Photovoltaic module defect visual inspection system based on unmanned aerial vehicle collected images

Abstract

The application relates to the technical field of inspection, and discloses a photovoltaic module defect visual detection system based on unmanned aerial vehicle image collection, which comprises a receiving unit, an image acquisition unit, a geometric reference feature extraction unit, a spatial dimension regularization unit and a target defect detection unit.The geometric reference feature extraction unit filters out texture interference according to the direction entropy of a gradient direction moment feature, and determines an image vanishing point; the spatial dimension regularization unit constructs a homography transformation matrix, inversely maps perspective distortion to a Euclidean plane to generate an orthographic feature map; and the target defect detection unit constructs a logical grid coordinate system according to the orthographic feature map, extracts a unitized image and identifies morphological abnormalities.The application uses a target object physical structure prior to construct a geometric rigidity constraint, eliminates perspective deformation, ensures that pixel displacement is consistent with physical span, solves the problem of detection precision deviation, and improves the signal-to-noise ratio of defect features in a complex environment.

Description

Technical Field

[0001] This invention relates to a visual inspection system for defects in photovoltaic modules based on images acquired by drones, belonging to the field of inspection technology. Background Technology

[0002] Currently, using drones equipped with visual sensors to collect images of components and identify defects is the mainstream technical approach to replace manual inspection and improve maintenance efficiency. This approach uses visual sensors to obtain surface features of components and determine fault states such as hot spots, damage, and microcracks. However, in actual inspections, drones are subject to dynamic pose shifts due to airflow interference and are affected by the undulating slope of mountainous terrain, resulting in an uncertain angle between the camera's imaging optical axis and the photovoltaic module plane. This causes perspective distortion during projection imaging, leading to nonlinear collapse of the geometric shape of the component in image space.

[0003] To address perspective distortion, the industry typically compensates for the acquired pose by adding high-precision positioning modules or constructing image models. However, adding external hardware modules increases system deployment costs and poses a risk of signal instability in complex environments. Reconstruction methods relying on preset models require high processor computing power, making it difficult to achieve real-time geometric calibration in UAV embedded terminals. Existing improvements focus on indirect compensation of image quality through preset waypoint guidance or deep learning algorithm optimization. For example, Chinese invention patent CN116223511A discloses a method and device for diagnosing defects in distributed rooftop photovoltaic modules based on UAV automatic inspection. It uses gimbal control vectors to adjust the camera angle so that the lens is perpendicular to the photovoltaic panel plane for shooting. Such solutions rely on the mechanical following accuracy of the gimbal. When airflow disturbances cause high-frequency attitude changes, mechanical compensation has a response lag. Back-end recognition uses convolutional neural networks to fit residual distortion image features, but does not solve the projection collapse problem at the geometric mapping level. While improving algorithm stability to combat geometric distortion, it cannot ensure the consistency of pixel displacement with physical span measurement scale, and the detection accuracy is constrained by flight attitude fluctuations.

[0004] Therefore, the technical problem to be solved by this invention is to utilize the geometric prior of the photovoltaic array itself as a constraint to achieve adaptive correction of perspective distortion and reconstruction of the standard feature space while reducing hardware dependence. Summary of the Invention

[0005] To address the problems mentioned in the background art, the technical solution of the present invention is as follows: A visual inspection system for photovoltaic module defects based on images acquired by a drone, the system comprising: The raw image signal receiving unit is used to acquire the raw inspection images containing planar array targets collected by the imaging platform; The geometric reference feature extraction unit is used to identify the gradient direction moment features of the corresponding planar array target edges in the original inspection image, calculate the direction entropy of the gradient direction moment features in a preset neighborhood, lock the structured linear feature flow by filtering out non-target texture interference terms with direction entropy less than a preset threshold, and determine the vanishing point of the image space based on the intersection geometric constraints of the structured linear feature flow. The spatial dimension regularization unit is used to construct a homography transformation matrix based on the mapping relationship between the vanishing point and the preset rectangular reference, and to perform inverse mapping processing on the perspective distortion area in the original inspection image to the Euclidean coordinate plane to generate an orthophoto feature map with geometric rigidity. The target defect detection unit is used to extract the unitized target image corresponding to each grid index in the logical grid coordinate system constructed by the orthophoto feature map, calculate the deviation of the unitized target image relative to the local brightness background, determine the defect pixel coordinates based on the pixel area where the deviation exceeds the preset abnormal threshold, and identify the morphological abnormal features corresponding to the defect pixel coordinates.

[0006] Preferably, the geometric reference feature extraction unit is also used to calculate the gradient energy centroid of the structured linear feature flow in the edge normal direction, so as to refine the positioning coordinates of the gradient direction moment feature to the sub-pixel level, and use the least squares fitting algorithm to perform linear regression processing on the sub-pixel level positioning points to generate calibrated feature line parameters. The geometric reference feature extraction unit corrects the coordinate position of the vanishing point according to the calibrated feature line parameters.

[0007] Preferably, when determining the vanishing point, the geometric reference feature extraction unit is also used to perform the following steps: using the vanishing point as the origin of radiation, generating a virtual ray cluster in the space of the original inspection image; performing geometric consensus verification between the virtual ray cluster and the gradient direction moment feature, and removing linear noise components that do not conform to the constraints of the origin of radiation.

[0008] Preferably, in the process of generating orthophoto maps, the spatial dimension regularization unit is also used to construct geometric virtual constraints by utilizing the translational symmetry of the planar array targets, and to perform feedback correction on the mapping parameters of the homography transformation matrix by comparing the pixel span consistency of adjacent unitized target images in the orthophoto map.

[0009] Preferably, the logic for performing feedback correction in the spatial dimension regularization unit includes: calculating the reprojection residual R for pixel span consistency, using the following formula: Where xi is the preset theoretical center x-coordinate of the i-th unitized target image in the standard Euclidean coordinate plane, and yi is the corresponding preset theoretical center y-coordinate; xi is the actual center x-coordinate in the pixel space after homography transformation matrix mapping, and yi is the corresponding actual center y-coordinate; n is the total number of unitized target images; when the reprojection residual R is greater than the preset deviation threshold, the local coordinate perturbation algorithm is executed to compensate for the geometric distortion caused by the non-coplanarity of the planar array targets.

[0010] Preferably, when identifying morphological abnormality features, the target defect detection unit is also used to perform the following steps: extracting the confluence texture distribution features inside the unitized target image; determining the geometric topological reference of the confluence texture distribution features based on the logical grid coordinate system; and determining that there is structural damage inside the unitized target image by identifying regions that have morphological deviations or brightness abrupt changes relative to the geometric topological reference.

[0011] Preferably, the target defect detection unit is also used to: calculate the average gray value of each grid cell in the statistical logical grid coordinate system, calculate the ratio of the average gray value to the preset global background brightness, and perform gray-scale mapping enhancement on the orthophoto feature map according to the ratio.

[0012] Preferably, the system also includes a trend monitoring module, which records data on the morphological anomalies of the same unitized target image in different inspection cycles, calculates the data evolution rate over time, and outputs a fault warning signal for the corresponding unitized target image when the evolution rate exceeds a preset degradation threshold.

[0013] Preferably, the original image signal receiving unit is further configured to: receive the external orientation element metadata of the imaging platform at the acquisition time; determine the initial spatial projection angle of the optical axis of the imaging platform relative to the planar array target based on the external orientation element metadata, and use the initial spatial projection angle as the seed parameter for constructing the homography transformation matrix by the spatial dimension warping unit.

[0014] Preferably, the system also includes a coordinate transformation module, which is used to map the defect pixel coordinates determined by the target defect detection unit to the geospatial coordinate system of the global navigation satellite system, and generate an inspection result report containing the defect type and its geospatial coordinates.

[0015] Compared with the prior art, the beneficial effects of the present invention are: 1. In the visual inspection of defects in photovoltaic modules, geometric constraints generated by the gradient features of the photovoltaic array frame are used to establish an inverse mapping from the image space to the standard Euclidean coordinate plane. The homography transformation matrix is ​​derived using the vanishing point and the feature space is reconstructed, so that the perspective distortion caused by the field of view offset of the acquisition device in the image is eliminated before the inspection is intervened. This coordinate regularization action based on the prior knowledge of the physical structure ensures the consistency between pixel displacement and the physical span of the module, so that the subsequent defect identification process for the cell area is no longer affected by flight attitude or terrain slope.

[0016] 2. Based on the virtual ray clusters radiating to the image plane formed by the vanishing points, the original border feature flow is subjected to poloidal geometric consensus verification. The geometric parameter calculation results are reused in reverse as feature filters to eliminate linear interference terms unrelated to the photovoltaic physical structure. This enables the system to identify and isolate interference from component surface shadows, support frame edges, or debris, ensuring the purity of the physical meaning of the homography transformation matrix solution process and avoiding geometric calibration offset caused by structured noise.

[0017] 3. By combining first-order statistical moment analysis of gradient distribution curves with local brightness normalization processing guided by logical grids, sub-pixel-level geometric refinement and environmental adaptive enhancement are achieved. By extracting the gradient energy centroid in the edge normal direction, the limitation of physical sensor resolution on positioning accuracy is broken. At the same time, the grid template generated by spatial reconstruction is used to perform unitized statistical normalization. This dynamic linkage between the geometric framework and radiation intensity enables the system to still lock high-frequency defect features such as microcracks under conditions of strong reflection or local underexposure, effectively improving the signal-to-noise ratio of the detection signal. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the core image processing and defect recognition logic of the system of this invention. Figure 2 This is a diagram showing the overall system architecture and data signal flow of the present invention. Detailed Implementation

[0019] This specific embodiment is intended to describe the present invention in detail. The following embodiments are intended for explanation and illustration, and are not intended to limit the scope of protection of the present invention.

[0020] A visual inspection system for photovoltaic module defects based on images acquired by a drone includes a raw image signal receiving unit, a geometric reference feature extraction unit, a spatial dimension regularization unit, and a target defect detection unit. The raw image signal receiving unit is connected to the geometric reference feature extraction unit and is used to acquire raw inspection images containing planar array targets acquired by an imaging platform. The imaging platform is configured as a drone equipped with a visual sensor. During the acquisition process, the raw image signal receiving unit receives the external orientation element metadata of the imaging platform at the acquisition time. The external orientation element metadata includes the flight attitude angle and spatial coordinates at the imaging time. The system determines the initial spatial projection angle of the optical axis of the imaging platform relative to the planar array target based on the external orientation element metadata, and uses it as the seed parameter for subsequent construction of the transformation matrix. The system extracts the pitch angle component and roll angle component output by the flight control module in real time, performs lookup calculations using a linearized discrete table of sine and cosine functions, and maps them to phase... For the rotational offset of the optical axis center, and combined with the 35mm equivalent focal length pixel value preset by the vision sensor, eight initial elements representing rotation and translation are generated in the homography transformation matrix according to a 1.0 scaling factor, to guide the algorithm to lock the starting point of the search space within 50 milliseconds; in the initial offset engineering calibration of the optical axis of the imaging platform, before the inspection operation starts, the UAV imaging platform is controlled to align with the rectangular reference part with a known geometric ratio on the horizontal ground, and the original image of the vision sensor under zero attitude command is obtained. The pixel coordinates of the four corners of the rectangular reference part in the image coordinate system are extracted, the perspective shrinkage ratio is calculated, and the real-time pitch angle and roll angle values ​​output by the inertial measurement unit are read simultaneously. The theoretical projection tilt angle obtained by inversely deducing the image is calculated with the difference between the actual attitude angle measured by the sensor, and the installation deviation component of the vision sensor relative to the body coordinate system is determined. This component is injected as a static compensation parameter into the initial calculation logic of the homography transformation matrix H to compensate for the initial geometric distortion deviation caused by the assembly tolerance of the UAV structural components.

[0021] The geometric reference feature extraction unit processes the original inspection image to extract geometric constraints. Addressing the issue of background interference affecting the photovoltaic module frame features, the geometric reference feature extraction unit identifies the gradient direction moment features of the corresponding planar array target edges in the original inspection image. The system calculates the direction entropy of the gradient direction moment features within a preset neighborhood. Specifically, a 3×3 Sobel operator is used to extract the gradient magnitude and direction of each pixel. The probability density pj of the gradient direction distribution is statistically analyzed within a 5×5 neighborhood window, and the direction entropy is calculated according to the following formula. Where E is the directional entropy, pj is the probability distribution of the j-th directional interval, and m is the total number of directional intervals; the geometric reference feature extraction unit locks the structured linear feature flow by filtering out non-target texture interference terms with directional entropy greater than a preset threshold Te; where Te is set to 0.6; the system determines the vanishing point of the image space based on the intersection geometric constraints of the structured linear feature flow; in the directional entropy calculation sampling accuracy and threshold calibration, the system divides the [0,π] gradient angle space into 18 equally spaced sub-intervals according to the discrete angle resolution of the visual sensor, so that the directional interval angle span is constant at 10°, and the probability distribution pj of each pixel gradient direction falling into the sub-interval is statistically analyzed within a 5×5 sliding neighborhood window. Based on the physical consistency of the gradient vector at the edge of the photovoltaic module, the obtained directional entropy E is lower than the directional entropy of the isotropic background texture of vegetation. Under typical working conditions, 1 The system uses 00 sets of samples comparing component edges with cluttered environments to calculate the maximum segmentation value that maximizes the sum of background texture filtering rate and target edge retention rate. This value is set as the directional entropy threshold Te, enabling automated extraction of structured linear feature flows. To improve positioning accuracy, the geometric reference feature extraction unit calculates the gradient energy centroid of the structured linear feature flow in the edge normal direction, refining the positioning coordinates of the gradient direction moment features to the sub-pixel level. The system uses a least squares fitting algorithm to perform linear regression processing on the sub-pixel level positioning points, generating calibrated feature line parameters. The geometric reference feature extraction unit corrects the coordinate position of the vanishing point based on the calibrated feature line parameters. When determining the vanishing point, the system generates a virtual ray cluster with the vanishing point as the radiation origin and performs geometric consensus verification between the virtual ray cluster and the gradient direction moment features, eliminating linear noise components that do not conform to the radiation origin constraint.

[0022] Spatial dimension warping units are used to correct perspective distortion. These units construct a homography transformation matrix H based on the mapping relationship between the vanishing point and a preset rectangular reference. The system initializes the homography transformation matrix H using seed parameters and performs an inverse mapping process to the Euclidean coordinate plane on the perspective distortion regions in the original inspection image, generating an orthophoto feature map with geometric rigidity. During generation, the system utilizes the translational symmetry of the planar array targets to construct geometric virtual constraints. By comparing the pixel span consistency of adjacent unitized target images in the orthophoto feature map, the system calculates the reprojection residual R, using the following formula: Where: R is the reprojection residual, xi is the preset theoretical center x-coordinate of the i-th unitized target image in the standard Euclidean coordinate plane, yi is the corresponding preset theoretical center y-coordinate, xi is the actual center x-coordinate in pixel space after homography transformation matrix mapping, yi is the corresponding actual center y-coordinate, and n is the total number of unitized target images; when the reprojection residual R is greater than the preset deviation threshold Tr, the system executes a local coordinate perturbation algorithm to compensate for the geometric distortion caused by the non-coplanarity of the planar array targets; where Tr is set to 1.5 pixels.

[0023] The target defect detection unit is used to identify component damage; it extracts the unitized target image corresponding to each grid index in the logical grid coordinate system constructed by the orthophoto feature map; the system calculates the deviation of the unitized target image relative to the local brightness background; the target defect detection unit statistically analyzes the average gray value of each grid unit in the logical grid coordinate system, calculates the ratio of the average gray value to the preset global background brightness, and performs gray-scale mapping enhancement on the orthophoto feature map based on the ratio; it determines the defect pixel coordinates based on the pixel region where the deviation exceeds the preset anomaly threshold; when identifying morphological anomalies, the system extracts the confluence texture distribution features inside the unitized target image and determines the geometric topological reference of the confluence texture distribution features based on the logical grid coordinate system; it identifies morphological changes relative to the geometric topological reference. For areas with deviations or sudden brightness changes, structural damage is determined to exist within the unitized target image. The system also includes a trend monitoring module and a coordinate transformation module. The trend monitoring module records the morphological anomaly features of the same unitized target image in different inspection cycles and calculates the data evolution rate over time. When the evolution rate exceeds a preset degradation threshold, a fault warning signal is output. The coordinate transformation module is used to map the defect pixel coordinates determined by the target defect detection unit to the geospatial coordinate system of the Global Navigation Satellite System. The system generates an inspection result report containing the defect type and its geospatial coordinates. Through the logical collaboration between the above units, the distorted inspection image is converted into a standard coordinate space, ensuring that the pixel displacement is consistent with the physical span of the component, thus achieving precise positioning of sub-centimeter level defects.

[0024] Example 1: In the inspection scenario of a mountainous photovoltaic power station at an altitude greater than 1500m, the roll and pitch angles of the drone fluctuate due to airflow disturbances during flight, causing projection scaling and trapezoidal distortion in the original inspection images acquired by the imaging platform. This results in a nonlinear mapping relationship between the pixel step size of the photovoltaic module's frame in the image space and its physical spatial span. Furthermore, due to interference from cluttered signal components generated by surface vegetation and specular reflections on the module surface, the feature point extraction method based on global grayscale thresholds faces problems of low signal-to-noise ratio and geometric topological collapse when locating geometric anchor points. This leads to false alarms in the identification process of sub-centimeter-level defects due to distortion of the feature space ratio. To address the difficulties in feature extraction in the aforementioned scenario, the system utilizes the rectangular geometric rigidity of the photovoltaic array as a priori feature. The system identifies the gradient direction moment features of the corresponding planar array target edges in the original inspection image through a geometric reference feature extraction unit. It then uses a 3×3 Sobel operator to extract the gradient magnitude and direction of each pixel and calculates the probability density pj of the gradient direction distribution within a 5×5 neighborhood window to compute the local direction entropy E, where pj is the probability distribution of the j-th direction interval and E is the local direction entropy. Based on the direction entropy E, the system filters out texture interference terms with a direction entropy greater than 0.6 bits. It utilizes the high directional determinism of the component bounding box gradient direction locally to lock the structured linear feature flow and determines the vanishing point Vp in the image space based on the intersection geometric constraints of the structured linear feature flow. This method, which enhances the stability of the global geometric reference based on local statistical features, eliminates non-target noise components.

[0025] After obtaining the vanishing point Vp coordinates, the spatial dimension warping unit constructs a homography transformation matrix H using the mapping relationship between the vanishing point Vp and the preset rectangular reference. It then generates an orthophoto map by performing an inverse mapping process on the regions affected by perspective distortion in the original inspection image to the Euclidean coordinate plane. This process reshapes the complex nonlinear perspective transformation into linear alignment in the standard coordinate system. Utilizing the geometric prior of the photovoltaic module itself, it eliminates the dependence of the positioning process on external positioning hardware, enabling the system to reconstruct a metric-based two-dimensional analysis plane under conditions of uncertain pose in three-dimensional space. This results in a 1:1 linear correspondence between the pixel distance in the orthophoto map and the physical span of the photovoltaic module. When the system enters the target defect detection unit's processing stage, it constructs a logical grid coordinate system in the orthophoto feature map, calculates the deviation of each unitized target image relative to the local brightness background, and, since the previous steps have eliminated the projection noise caused by geometric deformation, the system locks the geometric topological reference of the confluence texture under the constraint of the logical grid coordinate system, and confirms the defect pixel coordinates within an accuracy range where the reprojection residual R is less than 1.5 pixels. Thus, it maintains sub-pixel-level recognition accuracy under the objective reality of uneven illumination distribution, and finally outputs an inspection result report containing defect type and geospatial coordinates, enabling the system to achieve adaptive closure of the detection logic through geometric rigid constraints in a complex physical environment.

[0026] Example 2: In a test environment containing 100 planar array targets, an imaging platform with a survey height of 10m was used to acquire raw survey images with a resolution of 3024×4032 pixels. The test environment was designed to simulate trapezoidal distortion caused by UAV motion by adjusting the flight attitude to introduce pitch angle deflections from 15° to 45°. During the acquisition process, Gaussian white noise with a signal-to-noise ratio of 20dB and isotropic reflection signals generated by simulated surface vegetation were superimposed to construct the original input data baseline. This test setup simulated high-altitude mountainous terrain. Airflow disturbances and complex background interference during inspection; In the geometric reference extraction stage, the sample group of this invention uses the geometric reference feature extraction unit to process the original inspection image, setting the directional entropy threshold Te to 0.6 to balance edge preservation integrity and non-target texture filtering rate. The measured data shows that the average directional entropy E of the background vegetation area is 0.82 bits, while the directional entropy E of the target edge area of ​​the planar array is stable at 0.28 bits. After filtering out texture interference terms with directional entropy greater than Te by directional entropy E, the purity of the structured linear feature flow extracted by the sample group of this invention is improved by 74.5% compared with the control group that only uses the standard Sobel operator. Due to the elimination of vegetation noise interference on edge detection, the reconstruction error of the vanishing point Vp is reduced from 12.6 pixels in the control group to 1.2 pixels. Where Te is the directional entropy threshold, E is the directional entropy, and Vp is the vanishing point. In the spatial dimension regularization stage, the system uses the mapping relationship between the vanishing point Vp and the preset rectangular reference to construct the homography transformation matrix H and generate an orthophoto feature map. By introducing a problem intensity gradient comparison system, at 15°, 30° and 45°, The geometric consistency was tested under three pitch angle deflection gradients. The experimental results show that as the deflection angle increases, the reprojection residual R is 0.43 pixels, 0.76 pixels and 1.34 pixels respectively, showing linear adaptive characteristics. When the deflection angle increases to the overrange boundary of 60°, the reprojection residual R increases to 5.67 pixels due to the intensified nonlinear collapse effect of projection imaging. This confirms the effectiveness of the homography transformation matrix H in correcting the deviation within the 15∘ to 45∘ window, and achieves physical scaling alignment between the orthophoto feature map and the standard Euclidean plane.

[0027] In the target defect detection stage, the system extracts the unitized target image corresponding to each grid index from the orthophoto feature map. Based on the deviation formula D=|Gavg−Gbg| / Gbg, abnormal states are determined. For non-uniform brightness conditions with local cloud occlusion, an abnormal threshold Td is set to 15%. The sample group of this invention locks a defect feature with a size of 0.85cm, and its positioning accuracy in the logical grid coordinate system converges to 1.1 pixels. In contrast, the control group, which did not perform spatial dimension regularization, is affected by projection scaling, and its unitized image morphological deviation D fluctuates randomly, resulting in a false alarm rate of 23.1% at the same threshold. This indicates that eliminating projection noise through geometric rigid constraints is a way to improve... The key to defect identification signal-to-noise ratio is shown in the figure, where D is the deviation, Gavg is the average gray value of the current grid, Gbg is the average gray value of the background, and Td is the anomaly threshold. Summarizing the measured indicators of each experimental group, this experiment demonstrates that by using the directional entropy E of the gradient directional moment feature to isolate signal interference, and in conjunction with the homography transformation matrix H based on the vanishing point Vp to reconstruct the feature space, the system reduces the physical spatial positioning error of defect identification from 5.23 cm in the control group to 0.94 cm without the need for external positioning hardware support. This proves that the deviation correction mechanism based on physical structure priors can stably produce the expected beneficial effects, achieving a reliable conversion from image space to a standard analysis space with metric attributes.

[0028] Example 3: This example combines Figures 1 to 2 This section describes a visual inspection system for photovoltaic module defects based on images acquired by drones, such as... Figure 1 As shown, the system's processing logic begins with the imaging platform (UAV), which performs the action of acquiring images containing planar array targets; the raw image signal receiving unit acquires the raw inspection image and transmits the image data stream to the next node; the geometric reference feature extraction unit filters out texture interference items based on the image data stream and determines the image vanishing point, outputting the vanishing point parameters; the spatial dimension regularization unit performs perspective distortion inverse mapping based on the vanishing point parameters to generate an orthophoto map; the target defect detection unit receives the orthophoto map and identifies morphological anomalies in the logical grid coordinate system; the system finally outputs the inspection results containing defect pixel coordinates and anomaly features.

[0029] like Figure 2As shown, the photovoltaic array's on-site environment includes photovoltaic modules as the planar array target and geometric reference components for calibration. These are projected onto a UAV imaging acquisition platform via an optical imaging path. The UAV imaging acquisition platform is equipped with a visual sensor for image acquisition and a flight control module for providing external orientation elements. The raw inspection images and external orientation element metadata generated by the UAV imaging acquisition platform are transmitted to a defect detection computing terminal. The defect detection computing terminal is internally divided into a core processing system and auxiliary and output modules. The core processing system includes a raw image signal receiving unit, a geometric reference feature extraction unit, a spatial dimension normalization unit, and a target defect detection unit. The auxiliary and output modules include a trend monitoring module and a coordinate transformation module. The defect detection computing terminal ultimately outputs an inspection result report.

[0030] Example 4: In the inspection scenario of photovoltaic arrays on the roof of an industrial plant, due to the nonlinear settlement of the building structure and the high-frequency flutter generated by the imaging platform during close-range operation, local geometric inconsistencies are generated in the original inspection image. Furthermore, due to the interference of secondary reflected light generated by the metal components of the roof, random texture components are mixed into the gradient signal of the edge of the planar array target. As a result, the fixed vanishing point solution method is prone to feature space reconstruction accuracy deviation when dealing with scenes with local disturbances, causing the center point of the unitized target image to drift in the logical grid coordinate system. In order to determine the sampling accuracy in the direction entropy calculation, the geometric reference feature extraction unit determines the value of the total number of direction intervals m in the execution procedure. According to the discrete angular resolution of the visual sensor, the system divides the angle space of [0,π] into 18 equally spaced sub-intervals, that is, the angular span of each direction interval is 10∘. 18 is selected as the value of m to ensure that the gradient response of the planar array target border in the vertical and horizontal directions can fall accurately into the independent feature bucket, thereby suppressing the isotropic background noise texture when calculating the direction entropy E.

[0031] During the construction of the homography transformation matrix H using spatial dimension regularization units, the system executes a parameter solution procedure based on geometric constraints. It extracts two mutually orthogonal vanishing points, Vp1 and Vp2, generated by the intersection of structured linear feature flows in the image coordinate system. Combining these points with the aspect ratio γ of the planar array targets identified in the original inspection image, a homogeneous linear equation system is constructed. The value of γ is determined based on the physical specifications of the photovoltaic modules and is set to 1.65. The system uses the vanishing point coordinates as input and, under the orthogonal constraint of the projection transformation, performs singular value decomposition on the coefficient matrix to determine the transformation... The parameters are calculated to obtain the 8 degrees of freedom components of the homography transformation matrix H, which maps the quadrilateral region in the image space to the standard rectangular region in the orthophoto feature map, realizing the transformation from the perceptual space to the metric space. Here, H is the homography transformation matrix, Vp1 and Vp2 are the vanishing points, and γ is the aspect ratio. To address the stability of the target defect detection unit under varying light intensity, the system adopts an adaptive threshold calibration procedure based on background statistical features. It calculates the average gray value Gaavg and gray standard deviation σ of the neighboring pixels in the logical grid coordinate system of the unitized target image, and determines the dynamic anomaly threshold according to the following formula. Where Td is the anomaly threshold, k is the sensitivity coefficient (3.5 under this condition), σ is the grayscale standard deviation, and Gavg is the average grayscale value. When the deviation D is greater than the dynamically generated Td, the system determines that the corresponding pixel area has abnormal morphological features. This method of determining the judgment criteria through local statistical distribution enables the system to stabilize the reprojection residual R within 0.8 pixels when the background brightness fluctuates by more than 25%.

[0032] Example 5: In response to the increased computation time of the transformation matrix caused by the initial offset of the visual sensor, the system performs a pre-calibration procedure before the inspection starts. The imaging platform is aligned with a planar geometric reference component at a preset position on the ground. The geometric reference feature extraction unit calculates the directional entropy E of the edge of the planar geometric reference component, where E is the directional entropy. The flight attitude angle output by the imaging platform at this time is compared with its theoretical projection angle to obtain the static deviation component. The static deviation component is used to perform real-time correction on the external orientation element metadata obtained by the original image signal receiving unit, providing physically aligned seed parameters for the construction of the homography transformation matrix H.

[0033] During the deployment of components of different specifications, in order to eliminate the feature space ratio distortion caused by the aspect ratio γ setting deviation, the system uses the target defect detection unit to perform a baseline alignment procedure before constructing the logical grid coordinate system. It extracts the photovoltaic array feature flow containing k groups of units in the original inspection image, and performs a consistency back calculation between the extracted vanishing point Vp coordinates and the currently stored aspect ratio γ, where Vp is the vanishing point, γ is the aspect ratio, and k is the number of unit groups. The mapping deviation of the homography transformation matrix H is verified by calculating the second difference variance of the pixel span of adjacent components, and the matching of pixel displacement and physical span is realized in the orthophoto feature map.

[0034] Example 6: When the focal length f of the imaging system shifts due to environmental temperature differences, causing the vanishing point Vp to deviate from the principal point in the image coordinate system, the system performs a pre-calibration procedure before the inspection operation starts. The imaging platform is aligned with a rectangular reference component of known physical span. The geometric reference feature extraction unit extracts the gradient direction moment features of the rectangular reference component's edge and calculates the direction entropy E distribution under different pitch angle gradients, where f is the focal length and E is the direction entropy. The system extracts the maximum inter-class variance between the target edge component and the background texture component from the statistical distribution of the direction entropy E to determine the current calibration value of the direction entropy threshold Te. Based on the projection constraint of the vanishing point Vp, the system calculates the real-time offset of the focal length f. The calibration value and the real-time offset are used to update the initial weight components of the homography transformation matrix H, achieving physical compensation for fluctuations in the internal parameters of the imaging unit, where Vp is the vanishing point and H is the homography transformation matrix.

[0035] In cases involving geographic alignment of inspection data, the coordinate transformation module establishes a mapping relationship between the logical grid coordinates in the orthophoto map and the geospatial coordinates of the Global Navigation Satellite System at the imaging time. The system selects four physical control points within the photovoltaic power station, obtains the pixel positions of the physical control points in the image coordinate system through spatial dimension warping units, and calculates the spatial scale scaling factor λ, where λ is the scaling factor. The system determines the deflection angle component θ of the logical grid coordinate system relative to geographic north and performs a rotation and translation transformation to map the defect pixel coordinates to the geospatial coordinate system, where θ is the deflection angle component. This procedure eliminates the cumulative translational deviation caused by UAV movement, enabling the inspection result report to have geographic anchoring attributes. When the system faces the need for monitoring the damage evolution of planar array targets, the trend monitoring module executes a time-series-based feature evaluation procedure. The system acquires morphological anomaly feature data extracted from the same unitized target image in two consecutive inspection cycles t1 and t2, calculates the area components A1 and A2 of the abnormal pixel region, and determines the evolution rate v according to the following formula: v = A2 − A1t2 − t1, where v is the evolution rate, A1 is the area at time t1, and A2 is the area at time t2. The system sets the degradation threshold Tv to 0.005 square centimeters per hour according to the component degradation mechanism. When the evolution rate v exceeds Tv, the target defect detection unit triggers a fault warning signal and outputs geospatial coordinates.

[0036] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0037] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A visual inspection system for photovoltaic module defects based on images acquired by unmanned aerial vehicles (UAVs), characterized in that, The system includes: The raw image signal receiving unit is used to acquire the raw inspection images containing planar array targets collected by the imaging platform; The geometric reference feature extraction unit is used to identify the gradient direction moment features of the corresponding planar array target edges in the original inspection image, calculate the direction entropy of the gradient direction moment features in a preset neighborhood, lock the structured linear feature flow by filtering out non-target texture interference terms with direction entropy less than a preset threshold, and determine the vanishing point of the image space based on the intersection geometric constraints of the structured linear feature flow. The spatial dimension regularization unit is used to construct a homography transformation matrix based on the mapping relationship between the vanishing point and the preset rectangular reference, and to perform inverse mapping processing on the perspective distortion area in the original inspection image to the Euclidean coordinate plane to generate an orthophoto feature map with geometric rigidity. The target defect detection unit is used to extract the unitized target image corresponding to each grid index in the logical grid coordinate system constructed by the orthophoto feature map, calculate the deviation of the unitized target image relative to the local brightness background, determine the defect pixel coordinates based on the pixel area where the deviation exceeds the preset abnormal threshold, and identify the morphological abnormal features corresponding to the defect pixel coordinates.

2. The photovoltaic module defect visual inspection system based on UAV image acquisition according to claim 1, characterized in that, The geometric reference feature extraction unit is also used to calculate the gradient energy centroid of the structured linear feature flow in the edge normal direction to refine the positioning coordinates of the gradient direction moment feature to the sub-pixel level, and to perform linear regression processing on the sub-pixel level positioning points using the least squares fitting algorithm to generate calibrated feature line parameters. The geometric reference feature extraction unit corrects the coordinate position of the vanishing point based on the calibrated feature line parameters.

3. The photovoltaic module defect visual inspection system based on UAV image acquisition according to claim 1, characterized in that, When determining the vanishing point, the geometric reference feature extraction unit is also used to perform the following steps: using the vanishing point as the origin of radiation, a virtual ray cluster is generated in the space of the original inspection image; the virtual ray cluster and the gradient direction moment feature are subjected to geometric consensus verification to remove linear noise components that do not conform to the constraints of the origin of radiation.

4. The photovoltaic module defect visual inspection system based on UAV image acquisition according to claim 1, characterized in that, In the process of generating orthophoto maps, spatial dimension regularization units are also used to construct geometric virtual constraints by utilizing the translational symmetry of planar array targets. By comparing the pixel span consistency of adjacent unitized target images in the orthophoto map, feedback correction is performed on the mapping parameters of the homography transformation matrix.

5. A photovoltaic module defect visual inspection system based on UAV-acquired images according to claim 4, characterized in that, The logic for feedback correction performed by the spatial dimension warping unit includes: calculating the reprojection residual R for pixel span consistency, using the following formula: Where xi is the preset theoretical center x-coordinate of the i-th unitized target image in the standard Euclidean coordinate plane, and yi is the corresponding preset theoretical center y-coordinate; xi is the actual center x-coordinate in the pixel space after homography transformation matrix mapping, and yi is the corresponding actual center y-coordinate; n is the total number of unitized target images; when the reprojection residual R is greater than the preset deviation threshold, the local coordinate perturbation algorithm is executed to compensate for the geometric distortion caused by the non-coplanarity of the planar array targets.

6. A photovoltaic module defect visual inspection system based on UAV-acquired images according to claim 1, characterized in that, When identifying morphological abnormalities, the target defect detection unit also performs the following steps: extracting the confluence texture distribution features inside the unitized target image; determining the geometric topological reference of the confluence texture distribution features based on the logical grid coordinate system; and determining the presence of structural damage inside the unitized target image by identifying regions that deviate in shape or experience sudden changes in brightness relative to the geometric topological reference.

7. A photovoltaic module defect visual inspection system based on UAV-acquired images according to claim 1, characterized in that, The target defect detection unit is also used to: calculate the average gray value of each grid cell in the statistical logical grid coordinate system, calculate the ratio of the average gray value to the preset global background brightness, and perform gray-scale mapping enhancement on the orthophoto feature map based on the ratio.

8. A photovoltaic module defect visual inspection system based on UAV-acquired images according to claim 1, characterized in that, The system also includes a trend monitoring module, which records data on the morphological anomalies of the same unitized target image in different inspection cycles, calculates the data evolution rate over time, and outputs a fault warning signal for the corresponding unitized target image when the evolution rate exceeds a preset degradation threshold.

9. A photovoltaic module defect visual inspection system based on UAV-acquired images according to claim 1, characterized in that, The raw image signal receiving unit is also used to: receive the external orientation element metadata of the imaging platform at the acquisition time; determine the initial spatial projection angle of the optical axis of the imaging platform relative to the planar array target based on the external orientation element metadata, and use the initial spatial projection angle as the seed parameter for constructing the homography transformation matrix by the spatial dimension warping unit.

10. A photovoltaic module defect visual inspection system based on UAV-acquired images according to claim 1, characterized in that, The system also includes a coordinate transformation module, which maps the defect pixel coordinates determined by the target defect detection unit to the geospatial coordinate system of the global navigation satellite system, and generates an inspection result report containing the defect type and its geospatial coordinates.

Images