200nm near-infrared band. After the second inspection, defect B disappeared, and defect C was confirmed to exist but its area was reduced. After the joint confidence level was updated, it was converted into a reliable area and a manual review warning was triggered.

[0042] In an alternative approach, the method further includes: When any photovoltaic panel module is marked as a highly suspicious area or a medium-suspicious defect area, the flight altitude, shooting angle and sensor exposure parameters of the inspection equipment are controlled to re-acquire image data of that area; During the resampling process, the 1200±20nm near-infrared band image is prioritized to penetrate the salt spray surface; the two detection results are jointly updated with confidence, and if the area is considered a reliable detection area, a manual review warning is triggered.

[0043] In this embodiment, preliminary detection results are not used as the final conclusion. Instead, targeted resampling is proactively triggered through uncertainty assessment, significantly improving the adaptability in complex marine environments. The 1200±20nm near-infrared band is prioritized, utilizing its stronger penetration of water vapor and salt spray to obtain image information closer to the actual surface condition, thereby distinguishing between real defects and artifacts caused by surface deposits. The results of the initial detection (visible light + enhanced image) and the secondary resampling (optimized viewing angle + near-infrared) are jointly updated with confidence, avoiding random errors from single modalities or single observations and improving the accuracy of the final judgment. Resampling is triggered only for "medium / highly suspicious areas" instead of full-frame re-imaging, significantly reducing energy consumption and time costs; simultaneously, optimal secondary imaging quality is ensured by dynamically adjusting flight altitude, angle, and exposure. Manual review and warning are only triggered when the joint assessment still determines it to be a "reliable defect," reducing the burden of manual intervention while ensuring that critical defects are not overlooked.

[0044] According to the solution provided by the present invention, a multi-sensor module installed on an inspection device synchronously acquires raw visible light images and preset band near-infrared images of the target photovoltaic array; a polarization imaging unit acquires linear polarization degree images and polarization angle images; and a micro weather station acquires light intensity, haze concentration index, and relative humidity in real time. The raw visible light images, preset band near-infrared images, linear polarization degree images, and polarization angle images are spatially registered and timestamped to form an environmental state perception data set. The environmental state perception data set, light intensity, haze concentration index, and relative humidity are input to an environmental element feature encoder, which outputs an environmental state vector. This environmental state vector includes quantized specular reflection intensity, atmospheric scattering coefficient, overall illuminance level, and environmental corrosion factor of the current scene. Based on the environmental state vector, the optimal detection model pipeline matching the current environmental conditions is dynamically selected, and a preprocessing channel is activated based on the threshold values ​​of each component in the environmental state vector. The detection model pipeline includes a YOLO-World-R model optimized for high-reflection scenes and a YOLO-World-R model optimized for low-illuminance haze scenes. The OLO-World-H model and the YOLO-World-C model optimized for high-humidity salt spray corrosion scenarios are used. When the specular reflection intensity component exceeds a preset threshold, the linear polarization degree image, polarization angle image, preset band near-infrared image, and original visible light image are input into the reflection component analysis network to output a semantic segmentation mask for the specular reflection region. When the atmospheric scattering coefficient component exceeds a preset threshold, the original visible light image and haze concentration index are dehazed to obtain a dehazed enhanced image. The relative humidity is combined with the overall illuminance level component in the environmental state vector to generate a wet reflectivity probability map. The semantic segmentation mask for the specular reflection region, the dehazed enhanced image, and the wet reflectivity probability map are input into a multi-condition generative adversarial network to transform the original visible light image from the current harsh environment domain to the pseudo-standard illumination image domain to obtain the final enhanced image. The final enhanced image and the semantic segmentation mask for the specular reflection region are input into the optimal detection model pipeline for defect detection, and the preliminary defect detection results with localization and classification information are output. The uncertainty quantification assessment of the preliminary defect detection results is performed to identify areas with questionable reliability. This invention improves the accuracy of visual inspection of photovoltaic panels in harsh marine environments by automatically selecting the optimal detection model pipeline through real-time environmental state vectors.

[0045] Figure 5 A schematic diagram of the frame of an adaptive optimization device for visual inspection of photovoltaic panels in harsh marine environments, according to an embodiment of the present invention, is shown. The adaptive optimization device for visual inspection of photovoltaic panels in harsh marine environments includes: The environmental perception data acquisition module 510 is used to simultaneously acquire the original visible light image and the preset band near-infrared image of the target photovoltaic array through a multi-sensor module installed on the inspection equipment; acquire the linear polarization degree image and the polarization angle image through the polarization imaging unit; and acquire the light intensity, haze concentration index and relative humidity in real time through the micro weather station; and perform spatial registration and timestamp alignment of the original visible light image, the preset band near-infrared image, the linear polarization degree image and the polarization angle image to form an environmental state perception data set. The environmental meta-feature quantization module 520 is used to input the environmental state perception data set, light intensity, haze concentration index and relative humidity into the environmental meta-feature encoder and output an environmental state vector, wherein the environmental state vector includes quantified specular reflection intensity, atmospheric scattering coefficient, overall illuminance level and environmental corrosion factor of the current scene. The adaptive detection pipeline module 530 is used to dynamically select the optimal detection model pipeline that matches the current environmental conditions based on the environmental state vector, and to start the preprocessing channel based on the threshold of each component in the environmental state vector. The detection model pipeline includes a YOLO-World-R model optimized for high-reflection scenarios, a YOLO-World-H model optimized for low-illuminance haze scenarios, and a YOLO-World-C model optimized for high-humidity salt spray corrosion scenarios. When the specular reflection intensity component exceeds a preset threshold, the linear polarization degree image, polarization angle image, preset band near-infrared image, and original visible light image are input into the reflectance component analysis network to output a semantic segmentation mask for the specular reflection region. When the atmospheric scattering coefficient component exceeds a preset threshold, the original visible light image and haze concentration index are defogging to obtain a defogging enhanced image. The relative humidity is combined with the overall illuminance level component in the environmental state vector to generate a humid reflectivity probability map. The image domain transformation and enhancement module 540 is used to input the semantic segmentation mask of the specular reflection region, the dehazing and enhancement image and the wet reflection probability map into the multi-condition generative adversarial network to transform the original visible light image from the current harsh environment domain to the pseudo-standard illumination image domain to obtain the final enhanced image. The defect detection and evaluation module 550 is used to input the final enhanced image and the semantic segmentation mask of the specular reflection region into the optimal detection model pipeline for defect detection, and output preliminary defect detection results with localization and classification information; perform uncertainty quantification evaluation on the preliminary defect detection results, and identify areas with questionable reliability.

[0046] Figure 6 The diagram shows a structural schematic of an embodiment of the computing device of the present invention. The specific embodiments of the present invention do not limit the specific implementation of the computing device.

[0047] like Figure 6As shown, the computing device may include: a processor 602, a communications interface 604, a memory 606, and a communications bus 608.

[0048] The processor 602, communication interface 604, and memory 606 communicate with each other via communication bus 608. Communication interface 604 is used to communicate with other network elements, such as clients or other servers. Processor 602 executes program 610, specifically performing the relevant steps in the above-described embodiment of the adaptive optimization method for photovoltaic panel visual inspection in harsh marine environments.

[0049] Specifically, program 610 may include program code that includes computer operation instructions.

[0050] Processor 602 may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement embodiments of the present invention. The computing device includes one or more processors, which may be processors of the same type, such as one or more CPUs; or processors of different types, such as one or more CPUs and one or more ASICs.

[0051] Memory 606 is used to store program 610. Memory 606 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk storage device.

[0052] According to the solution provided by the present invention, a multi-sensor module installed on the inspection equipment synchronously acquires the original visible light image and the preset band near-infrared image of the target photovoltaic array; a polarization imaging unit acquires the linear polarization degree image and the polarization angle image; and a micro weather station acquires the light intensity, haze concentration index, and relative humidity in real time. The original visible light image, the preset band near-infrared image, the linear polarization degree image, and the polarization angle image are spatially registered and time-stamped to form an environmental state perception data set. The environmental state perception data set, light intensity, haze concentration index, and relative humidity are input to an environmental meta-feature encoder, which outputs an environmental state vector. The environmental state vector includes quantized specular reflection intensity, atmospheric scattering coefficient, overall illuminance level, and environmental corrosion factor of the current scene. The optimal detection model pipeline matching the current environmental conditions is dynamically selected based on the environmental state vector, and the preprocessing channel is activated based on the threshold of each component in the environmental state vector. The detection model pipeline includes a YOLO-World-R model optimized for high-reflection scenes and a YOLO-World-R model optimized for low-illuminance haze scenes. The OLO-World-H model and the YOLO-World-C model optimized for high-humidity salt spray corrosion scenarios are used. When the specular reflection intensity component exceeds a preset threshold, the linear polarization degree image, polarization angle image, preset band near-infrared image, and original visible light image are input into the reflection component analysis network to output a semantic segmentation mask for the specular reflection region. When the atmospheric scattering coefficient component exceeds a preset threshold, the original visible light image and haze concentration index are dehazed to obtain a dehazed enhanced image. The relative humidity is combined with the overall illuminance level component in the environmental state vector to generate a wet reflectivity probability map. The semantic segmentation mask for the specular reflection region, the dehazed enhanced image, and the wet reflectivity probability map are input into a multi-condition generative adversarial network to transform the original visible light image from the current harsh environment domain to the pseudo-standard illumination image domain to obtain the final enhanced image. The final enhanced image and the semantic segmentation mask for the specular reflection region are input into the optimal detection model pipeline for defect detection, and the preliminary defect detection results with localization and classification information are output. The uncertainty quantification assessment of the preliminary defect detection results is performed to identify areas with questionable reliability. This invention improves the accuracy of visual inspection of photovoltaic panels in harsh marine environments by automatically selecting the optimal detection model pipeline through real-time environmental state vectors.

[0053] Those skilled in the art will understand that modules in the device of the embodiments can be adaptively changed and placed in one or more devices different from that embodiment. Modules, units, or components in the embodiments can be combined into a single module, unit, or component, and further, they can be divided into multiple sub-modules, sub-units, or sub-components. Except where at least some of such features and / or processes or units are mutually exclusive, any combination of all features disclosed in this specification (including the accompanying claims, abstract, and drawings) and all processes or units of any method or device so disclosed can be employed. Unless expressly stated otherwise, each feature disclosed in this specification (including the accompanying claims, abstract, and drawings) may be replaced by an alternative feature that serves the same, equivalent, or similar purpose. Furthermore, those skilled in the art will understand that although some embodiments herein include certain features included in other embodiments but not others, combinations of features from different embodiments are intended to be within the scope of the invention and form different embodiments. For example, in the following claims, any of the claimed embodiments can be used in any combination. The invention can be implemented by means of hardware comprising several different elements and by means of a suitably programmed computer. In the unit claims listing several devices, several of these devices may be embodied by the same hardware item. Unless otherwise specified, the steps in the above embodiments should not be construed as limiting the order of execution.

Claims

1. An adaptive optimization method for visual inspection of photovoltaic panels in harsh marine environments, characterized in that, include: The system synchronously acquires raw visible light images and preset band near-infrared images of the target photovoltaic array through a multi-sensor module installed on the inspection equipment; obtains linear polarization degree images and polarization angle images through a polarization imaging unit; and obtains light intensity, haze concentration index and relative humidity in real time through a micro weather station. The system then performs spatial registration and timestamp alignment of the raw visible light images, preset band near-infrared images, linear polarization degree images and polarization angle images to form an environmental state perception data set. The environmental state perception data set, light intensity, haze concentration index and relative humidity are input into the environmental meta-feature encoder, and the environmental state vector is output. The environmental state vector includes the quantized specular reflection intensity, atmospheric scattering coefficient, overall illuminance level and environmental corrosion factor of the current scene. The optimal detection model pipeline matching the current environmental conditions is dynamically selected based on the environmental state vector, and the preprocessing channel is activated based on the threshold values ​​of each component in the environmental state vector. The detection model pipeline includes a YOLO-World-R model optimized for high-reflection scenarios, a YOLO-World-H model optimized for low-illuminance haze scenarios, and a YOLO-World-C model optimized for high-humidity salt spray corrosion scenarios. When the specular reflection intensity component exceeds a preset threshold, the linear polarization degree image, polarization angle image, preset band near-infrared image, and original visible light image are input into the reflectance component analysis network, outputting a semantic segmentation mask for the specular reflection region. When the atmospheric scattering coefficient component exceeds a preset threshold, the original visible light image and haze concentration index are defogging to obtain a defogging enhanced image. The relative humidity is combined with the overall illuminance level component in the environmental state vector to generate a humid reflectivity probability map. The semantic segmentation mask of the specular reflection region, the dehazing and enhanced image and the wet reflection probability map are input into the multi-condition generative adversarial network to transform the original visible light image from the current harsh environment domain to the pseudo-standard illumination image domain to obtain the final enhanced image. The final enhanced image and the semantic segmentation mask of the specular reflection region are input into the optimal detection model pipeline for defect detection, and the output is a preliminary defect detection result with localization and classification information; the uncertainty of the preliminary defect detection result is evaluated by uncertainty quantification, and regions with questionable reliability are identified.

2. The adaptive optimization method for visual inspection of photovoltaic panels in harsh marine environments according to claim 1, characterized in that, The inspection equipment is an autonomous cruise drone or a surface robot; the multi-sensor module consists of a global shutter RGB visible light camera, two near-infrared cameras equipped with different optical bandpass filters, and a polarization imaging unit; the wavelength distribution of the two near-infrared cameras is 850±10nm and 1200±20nm.

3. The adaptive optimization method for visual inspection of photovoltaic panels in harsh marine environments according to claim 1, characterized in that, The polarization imaging unit employs a CMOS image sensor with an on-chip polarization filter array. Each superpixel unit of the on-chip polarization filter array is composed of subpixels covered by micro-polarization filters in four directions: 0°, 45°, 90°, and 135°.

4. The adaptive optimization method for visual inspection of photovoltaic panels in harsh marine environments according to claim 1, characterized in that, Acquiring linear polarization degree images and polarization angle images through a polarization imaging unit further includes: The original polarized Bayer image output by the polarization imaging unit is reconstructed into four independent polarization direction intensity images through interpolation of adjacent pixels; The first three Stokes parameters are calculated based on the four polarization direction intensity images, and the linear polarization degree image and polarization angle image are obtained by solving the first three Stokes parameters.

5. The adaptive optimization method for visual inspection of photovoltaic panels in harsh marine environments according to claim 1, characterized in that, The YOLO-World-R model, the YOLO-World-H model, and the YOLO-World-C model are all based on the YOLO-World-v2 framework. The YOLO-World-v2 framework includes an image encoder, a text encoder, and a reparameterizable visual prediction head. The image encoder uses a CSPDarknet network to extract multi-scale visual features from the input image. The text encoder uses a pre-trained CLIP text encoder to encode the category names of photovoltaic panel defects into text feature vectors. The visual prediction head is used to locate and classify the fused visual features.

6. The adaptive optimization method for visual inspection of photovoltaic panels in harsh marine environments according to claim 1, characterized in that, The generation of a damp reflectivity probability map by combining relative humidity with the overall illuminance level component in the environmental state vector further includes: Based on relative humidity, overall horizontal illuminance component, photovoltaic panel installation tilt angle and current solar altitude angle, the base reflectivity map is calculated using the Fresnel reflection model. The basic reflectivity map, linear polarization image, and environmental corrosion factor in the environmental state vector are input into a convolutional neural network, which outputs a wet reflectivity probability map in the range of [0,1].

7. The adaptive optimization method for visual inspection of photovoltaic panels in harsh marine environments according to claim 1, characterized in that, An uncertainty quantification assessment was performed on the preliminary defect detection results, and areas of questionable reliability were further identified, including: The preliminary results of statistical defect detection include the frequency of defect boxes, the variance of bounding box vertex coordinates, and the entropy value of the category prediction probability distribution. Calculate the local contrast of the defect box in the original visible light image, the polarization consistency in the linear polarization image, and the proportion of the overlap area with the semantic segmentation mask of the specular reflection region and the probability map of wet reflection. When the inspection equipment performs multi-angle or continuous frame observations on the same photovoltaic panel module, calculate the positional drift of the defect in the same spatial location in the detection results of 3 consecutive frames. Defect boxes with entropy values, overlap area ratios, and positional drift exceeding the first threshold are marked as highly suspicious regions; defect boxes between the first and second thresholds are marked as moderately suspicious regions; and those below the second threshold are considered reliable detection regions.

8. The adaptive optimization method for visual inspection of photovoltaic panels in harsh marine environments according to claim 7, characterized in that, The method further includes: When any photovoltaic panel module is marked as a highly suspicious area or a medium-suspicious defect area, the flight altitude, shooting angle and sensor exposure parameters of the inspection equipment are controlled to re-acquire image data of that area; During the resampling process, the 1200±20nm near-infrared band image is prioritized to penetrate the salt spray surface; the two detection results are jointly updated with confidence, and if the area is considered a reliable detection area, a manual review warning is triggered.

9. An adaptive optimization device for visual inspection of photovoltaic panels designed for harsh marine environments, characterized in that, include: The environmental perception data acquisition module is used to simultaneously acquire raw visible light images and preset band near-infrared images of the target photovoltaic array through a multi-sensor module installed on the inspection equipment; acquire linear polarization degree images and polarization angle images through a polarization imaging unit; and acquire light intensity, haze concentration index and relative humidity in real time through a micro weather station; and spatially register and timestamp align the raw visible light images, preset band near-infrared images, linear polarization degree images and polarization angle images to form an environmental state perception data set. The environmental meta-feature quantization module is used to input the environmental state perception data set, light intensity, haze concentration index and relative humidity into the environmental meta-feature encoder, and output an environmental state vector, wherein the environmental state vector includes quantified specular reflection intensity, atmospheric scattering coefficient, overall illuminance level and environmental corrosion factor of the current scene. An adaptive detection pipeline module is used to dynamically select the optimal detection model pipeline that matches the current environmental conditions based on the environmental state vector, and to activate the preprocessing channel based on the threshold values ​​of each component in the environmental state vector. The detection model pipeline includes a YOLO-World-R model optimized for high-reflection scenarios, a YOLO-World-H model optimized for low-illuminance haze scenarios, and a YOLO-World-C model optimized for high-humidity salt spray corrosion scenarios. When the specular reflection intensity component exceeds a preset threshold, the linear polarization degree image, polarization angle image, preset band near-infrared image, and original visible light image are input into the reflectance component analysis network to output a semantic segmentation mask for the specular reflection region. When the atmospheric scattering coefficient component exceeds a preset threshold, the original visible light image and haze concentration index are defogging to obtain a defogging enhanced image. The relative humidity is combined with the overall illuminance level component in the environmental state vector to generate a humid reflectivity probability map. The image domain transformation and enhancement module is used to input the semantic segmentation mask of the specular reflection region, the dehazing and enhancement image and the wet reflection probability map into the multi-condition generative adversarial network to transform the original visible light image from the current harsh environment domain to the pseudo-standard illumination image domain to obtain the final enhanced image; The defect detection and evaluation module is used to input the final enhanced image and the semantic segmentation mask of the specular reflection region into the optimal detection model pipeline for defect detection, and output preliminary defect detection results with localization and classification information; it performs uncertainty quantification evaluation on the preliminary defect detection results and identifies areas with questionable reliability.

10. A computing device, comprising: The processor, memory, communication interface, and communication bus are provided, wherein the processor, memory, and communication interface communicate with each other via the communication bus. The memory is used to store at least one executable instruction, which causes the processor to perform the operation corresponding to the above-described adaptive optimization method for visual inspection of photovoltaic panels in harsh marine environments.

Images