Visible light and infrared image combined photovoltaic defect detection method based on unmanned aerial vehicle

Abstract

The invention relates to the technical field of image detection, in particular to a visible light and infrared image combined photovoltaic defect detection method based on an unmanned aerial vehicle, which comprises the following steps of: determining a photovoltaic power station detection area, carrying out synchronous aerial photography by using a visible light camera carried by the unmanned aerial vehicle and an infrared thermal imager, and carrying out local division to extract texture and temperature characteristics; the screening area performs fitting affine parameter generation on an extraction center point, corrects an infrared image to extract contour lines and texture change features, and screens a defect area to extract a positioning coordinate set; according to the method, through synchronously screening generated visible light and infrared image pairs, the resolution of an abnormal region is enhanced, a complex background and an abnormal target are accurately separated, fine-grained adaptive registration is realized through central point extraction and affine parameter derivation, and defect region characteristics are verified bidirectionally through a temperature contour closing proportion and a texture density variable quantity; thermal features and texture features are fused in the defect screening process, the recognition capability of weak anomalies is enhanced, and the defect positioning accuracy is improved.

Description

Technical Field

[0001] The present invention relates to the field of image detection technology, and in particular to a photovoltaic defect detection method based on a combination of visible light and infrared images of an unmanned aerial vehicle (UAV). Background Art

[0002] The field of image detection technology includes methods and means for extracting key information and identifying the characteristics of target objects by processing and analyzing image data. The core content includes steps such as image acquisition, feature extraction, image segmentation, image recognition and target detection. It is widely used in multiple industries such as industrial inspection, medical diagnosis, security monitoring, and agricultural monitoring. By analyzing the physical characteristics or abnormal conditions of the target in the image, the state changes of the object can be judged and classified. It usually combines visible light image processing and infrared imaging technology, and uses the complementarity of image information in different bands to improve the accuracy and reliability of detection.

[0003] Among them, the photovoltaic defect detection method based on the combined visible light and infrared images of drones refers to the use of drones equipped with visible light imaging equipment and infrared imaging equipment to synchronously obtain visible light images and infrared thermal imaging images of the surface of photovoltaic modules, align different source images through image registration, analyze surface damage, pollution, occlusion and other structural abnormalities in visible light images through texture feature extraction, process local overheating, hot spot effect, local short circuit and other thermal abnormalities in infrared images through temperature feature analysis, and conduct comprehensive analysis of the extracted multi-source features in combination with image fusion to complete the identification and marking of photovoltaic module defect types and distribution locations.

[0004] In the existing photovoltaic defect detection process, the data acquisition stage mostly relies on the independent operation of each imaging device, which fails to effectively avoid the source image alignment deviation caused by time synchronization errors. The local feature extraction process relies solely on texture or temperature indicators, and cannot form multi-dimensional composite feature support, resulting in insufficient resolution of abnormal areas under complex backgrounds or weak features. Traditional alignment mostly uses a global affine model, which is prone to cumulative offset in local subtle deformation and occlusion scenarios, resulting in amplified subsequent detection errors. Thermal anomaly detection methods generally rely on a single temperature mutation feature, ignoring the joint analysis of temperature spatial distribution characteristics and local structural features of the image. When faced with slight thermal anomalies or structural texture variation areas, the detection accuracy is insufficient, and there are problems of decreased defect detection rate and increased false detection rate. For example, when a slight hot spot effect or pollution occlusion appears on the surface of the component, it is often missed due to the lack of prominent features, or false alarms are caused by background changes. The overall detection performance is relatively sensitive to environmental interference, and there are obvious deficiencies in engineering adaptability and reliability. Summary of the Invention

[0005] The purpose of the present invention is to solve the shortcomings of the existing technology and propose a photovoltaic defect detection method based on the combination of visible light and infrared images of drones.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a photovoltaic defect detection method based on the combination of visible light and infrared images of drones, comprising the following steps:

[0007] S1: Determine the inspection area of ​​the photovoltaic power station and use a drone equipped with a visible light camera and an infrared thermal imager to perform synchronized aerial photography. The image groups are filtered according to the time synchronization tolerance threshold to generate synchronized visible light and infrared image pairs.

[0008] S2: performing local segmentation based on the synchronized visible light and infrared image pair, calculating local texture statistics of the visible light image and local temperature statistical differences of the infrared image, and establishing a local texture and temperature change feature matrix;

[0009] S3: According to the local texture and temperature change feature matrix, screen local region pairs and extract a center point set, fit the center point set to derive local affine transformation parameters, perform local block-level transformation of the infrared image, and generate a local affine registration matrix;

[0010] S4: Correcting the infrared image using the local affine registration matrix, extracting temperature contours and calculating the closed contour ratio, and simultaneously calculating the relative change in texture density in the corresponding local area of ​​the visible light image to establish thermal closure and texture density change features;

[0011] S5: According to the thermal sealing and texture density change characteristics, the defect local area is screened, the coordinates of the center of the area are extracted, and the photovoltaic module defect location coordinate set is obtained by summarizing.

[0012] As a further solution of the present invention, the synchronized visible light and infrared image pair includes image timestamp synchronization difference, image spatial alignment error, and synchronization tolerance judgment result; the local texture and temperature change feature matrix includes local texture feature index, local temperature statistical difference value, and local area division number; the local affine registration matrix includes local affine transformation parameters, center point set coordinates, and feature matching similarity score; the thermal closure and texture density change characteristics include temperature contour line closure ratio, texture density change, and local brightness gradient statistics; the photovoltaic module defect positioning coordinate set includes defect area center coordinates, defect detection confidence, and defect area number.

[0013] As a further solution of the present invention, the step of acquiring the synchronized visible light and infrared image pair is specifically as follows:

[0014] S111: Obtain the PV power station inspection operation area, set the equipment system clock preliminary calibration parameters and record the flight start timestamp based on the requirements of synchronized aerial photography with the UAV-mounted visible light camera and infrared thermal imager, and generate a preliminary equipment time synchronization reference value;

[0015] S112: Based on the preliminary device time synchronization reference value, detect the acquisition timestamp of each group of images, establish a time record sequence for each visible light image and infrared image, calculate the time difference between the visible light image and the infrared image based on the GNSS timing synchronization threshold, select the image group that meets the synchronization conditions, and obtain a time-synchronized image sequence;

[0016] S113: performing image alignment processing according to the time-synchronized image sequence and the time-series index value, filtering out image groups that do not meet the synchronization conditions, obtaining one-to-one corresponding combinations of visible light images and infrared images, and generating synchronized visible light and infrared image pairs.

[0017] As a further solution of the present invention, the steps of obtaining the local texture and temperature change feature matrix are specifically as follows:

[0018] S211: Acquire the synchronized visible light and infrared image pairs, perform region division processing based on standard size parameters of photovoltaic modules, divide each image into multiple local regions, and generate a local image region unit set;

[0019] S212: Based on the local image region unit set, calculating the texture statistics of each visible light local region, establishing preliminary texture values ​​and variance values ​​based on the gray level co-occurrence matrix contrast and the local binary pattern variance, performing difference synthesis on the gray level characteristics and structural characteristics, calculating the local texture statistical synthesis value, completing the local texture feature quantization based on the difference in the synthesized texture characteristics, and generating a local texture statistics set;

[0020] S213: According to the local texture statistics set and the infrared image areas in the local image area unit set, based on the temperature distribution data of each area, the regional temperature standard deviation and range value are calculated respectively, the temperature discreteness is normalized, the local texture statistics are combined with the temperature difference, and a local texture and temperature change feature matrix is ​​established.

[0021] As a further solution of the present invention, the steps of obtaining the local affine registration matrix are specifically as follows:

[0022] S311: Obtain the local texture and temperature change feature matrix, calculate the Manhattan distance of the feature vector of each local area based on the texture statistics and temperature statistics data of each local area in the matrix, screen according to the similarity matching threshold set in the photovoltaic panel surface feature similarity experiment, and generate a feature matching area pair set;

[0023] S312: Based on the feature matching region pair set, extract the center point position of each pair of regions, establish a center point set based on all the center points, perform coordinate fitting on the center point set, derive a local affine transformation initial parameter matrix, calculate the local affine transformation comprehensive difference value, optimize the initial parameter set, and obtain a local affine transformation parameter matrix;

[0024] S313: performing an affine transformation operation on each local block according to the local affine transformation parameter matrix and based on the local block index of the infrared image, establishing a transformed regional block set, and generating a local affine registration matrix.

[0025] As a further solution of the present invention, the steps of obtaining the thermal sealing and texture density change characteristics are specifically as follows:

[0026] S411: Acquire the infrared image corrected by the local affine registration matrix, set a temperature threshold starting value, extract the temperature area greater than the temperature threshold to generate a temperature contour standard, establish a corresponding contour set, and generate a temperature contour extraction result;

[0027] S412: Based on the temperature contour extraction results, detecting the head-to-tail connection relationship of each temperature contour, counting the number of contour lines that meet the head-to-tail closure condition and the total number of extracted contour lines, calculating a closed contour feature metric value, and performing a local thermal closure area assessment based on the closed contour feature metric value to obtain a thermal closure feature index set;

[0028] S413: Based on the thermal closure feature index set and the corresponding local area of ​​the visible light image, the brightness gradient amplitude statistics are extracted, the grayscale co-occurrence matrix energy ratio of the central area and the peripheral area is calculated, the texture density change is counted, and the thermal closure and texture density change characteristics are established by combining the thermal closure feature index set and the texture density change.

[0029] As a further solution of the present invention, the steps for obtaining the photovoltaic module defect location coordinate set are specifically as follows:

[0030] S511: Obtain the thermal closure and texture density variation characteristics, detect whether the closed contour feature metric value of each region is greater than a set closed contour ratio threshold, and directly compare the closed contour metric value of each local region with the set closed contour ratio threshold to generate a closed ratio screening region set;

[0031] S512: Filter the region set based on the closure ratio, detect the texture density change corresponding to each region, compare the texture density change of the local region with the critical value, retain the region where the change is greater than the critical value as the valid region, and simultaneously perform a joint evaluation of the thermal closure feature value, closure ratio, center grayscale energy value, and edge grayscale energy value to comprehensively determine the regional characteristics and obtain the defect region screening results;

[0032] S513: extracting the center coordinates of each local area that meets the screening conditions according to the defect area screening result, and summarizing and arranging all the center coordinates into a set, and finally generating a photovoltaic module defect location coordinate set.

[0033] Compared with the prior art, the advantages and positive effects of the present invention are:

[0034] In the present invention, the visible light and infrared image pairs generated by synchronous screening are used to improve data consistency, local division is used to extract texture and temperature statistical features, and the resolution of abnormal areas is enhanced. A composite representation is constructed based on the feature matrix to accurately separate complex backgrounds and abnormal targets. Center point extraction and affine parameter derivation are used to achieve fine-grained adaptive registration, maintaining stable local registration accuracy. The temperature contour closure ratio and texture density change are used to bidirectionally verify the defect area characteristics to ensure detection constraints and accuracy. Thermal features and texture features are integrated in the defect screening process to enhance the recognition ability of weak anomalies such as small-scale damage, local overheating, and pollution occlusion. The defect positioning accuracy is improved, and false detection and missed detection phenomena are significantly reduced. The detection adaptability and environmental robustness are simultaneously enhanced. The overall operation has stronger resistance to illumination changes, background complexity, and changes in component surface details, effectively improving the accuracy and stability of photovoltaic defect detection in multiple scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 It is a flow chart of the main steps of the present invention;

[0036] Figure 2 This is a flow chart of step S1 of the present invention;

[0037] Figure 3 This is a flow chart of step S2 of the present invention;

[0038] Figure 4 This is a flow chart of step S3 of the present invention;

[0039] Figure 5 This is a flow chart of step S4 of the present invention;

[0040] Figure 6 This is a flow chart of step S5 of the present invention. DETAILED DESCRIPTION

[0041] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0042] In the description of the present invention, it should be understood that the terms "length," "width," "up," "down," "front," "back," "left," "right," "vertical," "horizontal," "top," "bottom," "inside," "outside," and the like, indicating positions or relationships, are based on the positions or relationships shown in the accompanying drawings and are intended only to facilitate the description of the present invention and simplify the description. They do not indicate or imply that the devices or elements referred to must have a specific orientation, be constructed, or operate in a specific orientation. Therefore, they should not be construed as limiting the present invention. Furthermore, in the description of the present invention, "plurality" means two or more, unless otherwise expressly and specifically defined.

[0043] See also Figure 1 The photovoltaic defect detection method based on the combined visible light and infrared images of UAV includes the following steps:

[0044] S1: Determine the inspection area for the PV power plant and use a drone equipped with a visible light camera and an infrared thermal imager to perform synchronized aerial photography. The image acquisition timestamps are detected and compared with the device clock offsets. The image groups are filtered based on the time synchronization tolerance threshold (the maximum time difference allowed between the acquisition times of the two devices, usually set based on the drone's flight speed and image resolution, for example, a threshold of ≤0.1 seconds when using GNSS timing synchronization) to generate synchronized visible light and infrared image pairs.

[0045] S2: Based on the synchronized visible light and infrared image pairs, local division is performed according to the standard size of the photovoltaic module. The local texture statistics of the visible light image are calculated (using the gray level co-occurrence matrix contrast or local binary pattern variance). At the same time, the local temperature statistical difference of the infrared image is calculated (obtained by calculating the regional temperature standard deviation or range). The local texture and temperature change feature matrix is ​​established.

[0046] S3: Based on the texture and temperature statistics of each local area in the local texture and temperature change feature matrix, perform absolute difference calculation (using Manhattan distance to measure the absolute difference between two eigenvectors) and screen local area pairs based on the similarity matching threshold (determined based on the similarity experiment of photovoltaic panel surface features), extract the center point set of the screened area, fit the center point set to derive the local affine transformation parameters, use the parameters to perform local block-level transformation of the infrared image, and generate a local affine registration matrix;

[0047] S4: Using the infrared image corrected by the local affine registration matrix, set the temperature contour generation standard according to the temperature threshold starting value (set to an initial value 5°C higher than the ambient temperature) and extract the temperature contours. Detect the head-to-tail connection relationship of each contour and calculate the closed contour ratio (the percentage of closed contours to the total extracted contours). Synchronously, extract the brightness gradient amplitude statistics and calculate the relative change in texture density (calculated by the ratio of the grayscale co-occurrence matrix energy values ​​of the central area and the surrounding area) in the corresponding local area of ​​the visible light image to establish the thermal closure and texture density change characteristics.

[0048] S5: Based on the characteristics of thermal sealing and texture density change, screen the local areas where the closing ratio is greater than the closing contour ratio threshold and the texture density decrease is greater than the texture density change critical value, extract the center coordinates of the qualified areas, and summarize them to obtain the photovoltaic module defect location coordinate set.

[0049] The synchronized visible light and infrared image pairs include the image timestamp synchronization difference, image spatial alignment error, and synchronization tolerance judgment result. The local texture and temperature change feature matrix includes local texture feature indicators, local temperature statistical difference values, and local area division numbers. The local affine registration matrix includes local affine transformation parameters, center point set coordinates, and feature matching similarity scores. The thermal closure and texture density change features include the temperature contour line closure ratio, texture density change, and local brightness gradient statistics. The photovoltaic module defect positioning coordinate set includes the defect area center coordinates, defect detection confidence, and defect area number.

[0050] See also Figure 2 , the specific steps of S1 are:

[0051] S111: Obtain the PV power station inspection operation area, set the equipment system clock preliminary calibration parameters and record the flight start timestamp based on the requirements of synchronized aerial photography with the UAV-mounted visible light camera and infrared thermal imager, and generate a preliminary equipment time synchronization reference value;

[0052] Obtain the inspection operation area of ​​the photovoltaic power station, call the power station GIS map data according to the aerial photography mission planning system, select the inspection area with complete component distribution, no obstruction, and flat terrain, and divide the aerial photography planning points into 10-meter grid units within the selected area. The flight path of each aerial photography point is based on the maximum shooting radius covered by the visible light camera and infrared thermal imager. At the same time, set the flight starting point and perform preliminary calibration of the equipment system clock. The specific operation is to connect the visible light camera and infrared thermal imager to the GNSS synchronization module respectively, read the synchronization module timing signal, transmit the timing signal to the built-in clocks of the two sensor modules, and use the calibrated clock time as the flight mission start timestamp record. For example, in actual operation, if the synchronization module timing signal reads Beijing time 08:00:00.000, the visible light camera and infrared thermal imager will be at the same time. The system clocks of all instruments and equipment are set to 08:00:00.000, and the initial timestamp data of the aerial photography flight starting point is recorded. The aerial photography starting point position is set to 45230 meters in X coordinate and 21450 meters in Y coordinate. The take-off time is recorded by the flight controller as 08:00:05.120, forming the preliminary equipment time synchronization reference value. During the calibration process, it is necessary to judge the clock synchronization error. If the error is less than 0.05 seconds, the synchronization is considered valid. Otherwise, recalibration is required. In an actual test, the time after calibration of the visible light camera is 08:00:00.002, and the time of the infrared thermal imager is 08:00:00.003. After comparison, the difference is 0.001 seconds, which is less than the allowable error threshold of 0.05 seconds. Therefore, the synchronization is considered successful and the preliminary equipment time synchronization reference value is recorded as 08:00:00.000.

[0053] S112: Based on the preliminary equipment time synchronization reference value, detect the acquisition timestamp of each group of images, establish time record sequences for visible light and infrared images respectively, calculate the time difference between the visible light image and the infrared image based on the GNSS timing synchronization threshold ≤ 0.1 second, select the image group that meets the synchronization conditions, and obtain the time synchronized image sequence;

[0054] Based on the preliminary equipment time synchronization benchmark value, the acquisition timestamp of each group of images is detected, and the UAV flight control data recording module is called to extract the specific time of each shutter release during the aerial photography process. At the same time, the acquisition time of the visible light image and the infrared image are obtained respectively, and the time record sequence is established respectively. The visible light image time sequence is set to {08:00:05.120, 08:00:06.720, 08:00:08.320}, and the infrared image time sequence is {08:00:05.118, 08:00:06.719, 08:00:08.317}. According to the GNSS timing synchronization tolerance threshold set to 0.1 second, the time difference of each pair of images is calculated. For example, the time difference of the first group of images is |08:00:05.120-08:00:05.118|=0.002 seconds, which is small. The image groups are selected as qualified image groups. All image groups are judged in turn. If the image time difference is within the threshold, they are retained, otherwise they are eliminated. During the screening process, if an image group with a time difference of 0.101 seconds appears, it is considered to be out of limit and eliminated. For example, in an actual detection, the acquisition time of an infrared image is 08:00:06.620, and the corresponding visible light image time is 08:00:06.720. The difference is 0.100 seconds, which is within the acceptable range. The image group is retained. After final screening, the time-synchronized image sequence {(08:00:05.120, 08:00:05.118), (08:00:06.720, 08:00:06.719), (08:00:08.320, 08:00:08.317)} is obtained, as shown in Table 1.

[0055] Table 1 Time synchronization image acquisition record table

[0056] Visible light image acquisition time (UTC time) Infrared image acquisition time (UTC time) Time difference (seconds) 08：00：05.120 08：00：05.118 0.002 08：00：06.720 08：00：06.719 0.001 08：00：08.320 08：00：08.317 0.003

[0057] Table 1 lists the time-synchronized image acquisition data that have been screened and retained. It can be seen that the time difference is less than 0.1 second, and the synchronization accuracy meets the requirements of aerial photography detection.

[0058] S113: performing image alignment processing based on the time-synchronized image sequence and the time-series index value, filtering out image groups that do not meet the synchronization conditions, obtaining a one-to-one corresponding combination of visible light images and infrared images, and generating a synchronized visible light and infrared image pair;

[0059] According to the time synchronization image sequence, image alignment processing is performed according to the time series index value. First, the file index numbers of the visible light image and the corresponding infrared image in each group of synchronized images are extracted. For example, the visible light image numbers are IMG_0001, IMG_0002, IMG_0003, and the infrared image numbers are IR_0001, IR_0002, IR_0003. According to the time synchronization sequence, the corresponding matches are made in sequence, and the image groups that do not meet the synchronization conditions are filtered out. That is, in the actual processing process, if it is found that the synchronization time difference of a group of images exceeds the threshold, the group of data is eliminated. For example, if the visible light image IMG is found in a certain detection, The time of image _0004 is 08:00:10.520, while the time of infrared image IR_0004 is 08:00:10.630. The time difference is 0.110 seconds, which exceeds the 0.1 second threshold. Therefore, it is excluded from the synchronization pair. Finally, only the synchronization pair is retained to form a combination relationship between the visible light image and the infrared image. The pairing list is established as shown below, namely (IMG_0001, IR_0001), (IMG_0002, IR_0002), and (IMG_0003, IR_0003). By generating synchronized visible light and infrared image pairs, subsequent image feature extraction and defect detection can be carried out.

[0060] See also Figure 3 , the specific steps of S2 are:

[0061] S211: Acquire a pair of synchronized visible light and infrared images, perform region division processing based on standard size parameters of photovoltaic modules, divide each image into multiple local regions, and generate a local image region unit set;

[0062] Acquire synchronized visible light and infrared image pairs. In implementation, first select a standard photovoltaic power station area, where the component size is uniformly 1.992 meters × 0.992 meters. Based on the size data, a coordinate grid segmentation method is used. Each set of images is divided equally into local sub-areas in the horizontal and vertical directions. For example, if the resolution of the drone image is 4096 × 2160 pixels, each local area is divided into approximately 204 × 108 pixels. When dividing, read the starting coordinate point of the upper left corner of the image, and extract the sub-area coordinate boundary at a fixed interval. Each local sub-area is numbered as a number index, such as R1, R2... During the acquisition process, image sub-block extraction is performed for each sub-area. , through programming to control the sub-block acquisition order and number, to avoid omission or repeated acquisition, for abnormal pixel areas (such as image edges) to perform elimination, the specific elimination condition is the area where the sub-block size is less than 80% of the standard size, set the elimination threshold area to 0.8 times the standard area, the calculated standard area is 204×108=22032 pixels, the elimination area threshold is 17625.6 pixels, retain the sub-area larger than this area, after completing each local division of an image, the area number, coordinate index, pixel size and other information are stored in the local image area unit set, for example, an image can generate 60 local image units, the area information is stored in an array structure, an example is shown in the following table:

[0063] Table 2 Example table of local image region unit set

[0064] Area Number Upper left corner coordinates Lower right corner coordinates Area width Area Height R1 (0，0) (204，108) 204 pixels 108 pixels R2 (204，0) (408，108) 204 pixels 108 pixels R3 (408，0) (612，108) 204 pixels 108 pixels

[0065] As shown in Table 2, each sub-region is divided according to a standard size, and the corresponding pixel position and size are recorded to establish a local image region unit set.

[0066] S212: Based on the local image region unit set, the texture statistics of each visible light local region are calculated. The preliminary texture value and variance value are established based on the gray level co-occurrence matrix contrast and the local binary pattern variance, respectively. The gray level characteristics and structural characteristics are synthesized using the formula:

[0067]

[0068] Calculate the local texture statistical comprehensive value L f , based on the comprehensive texture characteristic differences, local texture feature quantization is completed to generate a local texture statistics set, where C i Represents the gray level co-occurrence matrix contrast of the i-th local area, Represents the mean contrast value of the gray level co-occurrence matrix of all local regions, V i represents the local binary pattern variance of the i-th local region, represents the mean of the variance of the local binary patterns of all local regions, and m represents the number of local region units;

[0069] Based on the local image area unit set, the visible light image data is extracted and the contrast value of each local area is calculated by the gray level co-occurrence matrix. i The calculation formula is the sum of the products of the grayscale differences between each pixel pair, and then normalized. For a single region R1, the pixel grayscale matrix is ​​extracted, and the initial contrast value C1 = 0.37 is calculated by statistically analyzing the grayscale differences. Statistics are completed for all local regions in turn. At the same time, the local binary pattern features of the region are extracted and the local variance V is calculated. i The local binary pattern is obtained by comparing the grayscale of each pixel with the surrounding pixels to obtain a binary code, and the variance of all the code values ​​in the area is calculated. The local binary pattern variance V1 of the example area R_1 is 0.58. According to the C of all local areas i 、V i Calculate the overall mean Then, based on the formula, the combined operation of each participating item is performed. The specific process is as follows: first, each C i and mean Sum the squares of the differences and take the square root. Assuming there are m = 60 sub-regions, the actual summation result is The first item is Secondly, the mean and Take the absolute value |0.42-0.54|=0.12, and then i Take the square root and average, example The sum of the three items is L f =0.048+0.12+0.76=0.928. The final operation obtains the comprehensive value of local texture statistics, and the L_f value sets of all local areas are sorted out to generate a local texture statistics set.

[0070] The local texture statistical composite value represents a comprehensive quantitative measure of the grayscale contrast variation, overall brightness structure deviation, and detail texture discreteness within a single region. Its composite value is constructed using three independent dimensions: standard deviation, absolute difference, and square root mean. The standard deviation reflects the fluctuation of the local region's grayscale contrast and represents the uniformity of the light-dark contrast variation within the region. The absolute difference measures the overall difference between the local region's average contrast and the local binary pattern variance, indicating the consistency or deviation between the brightness characteristics and detail structure variations. The square root mean reflects the discreteness and richness of detail variations within the local region and reflects the complexity of the regional texture. Therefore, the local texture statistical composite value can uniformly measure the changing trends of the light intensity distribution and microstructural characteristics within the local region. Higher values ​​indicate more significant grayscale contrast fluctuations, greater brightness structure deviations, and richer and more dispersed detail textures within the local region. Lower values ​​indicate stable light intensity variations and uniform detail patterns within the region. This approach is suitable for improving the sensitivity of local features in local anomaly detection and micro-defect recognition tasks in photovoltaic module images.

[0071] The formula is integrated by summing the standard deviation of the contrast of the local gray-level co-occurrence matrix, the absolute difference between the contrast mean and the mean of the local binary pattern variance, and the mean of the square root of the local binary pattern variance. First, It is used to measure the fluctuation degree of the grayscale contrast of each local area and the overall mean, reflecting the consistency of the contrast changes between regions. Therefore, the standard deviation is calculated by square difference and square root to enhance the sensitivity to changes in abnormal areas. By calculating the absolute value difference between the overall grayscale contrast mean and the local binary pattern variance mean, the overall deviation between the overall brightness structure and the detail pattern is captured to avoid the positive and negative differences from offsetting the information. Finally, The square root of the local binary pattern variance in the local area is taken and the mean is calculated. The purpose is to weaken the influence of the local maximum variance value on the overall statistics, while retaining the wide distribution characteristics of local texture details. The above three items comprehensively measure the amplitude of the internal feature changes in the region from three aspects: local contrast fluctuation, overall structural difference, and local texture discreteness. Therefore, addition operation is used for integration to form a unified measurement index. The overall logic is to accurately characterize the local texture change trend and detail complexity through the joint accumulation of local fluctuation, overall offset, and local detail diffusion.

[0072] S213: Based on the local texture statistics set and the infrared image regions in the local image region unit set, based on the temperature distribution data of each region, the regional temperature standard deviation and range value are calculated respectively, the temperature discreteness is normalized, and the local texture statistics and the temperature difference are combined to establish a local texture and temperature change feature matrix;

[0073] According to the local texture statistics set and the infrared image data in the local image area unit set, the temperature pixel value set in each local area is extracted, and the standard deviation and range of the local area temperature are calculated respectively. For area R_1, the temperature data is [38.2, 38.5, 38.1, 38.6, 38.4] degrees Celsius, the standard deviation is calculated to be 0.18℃, and the range is calculated to be 0.5℃. Then normalization processing is performed, where the setting of the local temperature fluctuation normalization threshold σ is based on the surface temperature fluctuation range of the photovoltaic module under the standard environment when the photovoltaic module is working. According to actual tests, when the irradiance is stable at 800W / m 2 , Under the condition of maintaining the ambient temperature at 25℃, the fluctuation range of the component surface temperature is usually within ±0.2℃, so the normalized threshold σ is set to 0.2℃. This setting value fluctuates mainly with the ambient temperature change and irradiance stability. When the ambient temperature changes by ±5℃, the upper limit of the component temperature fluctuation changes by about

[0074] ±0.03℃, irradiance changes by ±100W / m 2 The upper limit of component temperature fluctuation changes by about ±0.04℃. Therefore, in practical applications, the σ value needs to be appropriately adjusted according to the meteorological conditions at the detection site to ensure the consistency and stability of the normalization standard. The normalization process uses the ratio of the temperature standard deviation to the set threshold σ for normalization. If the local area temperature standard deviation is less than the set σ value, the normalized value maintains the original ratio. Otherwise, it is compressed to the maximum standard value to avoid extreme temperature difference interference. For each area, the temperature discrete index after normalization is selected and combined with the corresponding local texture statistics in two dimensions. The matrix unit format is set to [L_f, σ], that is, the data unit of a local area is a combination of its texture statistics and temperature difference value. For example, the combination unit of area R_1 is [0.928, 0.18]. All local areas are processed in turn, and finally a complete matrix is ​​formed to establish the local texture and temperature change feature matrix.

[0075] See also Figure 4 , the specific steps of S3 are:

[0076] S311: Obtain a local texture and temperature change feature matrix, calculate the Manhattan distance of the feature vector of each local region based on the texture statistics and temperature statistics data of each local region in the matrix, screen according to the similarity matching threshold set in the photovoltaic panel surface feature similarity experiment, and generate a feature matching region pair set;

[0077] Obtain the local texture and temperature change feature matrix, extract the corresponding texture statistics and temperature statistics for each local area, and take out the grayscale co-occurrence matrix contrast value and local temperature standard deviation of each local unit respectively. In this example, the photovoltaic module standard division size is selected as 156mm×156mm unit, and a visible light image is divided into 40×40 local areas. Each area is numbered from R1 to R1600. Subsequently, a feature vector is established for each pair of local areas. The grayscale co-occurrence matrix contrast value ranges from 0 to 5, and the regional temperature standard deviation ranges from 0 to 3 degrees Celsius. For example, the feature vector of the R23 area is (2.3, 1.2), and the feature vector of the R57 area is (2.5, 1.4). The Manhattan distance between the two vectors is calculated, and the distance value is

[0078] |2.3-2.5|+|1.2-1.4|=0.2+0.2=0.4. The threshold for similarity matching of photovoltaic panel features is set to 0.5. When the Manhattan distance is less than 0.5, it is determined to be a similar region pair, forming a feature matching region pair set. Table 3 shows some actual region features and screening results.

[0079] Table 3 Feature vectors and screening results

[0080]

[0081]

[0082] As shown in Table 3, the region pairs with Manhattan distance less than 0.5 are retained for subsequent center point extraction and affine transformation derivation. After the screening is completed, a set of feature matching region pairs is generated.

[0083] S312: Based on the feature matching region pair set, extract the center point position of each pair of regions, establish a center point set based on all center points, perform coordinate fitting on the center point set, and derive the local affine transformation initial parameter matrix using the formula:

[0084]

[0085] Calculate the local affine transformation comprehensive difference value A fit , optimize the initial parameter set according to the calculation results and obtain the local affine transformation parameter matrix, where x j Represents the horizontal coordinate of the center point of the jth original local area, y j Represents the ordinate of the center point of the jth original local area, x′ j Represents the horizontal coordinate of the center point of the j-th target local area, y′ j Represents the ordinate of the center point of the jth target local area, Represents the mean value of the horizontal coordinate of the center point of the original local area, represents the mean of the horizontal coordinates of the center points of the local target area, and n represents the number of center point pairs;

[0086] Based on the feature matching region pair set, the center point position of each pair of regions is extracted. In the case of photovoltaic module standard unit division, it is assumed that the center coordinates of the R23 region are (1200, 800) and the center coordinates of the R57 region are (1220, 820). Each pair of matching region center points are numbered in sequence to form a center point set P. Let P = {P1, P2, ..., Pn}, where each P point contains the horizontal and vertical coordinates of the original image and the target image. The local affine transformation initial parameter matrix is ​​derived from all center point sets. On this basis, for each pair of coordinates in the center point set, the following formula is used to calculate the local affine transformation comprehensive difference value. Taking n = 3 as an example, it is assumed that the center point set data is as follows:

[0087] Table 4 Matching center point coordinates

[0088] Matching pair number Original image horizontal coordinate Original image vertical coordinate Target graph horizontal axis Target graph vertical coordinate P1 1200 800 1220 820 P2 2400 1600 2425 1625 P3 3600 2400 3620 2410

[0089] Refer to Table 4. First, calculate the absolute difference between the horizontal and vertical coordinates of each pair of center points:

[0090] P1: |1200-1220|+|800-820|=20+20=40;

[0091] P2: |2400-2425|+|1600-1625|=25+25=50;

[0092] P3: |3600-3620|+|2400-2410|=20+10=30;

[0093] Then calculate the square root difference:

[0094] P1:

[0095] P2:

[0096] P3:

[0097] Calculate the mean:

[0098]

[0099]

[0100] The mean difference is |2400-2421.67|=21.67;

[0101] Substituting into the formula:

[0102]

[0103] The comprehensive difference value of the local affine transformation is 55.99, which shows that there is a certain degree of displacement deviation between the feature region pairs, providing a data basis for the subsequent optimization of the local affine parameter matrix, thereby obtaining the local affine transformation parameter matrix.

[0104] The local affine transformation comprehensive difference value is used to quantify the combined degree of positional offset, overall translational change, and local structural change between the center points of a local region during the matching and alignment process. It is calculated by jointly calculating the absolute offset, Euclidean distance, and mean shift. This value reflects the consistency of the relative position information between local regions before and after the affine transformation. The absolute offset mainly captures the local detail errors in the horizontal and vertical directions, the Euclidean distance reflects the geometric amount of the overall spatial deviation of the local region, and the mean shift reflects whether there is a systematic translational offset in the overall block. Therefore, the smaller the local affine transformation comprehensive difference value, the closer the overall morphology and spatial arrangement between the matching regions, and the better the alignment and superposition of the local region images after the affine transformation. Conversely, the larger the comprehensive difference value, the more significant the misalignment or deformation of the local region during the matching process, and the need for further optimization of the affine parameters. Therefore, this comprehensive difference value not only serves as a direct basis for the derivation of local affine parameters but also effectively guides the subsequent refinement and transformation quality evaluation of local block-level transformations of infrared images.

[0105] The formula first performs the absolute value operation |x j -x′ j |+|y j -y′ j |Calculate the sum of the individual offsets of each pair of center points in the horizontal and vertical positions, and then average them after summing them up, that is, The operation reflects the concentration level of local offsets in the overall area, and then takes the square root of the average offset to balance the dimensional influence of the offset size and enhance the sensitivity of offset fluctuations. Subsequently, the The square root operation of this Euclidean distance measures the actual spatial displacement of each pair of center points, through The average value reflects the overall geometric change trend of the region. At the same time, in order to strengthen the overall level change, the absolute value of the difference between the mean of the horizontal coordinate of the original center point and the mean of the horizontal coordinate of the target center point is further introduced. This item is used to capture the overall translation bias of the region. The three parts of the results are directly added to form the local affine transformation comprehensive difference value A fitThe overall design reflects the joint consideration of local micro-offsets, overall position drift and local geometric structure changes. The addition operation is responsible for superimposing the offset contributions from different sources, the square root operation is used to measure the change in spatial distance, the absolute value operation is used to capture the direction-independent offset, and the average operation is used to unify the scale standard to ensure that the offset information of different dimensions can be synthesized into the comprehensive difference value on the same scale, thereby providing a reliable error quantification basis for subsequent affine parameter fitting.

[0106] S313: performing an affine transformation operation on each local block according to the local affine transformation parameter matrix and the local block index of the infrared image, establishing a transformed regional block set, and generating a local affine registration matrix;

[0107] According to the local affine transformation parameter matrix, according to the corresponding affine transformation parameters of each region, affine transformation processing is performed on the infrared image blocks. Each local block performs coordinate translation, rotation and scale change operations. Assuming that the starting coordinates of a single local block infrared image are (100, 100), the affine matrix is:

[0108]

[0109] Perform local transformation processing, and the new position of a single pixel point (110, 115) after transformation is:

[0110] New horizontal coordinate = 0.98 × 110 + 0.05 × 115 = 107.8 + 5.75 = 113.55;

[0111] New ordinate = -0.03 × 110 + 1.02 × 115 = -3.3 + 117.3 = 114.00;

[0112] Thus, a local block coordinate mapping is completed, and each local block is transformed in batches to form an overall affine mapping of the local block, a new set of local blocks of the infrared image is established, and finally a local affine registration matrix is ​​generated.

[0113] See also Figure 5 , the specific steps of S4 are:

[0114] S411: Obtain the infrared image corrected by the local affine registration matrix, set the starting temperature threshold to be the ambient temperature plus 5 degrees Celsius, extract the temperature area greater than the threshold to generate a temperature contour standard, establish a corresponding contour set, and generate a temperature contour extraction result;

[0115] Obtain the infrared image after local affine registration matrix correction. First, select the ambient temperature measurement value of 25℃, and set the initial temperature threshold of 30℃ by adding 5℃. Based on this temperature threshold, mark all pixel areas with temperatures above 30℃ on the image to delineate the preliminary high-temperature area. Combined with regional grid division, the infrared image is divided into a standard grid of 30×30 pixels per grid. The maximum, minimum and average temperature values ​​of each grid are detected. If the temperature value of any pixel in the grid is greater than 30℃, the grid is regarded as a temperature abnormal grid, and the coordinates of the grid center point are recorded. For example, the coordinates of the grid center point detected in a certain area are (150, 200), and the highest temperature in the grid corresponding to this point is 32.5℃. The grids below the threshold are ignored. Then, the temperature contour is traced in the entire image with the local abnormal grid as the starting point to form a continuous temperature contour line. The sampled temperature data are shown in Table 5.

[0116] Table 5 Temperature threshold extraction and regional temperature statistics

[0117]

[0118] As shown in Table 5, according to the extraction results, it can be confirmed that several temperature contour lines are extracted, and the temperature contour line extraction results are obtained.

[0119] S412: Based on the temperature contour extraction results, detect the head-to-tail connection relationship of each temperature contour, count the number of contours that meet the head-to-tail closure conditions and the total number of extracted contours, using the formula:

[0120]

[0121] Calculate the closed contour feature metric R c , the local thermal closure area is evaluated according to the closed contour feature metric value, and the thermal closure feature index set is obtained, where N c Represents the number of closed contour lines, N t Represents the total number of extracted contour lines, p k represents the horizontal coordinate of the starting point of the kth contour line, q k Represents the abscissa of the end point of the kth contour line;

[0122] Based on the temperature contour extraction results, first extract the starting point coordinates and end point coordinates in each temperature contour, and set the starting point horizontal coordinate as p k The horizontal coordinate of the end point is q k, all the extracted contour lines are compared one by one to determine the head-to-tail connection relationship. It is defined that if the distance between two points is less than 3 pixels, the contour line is considered closed. In the actual measurement sample, if there is p1 = 152 and q1 = 154, the distance between the two points is 2 pixels, which is considered a closed line. The above closure judgment is performed on all 25 extracted contour lines, and 18 closed contour lines are detected. Based on this statistical result, the formula is used to calculate N c =18, N t =25. For the difference between the starting point and the end point, assuming that the difference between the five samples is 3, 2, 4, 3, and 2 respectively, substitute it into the formula and the specific calculation is as follows:

[0123] beg

[0124] 3+2+4+3+2=14;

[0125] beg

[0126] beg because So the total is still 14;

[0127] Evaluate the overall expression:

[0128]

[0129] The final closed contour feature metric value is 0.386, indicating good closing characteristics, which is then used to generate the thermal closure feature index set.

[0130] The closed contour feature metric is used to quantitatively describe the degree of closure and structural coherence of the temperature contour lines extracted from the infrared image in the local area. This feature metric reflects the continuity and enclosure integrity of the temperature anomaly area in the local area by comprehensively considering the proportion of closed contour lines, the absolute difference in distance between the starting and ending points of the contour lines, and the overall distance deviation. Among them, the proportion of closed contour lines directly reflects whether the local high-temperature area has a continuous enclosure structure. The absolute value and square root of the difference between the starting and ending points reflect the smoothness and proximity of the local curve. The overall multiplication linkage balances the two dimensions of closure and tightness. The closer the value of this indicator is to 1, the continuous connection and complete boundary closure of the temperature contour lines in the local area are, and the local high-temperature features are stable and regular in shape. The lower the value, the more broken, distorted or unclosed temperature structures exist in the local area. Therefore, this feature metric can provide a basic and quantitative structural description for subsequent texture density change analysis and local thermal anomaly detection.

[0131] This formula comprehensively reflects the closing characteristics of local temperature contours through multi-level operational relationships, among which, firstly, The direct ratio of the number of closed contour lines to the total number of extracted contour lines is calculated to reflect the basic level of overall closure. This ratio reflects the integrity of the basic structure. This item takes the square root of the sum and average of the absolute values ​​of the horizontal coordinate differences between the starting point and the end point of each contour line. The purpose is to reduce the pulling effect of large differences on the overall statistics through square root operation, emphasizing the stability change trend of the overall closure. Therefore, the square root operation is a process of smoothing outliers. The above two parts are combined by addition to form the first indicator, which comprehensively reflects the local overall closure status. Then, it is multiplied by an adjustment factor on the right side. This part sums the square roots of the Euclidean distances of all contour line starting and ending points, calculates the ratio after adjusting with the closure number, and then converts it into an inverse indicator of closure tightness through a one-subtraction method to reflect the degree of local deviation. The summation and square root processes correspond to the accumulation and nonlinear suppression differences respectively. Finally, through the overall linkage of multiplication, the uniformity and closure of the local temperature contour line structure are comprehensively reflected, ensuring the accuracy and refinement of the overall feature evaluation.

[0132] S413: Based on the thermal closure feature index set and the corresponding local area of ​​the visible light image, extract the brightness gradient amplitude statistics, calculate the gray level co-occurrence matrix energy ratio between the central area and the peripheral area, calculate the texture density change, and combine the thermal closure feature index set and the texture density change to establish the thermal closure and texture density change characteristics;

[0133] According to the thermal closure feature index set and the corresponding visible light local area in the infrared image, the central area is first defined as a circular area with a radius of 30 pixels, and the peripheral area is defined as the outer ring of the center to 60 pixels. The energy value of the gray level co-occurrence matrix is ​​extracted for the central area and set as E c , the energy value of the surrounding area is set to E s , for example, the energy value E of the central area is sampled c =0.85, surrounding energy value E s =0.65, then the texture density change is calculated as the energy ratio:

[0134]

[0135] It indicates that the texture density in the central area is about 30.8% higher than that in the surrounding area. Combined with the closure measurement value of the corresponding area in the thermal closure feature index, a local joint feature is formed, and finally the thermal closure and texture density change features are generated.

[0136] See also Figure 6 , the specific steps of S5 are:

[0137] S511: Acquire the thermal closure and texture density change features, and detect whether the closed contour feature metric value of each region is greater than the set closed contour ratio threshold. The closed contour ratio threshold is set to 0.35 based on experiments. If the closed contour metric value of a region is 0.42, it is determined to be a qualified region. Otherwise, it is eliminated. The screening process directly compares the closed contour metric value of each local region with the set threshold value. After execution, a closed contour ratio screening region set is generated;

[0138] After obtaining the thermal closure and texture density change characteristics and the thermal closure feature index and texture density change value of each local area, each local area is first screened according to the set closed contour ratio threshold. The threshold is set to 0.35, that is, if the closed contour feature metric value is greater than 0.35, the area is retained, and if it is less than or equal to it, it is eliminated. In actual operation, the closed contour feature value of area A is detected to be 0.42, which is greater than the threshold. Therefore, it is judged that area A meets the screening conditions, and the detection value of area B is 0.31, so it is eliminated and no longer enters the subsequent processing. During the detection, the area number is divided according to the actual photovoltaic module and the corresponding sampling is performed. For example, area A is numbered A01 and area B is numbered B02. In order to verify the screening logic, the closed contour feature sample data of the initial detection is recorded and listed in Table 7, as shown in Table 7:

[0139] Table 7 Local area closure feature data table

[0140] Area Number Closed contour eigenvalues A01 0.42 B02 0.31 C03 0.45 D04 0.36

[0141] As shown in Table 7, regions A01, C03, and D04 meet the requirement that the closure measure value is greater than 0.35, and region B02 is screened out. During the screening process, a direct comparison is strictly performed based on the size relationship between the closure characteristic value of each region and 0.35. All data are measured in the same acquisition batch and the temperature environment is stable. After screening, a closed ratio screening region set is formed.

[0142] S512: Filter the region set based on the closure ratio, detect the texture density change corresponding to each region, and set the texture density change threshold to 1.2. By directly comparing the texture density change of the local region with the threshold, the region with a change greater than 1.2 is retained as the valid region. At the same time, the thermal closure feature value, closure ratio, center grayscale energy value, and edge grayscale energy value of the local region are jointly evaluated to comprehensively determine the regional characteristics and obtain the defect region screening results;

[0143] Based on the closed ratio screening area set, the above-screened areas are further screened according to the texture density change. The texture density change critical value is set to 1.2. For each local area, the grayscale co-occurrence matrix energy value of the central area and the energy value of the surrounding area are extracted, and the energy change ratio is calculated as the texture density change. When the change is greater than 1.2, it is considered that the texture density change is significant and the area is retained. Otherwise, it is eliminated. In the actual sampling, the energy value of the central area of ​​area A01 is 0.85, the energy value of the surrounding area is 0.65, and the texture density change is calculated as D e =0.85 / 0.65≈1.31, which is greater than 1.2 and meets the retention condition. The central energy of area C03 is 0.78 and the peripheral energy is 0.75. The change D e =0.78 / 0.75≈1.04, which is less than 1.2, so it is determined to be eliminated. The energy of the center of area D04 is 0.90, and the energy of the periphery is 0.70. The change D e =0.90 / 0.70≈1.29, which meets the screening conditions. Therefore, the two areas A01 and D04 are retained in the final screening. The center grayscale energy, edge grayscale energy, and thermal sealing characteristic values ​​of these areas are synchronously recorded in the screening table for subsequent call processing. The defect area screening results are obtained by directly comparing the relationship between the texture density change and 1.2.

[0144] S513: Based on the defect area screening results, extract the center coordinates of each local area that meets the screening conditions, aggregate all the center coordinates into a set, and finally generate a photovoltaic module defect location coordinate set;

[0145] According to the defect area screening results, the center coordinates are extracted based on the position information of each retained area. The center coordinates of area A01 are located at (150, 200), and the center coordinates of area D04 are located at (450, 300). During the extraction, the original pixel position data of each area during detection is used as the benchmark, and the horizontal and vertical pixel coordinate values ​​are recorded separately without any coordinate translation or rotation processing to ensure that the coordinate values ​​are consistent with the original detection image. The center point coordinates of all eligible areas are collected to form a list. Each row in the list corresponds to the center position data of a defect area. Finally, all the center coordinates are sorted and archived, and a unified coordinate set is established to finally generate the photovoltaic module defect location coordinate set.

[0146] The above are merely preferred embodiments of the present invention and do not limit the present invention in any other form. Any technician familiar with the profession may use the technical content disclosed above to change or modify it into an equivalent embodiment with equivalent changes and apply it to other fields. However, any simple modification, equivalent change and modification made to the above embodiment based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall still fall within the scope of protection of the technical solution of the present invention.

Claims

1. A photovoltaic defect detection method based on the combination of visible light and infrared images from drones, characterized by: The following steps are involved: S1: Determine the inspection area of ​​the photovoltaic power station and use a drone equipped with a visible light camera and an infrared thermal imager to perform synchronized aerial photography. The image groups are filtered according to the time synchronization tolerance threshold to generate synchronized visible light and infrared image pairs. S2: performing local segmentation based on the synchronized visible light and infrared image pair, calculating local texture statistics of the visible light image and local temperature statistical differences of the infrared image, and establishing a local texture and temperature change feature matrix; S3: According to the local texture and temperature change feature matrix, screen local region pairs and extract a center point set, fit the center point set to derive local affine transformation parameters, perform local block-level transformation of the infrared image, and generate a local affine registration matrix; S4: Correcting the infrared image using the local affine registration matrix, extracting temperature contours and calculating the closed contour ratio, and simultaneously calculating the relative change in texture density in the corresponding local area of ​​the visible light image to establish thermal closure and texture density change features; S5: According to the thermal sealing and texture density change characteristics, the defect local area is screened, the coordinates of the center of the area are extracted, and the photovoltaic module defect location coordinate set is obtained by summarizing.

2. The photovoltaic defect detection method based on the combined visible light and infrared image of UAV according to claim 1 is characterized in that: The synchronized visible light and infrared image pair includes the image timestamp synchronization difference, the image space alignment error, and the synchronization tolerance judgment result; the local texture and temperature change feature matrix includes the local texture feature index, the local temperature statistical difference value, and the local area division number; the local affine registration matrix includes the local affine transformation parameters, the center point set coordinates, and the feature matching similarity score; the thermal closure and texture density change characteristics include the temperature contour line closure ratio, the texture density change amount, and the local brightness gradient statistics; the photovoltaic module defect positioning coordinate set includes the defect area center coordinates, the defect detection confidence, and the defect area number.

3. The photovoltaic defect detection method based on the combined visible light and infrared imaging of drones according to claim 1 is characterized in that: The steps of acquiring the synchronized visible light and infrared image pair are specifically as follows: S111: Obtain the PV power station inspection operation area, set the equipment system clock preliminary calibration parameters and record the flight start timestamp based on the requirements of synchronized aerial photography with the UAV-mounted visible light camera and infrared thermal imager, and generate a preliminary equipment time synchronization reference value; S112: Based on the preliminary device time synchronization reference value, detect the acquisition timestamp of each group of images, establish a time record sequence for each visible light image and infrared image, calculate the time difference between the visible light image and the infrared image based on the GNSS timing synchronization threshold, select the image group that meets the synchronization conditions, and obtain a time-synchronized image sequence; S113: performing image alignment processing according to the time-synchronized image sequence and the time-series index value, filtering out image groups that do not meet the synchronization conditions, obtaining one-to-one corresponding combinations of visible light images and infrared images, and generating synchronized visible light and infrared image pairs.

4. The photovoltaic defect detection method based on the combined visible light and infrared image of a drone according to claim 1 is characterized in that: The steps for obtaining the local texture and temperature change feature matrix are specifically as follows: S211: Acquire the synchronized visible light and infrared image pairs, perform region division processing based on standard size parameters of photovoltaic modules, divide each image into multiple local regions, and generate a local image region unit set; S212: Based on the local image region unit set, calculating the texture statistics of each visible light local region, establishing preliminary texture values ​​and variance values ​​based on the gray level co-occurrence matrix contrast and the local binary pattern variance, performing difference synthesis on the gray level characteristics and structural characteristics, calculating the local texture statistical synthesis value, completing the local texture feature quantization based on the difference in the synthesized texture characteristics, and generating a local texture statistics set; S213: According to the local texture statistics set and the infrared image areas in the local image area unit set, based on the temperature distribution data of each area, the regional temperature standard deviation and range value are calculated respectively, the temperature discreteness is normalized, the local texture statistics are combined with the temperature difference, and a local texture and temperature change feature matrix is ​​established.

5. The photovoltaic defect detection method based on the combined visible light and infrared image of UAV according to claim 1 is characterized in that: The steps for obtaining the local affine registration matrix are specifically as follows: S311: Obtain the local texture and temperature change feature matrix, calculate the Manhattan distance of the feature vector of each local area based on the texture statistics and temperature statistics data of each local area in the matrix, screen according to the similarity matching threshold set in the photovoltaic panel surface feature similarity experiment, and generate a feature matching area pair set; S312: Based on the feature matching region pair set, extract the center point position of each pair of regions, establish a center point set based on all the center points, perform coordinate fitting on the center point set, derive a local affine transformation initial parameter matrix, calculate the local affine transformation comprehensive difference value, optimize the initial parameter set, and obtain a local affine transformation parameter matrix; S313: performing an affine transformation operation on each local block according to the local affine transformation parameter matrix and based on the local block index of the infrared image, establishing a transformed regional block set, and generating a local affine registration matrix.

6. The photovoltaic defect detection method based on the combined visible light and infrared image of UAV according to claim 1 is characterized in that: The steps for obtaining the thermal sealing and texture density change characteristics are specifically as follows: S411: Acquire the infrared image corrected by the local affine registration matrix, set a temperature threshold starting value, extract the temperature area greater than the temperature threshold to generate a temperature contour standard, establish a corresponding contour set, and generate a temperature contour extraction result; S412: Based on the temperature contour extraction results, detecting the head-to-tail connection relationship of each temperature contour, counting the number of contour lines that meet the head-to-tail closure condition and the total number of extracted contour lines, calculating a closed contour feature metric value, and performing a local thermal closure area assessment based on the closed contour feature metric value to obtain a thermal closure feature index set; S413: Based on the thermal closure feature index set and the corresponding local area of ​​the visible light image, the brightness gradient amplitude statistics are extracted, the grayscale co-occurrence matrix energy ratio of the central area and the peripheral area is calculated, the texture density change is counted, and the thermal closure and texture density change characteristics are established by combining the thermal closure feature index set and the texture density change.

7. The photovoltaic defect detection method based on the combined visible light and infrared image of UAV according to claim 1 is characterized in that: The steps for obtaining the photovoltaic module defect location coordinate set are specifically as follows: S511: Obtain the thermal closure and texture density variation characteristics, detect whether the closed contour feature metric value of each region is greater than a set closed contour ratio threshold, and directly compare the closed contour metric value of each local region with the set closed contour ratio threshold to generate a closed ratio screening region set; S512: Filter the region set based on the closure ratio, detect the texture density change corresponding to each region, compare the texture density change of the local region with the critical value, retain the region where the change is greater than the critical value as the valid region, and simultaneously perform a joint evaluation of the thermal closure feature value, closure ratio, center grayscale energy value, and edge grayscale energy value to comprehensively determine the regional characteristics and obtain the defect region screening results; S513: extracting the center coordinates of each local area that meets the screening conditions according to the defect area screening result, and summarizing and arranging all the center coordinates into a set, and finally generating a photovoltaic module defect location coordinate set.

Images