A Solar Panel Anomaly Detection Method Based on YOLO V8 Model

Abstract

The photovoltaic panel anomaly detection method based on the YOLO V8 model comprises the following steps: collecting photovoltaic panel images through a drone; collecting and preprocessing the photovoltaic panel images; transmitting the preprocessed photovoltaic panel image data to a server; setting a prediction model on the server side based on the YOLO V8 model to predict the photovoltaic panel images transmitted by the server; setting a training model on the server side based on the YOLO V8 model, normalizing the positive and negative examples in the photovoltaic panel image data set through the path of the positive and negative examples, and training through the built-in objective function; through the above steps, the photovoltaic panel anomaly detection is realized. The method modifies the sampling method in the YOLO V8 model, combines the existing deformable convolution method and the modified activation function, and can well adapt to the problem of spot deformation diversity; compared with the conventional fixed shape convolution, the deformable convolution can better adapt to different receptive fields and improve the sampling effect.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic power plant inspection technology, specifically to a photovoltaic panel anomaly detection method based on the YOLO V8 model. Background Technology

[0002] Due to the large area and extensive photovoltaic panel layout of mountain photovoltaic power stations, manual inspection is costly and complex. Using drones for intelligent inspection has become a new option. However, drone platforms have limitations, including restrictions on the equipment they can carry and the capacity and power consumption of their batteries. Therefore, the identification of photovoltaic panels through infrared photography has traditionally relied on human visual inspection or simple image recognition algorithms such as feature engineering and frequency domain processing. Both of these approaches suffer from low accuracy and high workload, making it difficult to achieve satisfactory predictive results.

[0003] In addition, since solar photovoltaic panel stations are located at great distances, basic communication services such as wireless base stations and data transmission signals are subject to unstable quality. Ensuring stable and effective data transmission to remote servers is a significant challenge.

[0004] Due to the specialized nature of the power electronics industry, a lack of understanding of dataset design and objective function formulation, as well as improper data arrangement within the dataset, can negatively impact the final model's prediction performance. Therefore, in image detection, one option is to use the mature YOLO V8 model to detect the transmitted images. However, in existing technologies, due to the significant variations in light spots, conventional convolutional samplers often struggle to achieve satisfactory sampling results. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention proposes a photovoltaic panel anomaly detection method based on the YOLO V8 model. This method modifies the sampling method in the YOLO V8 model and combines the existing deformable convolution method with the modified activation function, which can well adapt to the problem of diverse light spot deformation. Deformable convolution can better adapt to different receptive fields than conventional fixed-shape convolution, thus improving the sampling effect.

[0006] The technical solution adopted in this invention is as follows:

[0007] A photovoltaic panel anomaly detection method based on the YOLO V8 model includes the following steps:

[0008] Step 1: Acquire images of photovoltaic panels using a drone;

[0009] Step 2: Collect and preprocess images of photovoltaic panels;

[0010] Step 3: Transmit the photovoltaic panel image data after the preprocessing in Step 2 to the server;

[0011] Step 4: Set up a prediction model on the server based on the YOLO V8 model to predict the photovoltaic panel images transmitted from the server in Step 3;

[0012] Step 5: Set up the training model on the server side based on the YOLO V8 model. By using the paths of the positive and negative samples in the photovoltaic panel image dataset, the training model will normalize the positive and negative samples in the photovoltaic panel image dataset and train it through the built-in objective function.

[0013] The above steps enable the detection of photovoltaic panel anomalies.

[0014] Step 2, the photovoltaic panel image preprocessing includes the following steps:

[0015] S2.1: Determine if the photovoltaic panel image is qualified. If yes, proceed to S2.2; otherwise, discard it directly.

[0016] S2.2: Determine whether the image is an abnormal hotspot. If so, mark the abnormal location; otherwise, mark it as a normal image and proceed to S2.3.

[0017] S2.3: Put all abnormal images in one path folder, and all normal images in another path folder.

[0018] Step 4, predicting the photovoltaic panel image, includes the following steps:

[0019] S4.1. Algorithm Training: Based on the YOLO V8 model, the positive and negative samples in the dataset are normalized and trained using the built-in objective function;

[0020] S4.2. Model Training: The YOLO V8 model was trained using a large amount of labeled data to achieve the expected performance.

[0021] S4.3. Image preprocessing: Determine if the image is abnormal and label it; images that do not meet the quality requirements will be discarded directly.

[0022] S4.4. Server-side prediction: Performed after the YOLO V8 model is fully deployed.

[0023] Step 5 includes the following steps:

[0024] S5.1: Configure the validation set, training set, and learning rate in the configuration file, and then read the configuration file;

[0025] S5.2: Set termination conditions;

[0026] S5.3: Determine if the termination condition has been met. If yes, proceed to S5.4; otherwise, terminate the training.

[0027] S5.4: Image Segmentation: Divide an image into multiple segmentation units;

[0028] S5.5: Calculate partition confidence:

[0029] The formula for confidence prediction is based on the following formula:

[0030]

[0031] This formula calculates the confidence level of each cell containing a particular object type. The cell's type is determined by the object type with the highest confidence level. `Pr(Object)` indicates whether an object exists in the bounding box, and `Pr(Class)`... i |Object) indicates that this object is a Class i The probability. If an image is the background, then this probability should be 0; This represents the ratio of the interaction area between the true predicted bounding box (dataset calibration) and the predicted bounding box predicted using the neural network;

[0032]

[0033] Where A∩B refers to the area where the prediction box and the calibration box intersect, and A∪B refers to the area where the prediction box and the calibration box merge.

[0034] For each prediction box, the classification with the highest confidence in each class is calculated, and then the final prediction result is obtained based on the classification.

[0035] Because the hot spot area captured by the drone during detection is relatively small, the coefficient of IOU is increased when calculating the detection confidence level, as follows:

[0036]

[0037] The coefficient λ is adjusted based on the images acquired on-site.

[0038] S5.6: Calculate the objective function and return to S5.3;

[0039] The objective function considers the current prediction quality from two aspects: ①. Whether the target exists within the current bounding box. ②. The difference between the center coordinates of the bounding box and the centrally labeled coordinates in the dataset. The squared error function is used to measure the bounding box center error because it is easy to optimize. Since using the squared error function can affect accuracy, the coefficient λ for the bounding box error is increased when considering the coefficients. coord However, it reduced the coefficient λ for classification confidence. noobj .

[0040] The objective function is as follows:

[0041]

[0042] Where: λ coord The coefficient representing the bounding box error; (x i y i () refers to the center point position of the i-th predicted box calculated by the neural network; w refers to the position of the center point of the i-th predicted bounding box in the dataset. i h i Let represent the width and height of the i-th predicted bounding box, respectively, predicted by the neural network. These represent the width and height of the predicted bounding boxes labeled in the dataset, respectively. S represents the number of horizontal and vertical squares in the image segmentation; B represents the number of bounding boxes in each grid cell after image segmentation. This indicates whether the target exists in the i-th row and j-th cell.

[0043] In step 5, the convolution method of the receptive field in the YOLO V8 model is modified. The specific formula is as follows:

[0044] f (i，j) =∫ p∈R heat((i,j)+p)dp;

[0045] Where: P represents the receptive field of the deformation; (i, j) represents the specific coordinates; f(i, j) represents the value of (i, j) in the feature map;

[0046] The receptive field R of a feeling is represented as follows:

[0047] R={(x, y)|x, y∈(-1, 1)}.

[0048] This invention provides a photovoltaic panel anomaly detection method based on the YOLO V8 model, with the following technical advantages:

[0049] 1) In step 1 of this invention, the photovoltaic panel images collected by DJI drones have high resolution, which facilitates clear classification and is essential for subsequent model training.

[0050] 2) In step 3 of this invention, a specific method for image transmission and compression is designed based on the DJI platform to process the acquired images, effectively reducing costs.

[0051] 3) In step 3 of this invention, after the image is captured by the DJI drone, it will be efficiently transmitted to the server for prediction and recognition with the highest transmission quality. This greatly improves the transmission efficiency and the final prediction quality.

[0052] 4) In step 4 of this invention, the prediction system organically combines drones, ground base stations or wireless access points, and remote servers. Based on the YOLO V8 model, it can significantly improve the accuracy and efficiency of prediction.

[0053] 5) In step 5 of this invention, the correlation logic between the objective function, data labeling, and training parameters is organically combined, using the YOLO V8 model. This effectively reduces the learning curve for operators and ensures that the final data prediction reaches a satisfactory level.

[0054] 6) In the detection method of the present invention, in step 5, a method combining deformable convolution and modifying the objective function was adopted to modify the sampler, making it more suitable for hot spot detection. Attached Figure Description

[0055] The present invention will be further described below with reference to the accompanying drawings and embodiments:

[0056] Figure 1 This is a schematic diagram of the anomaly detection process of the present invention.

[0057] Figure 2 This is a schematic diagram of the training process of the YOLO V8 model for anomaly detection in this invention. Figure 1 .

[0058] Figure 3 This is a diagram illustrating the division of an image into multiple segments.

[0059] Figure 4 A schematic diagram illustrating the data transmission of an unmanned aerial vehicle (UAV) system.

[0060] Figure 5 This is a timing diagram for the transmitted signals.

[0061] Figure 6 This is a schematic diagram of the training process of the YOLO V8 model for anomaly detection in this invention. Figure 2 .

[0062] Figure 7 This is a flowchart of the image preprocessing process. Detailed Implementation

[0063] A photovoltaic panel anomaly detection method based on the YOLO V8 model, comprising the following parts:

[0064] (I) Data Collection and Labeling / Classification:

[0065] For training a neural network model, a high-quality dataset with well-defined categories is crucial. Once data collection is complete, the process automatically classifies all data into the correct categories. This is a fundamental requirement for later model training; correct classification is the basis for automated model training.

[0066] (II) Image Transmission and Compression Methods:

[0067] Due to cost constraints, images will be processed and compressed before being transmitted to the server for further processing. The method of image compression and transmission significantly impacts the final prediction quality. This invention presents a specific method for image transmission and compression on the DJI platform, ensuring that images captured by DJI drones are efficiently transmitted to the prediction and recognition server with maximum transmission quality.

[0068] (III) Model Reasoning and Training:

[0069] Model inference and training are two separate steps. Given a dataset, the model will automatically train based on a given single-objective function and optimization method, using data types labeled within the given dataset. After the trained model is deployed, the server will receive the image and pass it to the model's input. The model will then classify and predict the current image based on the parameters.

[0070] Model data calibration, objective function verification, and specific optimization training methods include:

[0071] a. Mark the data as positive or negative based on the predefined labels in the dataset;

[0072] b. Configure the validation set, training set, and learning rate in the configuration file;

[0073] c. Configuration related to starting training.

[0074] (iv). This invention designs a prediction system for an anomaly detection algorithm.

[0075] This prediction system organically combines drones, ground base stations or wireless access points, and remote servers to achieve wireless communication at remote sites. Based on the YOLO V8 model, it designs a deployment and training module for an algorithm model that can easily and effectively classify data and train efficiently. The prediction system focuses on efficiently and stably transmitting relevant data from remote sites to the server, and then performing efficient and accurate predictions on the transmitted images at the server end.

[0076] In terms of algorithm training, the correlation logic between the objective function, data labeling, and training parameters is organically combined. Using the YOLO V8 model, an efficient and simple automated training model is designed. By following the paths of positive and negative samples in the dataset, YOLO V8 normalizes the positive and negative samples in the dataset and trains them using the built-in objective function, reducing the learning curve for operators and ensuring that the final data predictions reach a satisfactory level.

[0077] The model training process is as follows: Figure 6 As shown: Model training is a process of training a model using a large amount of collected labeled data to achieve the expected performance. First, an image preprocessing workflow is used to filter out qualified and unqualified images. Then, the qualified images are labeled and placed in the correct path folder. Training will be started from the command line. After training is completed, the trained model is placed in the appropriate path folder.

[0078] The image preprocessing process is as follows: Figure 7 As shown, images are labeled based on their quality to determine if they are abnormal. Images with substandard quality are discarded. Additionally, images containing hotspot anomalies require labeling of the abnormal regions using a tool; otherwise, simply mark them as normal. All images should be placed in their respective folders. All abnormal images should be placed in one folder, and all normal images in another, to facilitate subsequent model training.

[0079] The server-side prediction portion will be performed after the model is fully deployed. The model needs to be trained on a separate cluster and can only be deployed to the server after all metrics meet predetermined standards. The training and deployment process is generally carried out on separate servers. The model uses the YOLOv8 model for prediction and training, and the recommended deep learning framework is PyTorch.

[0080] The YoLoV8 model offers three options:

[0081] ①.YOLOV8n(NanoModel);

[0082] ②.YOLOv8s(SmallModel);

[0083] ③.YOLOV8m(MediumModel);

[0084] The model selection can be ultimately determined based on the specific needs of the project.

[0085] (V) Technical Principles:

[0086] During training, the entire process is as follows: Figure 2 As shown, this section relates to the objective function, optimization algorithm, and termination condition. The configuration file stores the training dataset, positive and negative samples, and related training parameters. These are then read and used as training parameters for the algorithm. Object detection is redefined as a single regression problem, directly from image pixels to bounding box coordinates and class probabilities. Using the above prediction system, the presence and location of objects can be predicted with only a single look at the (YOLO) image.

[0087] 1. When partitioning an image, first divide the image into multiple segments, such as... Figure 3 As shown, the initial image is divided into 7x7 partitions. A confidence score is calculated for each partition, reflecting the most reliable classification result. First, YOLO is extremely fast. Because this invention treats the detection box as a regression problem, complex pipelines are unnecessary. The neural network is simply run on new images during testing to predict detections. The base network runs at 45 frames per second without batch processing on a Titan X GPU, while the fast version runs at over 150 fps. Therefore, streaming video can be processed in real-time with a latency of less than 25 milliseconds. Furthermore, YOLO's average accuracy is more than twice that of other real-time systems. Second, YOLO performs global inference on the image during prediction. Unlike techniques based on sliding windows and region proposals, YOLO sees the entire image during training and testing, thus implicitly encoding contextual information about the category and its appearance. Fast R-CNN, a top-tier detection method, misidentifies background patches in an image as objects because it doesn't see the larger context. Compared to Fast R-CNN, YOLO produces less than half the number of background errors.

[0088] 2. During detection, the independent components of object detection are unified into a single neural network. This invention's network uses features from the entire image to predict each bounding box. It also simultaneously predicts all bounding boxes for all categories in the image. This invention's network performs global inference on the entire image and all objects within it. The YOLO design supports end-to-end training and real-time speed while maintaining high average accuracy. Generally, the confidence prediction formula is based on the following formula:

[0089]

[0090] This formula calculates the confidence level of each cell containing a particular object type. The cell's type is determined by the object category with the highest confidence level. First, for an input image, we divide it into regions based on previously defined parameters. Figure 3 For example, we will first divide this image into a 7x7 grid, and answer two questions in each grid:

[0091] ①. Is this grid the background or does it contain a target to be detected?

[0092] ②. What type of items are being tested?

[0093] The answers to both questions are reflected in the formulas: Pr(Object) indicates whether an object exists in the bounding box, and Pr(Class)... i |Object) indicates that this object is a Class iThe probability. If an image is the background, then this probability should be 0. This represents the ratio of the interaction area between the true predicted bounding box (dataset calibration) and the predicted bounding box predicted using the neural network. Where A∩B refers to the area where the predicted bounding box and the calibration box intersect, and A∪B refers to the area where the predicted bounding box and the calibration box merge.

[0094]

[0095] For each prediction box, the highest confidence level in each class is calculated, and the final prediction result is obtained based on the classification. Probabilistic prediction is performed by a neural network. Yolo uses a convolutional network to extract features and then uses fully connected layers to obtain the predicted values. Because the hotspot area captured by the drone during detection is relatively small, the IOU coefficient is emphasized when determining the detection confidence.

[0096]

[0097] The coefficient λ is adjusted based on the images collected on-site to ensure that the drone can still effectively detect variables when adjusting the drone usage method later.

[0098] The classic YOLO objective function is defined as follows: During training, a multi-part objective function is used to measure the reliability of the final prediction quality of the current neural network parameters. The objective function considers the current prediction quality from two aspects: a. whether the target exists within the current bounding box, and b. the difference between the center coordinates of the bounding box and the calibration in the dataset. We use the squared error function to measure the bounding box center error because this function is convenient for optimization. Since using the squared error function can affect accuracy, the coefficient λ for the bounding box error is increased when considering the coefficients. coord However, it reduced the coefficient λ for classification confidence. noobj . λ coord It is generally 5, but λ noobj The value is typically 0.5. The objective function is shown below:

[0099]

[0100] This section is completed in two parts: the first part represents the error of the bounding box center coordinates, measured using the second norm; the second part measures the error of the bounding box height and width, including the confidence error term for bounding boxes containing the target. The fourth part is the confidence error term for bounding boxes not containing the target. In the evaluation, because it is necessary to highlight that small fluctuations within large predicted boxes are more important than fluctuations within small boxes, the height and width of the predicted boxes are measured separately to meet the final prediction criteria. This indicates whether an object exists in the j-th detection box within the i-th segmentation unit. This indicates whether an object exists in the i-th segmentation unit; if it exists, this value is 1, otherwise it is 0. (x i y i () refers to the center point position of the i-th predicted bounding box calculated by the neural network. w refers to the position of the center point of the i-th predicted bounding box in the dataset. i h i These refer to the width and height of the i-th predicted bounding box, respectively, predicted by the neural network. C represents the width and height of the predicted bounding boxes labeled in the dataset, respectively. i , These represent the predicted and actual confidence scores, respectively. i (c), This represents the probability of a certain category appearing in the i-th segmentation unit. In practical applications, since the number of target categories is relatively small and the area of ​​the predicted bounding box is generally small during hotspot detection, this invention makes the following modifications:

[0101] 1) Replacement and As an indicator variable, use the existence measure function An existence measurement is performed. If no hotspot exists in the current region and detection frame, the indicator function is 0. If a hotspot exists, the function value is proportional to the hotspot's temperature. The specific calculation method is as follows:

[0102]

[0103] Based on discrete coordinates, a double integral operation is performed on the heatmap to highlight the influence of heat values ​​on the function. ij Let represent the pixels in the prediction box, and heat(x) represent the heat value of the corresponding pixel. The double integral in discrete space is defined as:

[0104]

[0105] Where x represents the row and y represents the column; the integration region is defined as [x...]. l x r ]×[y l y r ];

[0106] 2) When calculating the bias of the prediction box, w i h i , We no longer perform square root operations because the predicted bounding boxes are generally small, and square root operations would excessively reduce the accuracy of the detected bounding boxes.

[0107] 3) Remove confidence level and target existence detection;

[0108] 4) Since the shape of hot spots is quite diverse in regional hot spot detection, a variable convolution method is added to the conventional block division to adapt the shape.

[0109] The optimized objective function is:

[0110]

[0111] The YOLO model predicts multiple classification criteria for each segmented region, but during training, we need only one bounded box to be responsible for the predicted object. Therefore, the concept of Intersection over Union (IOU) is introduced; only predictions that match the IOS are reflected in the objective function as having positive value. The loss function only penalizes classification if the object exists in that grid cell; otherwise, it's incorrect. The prediction variable is considered "responsible" for the ground truth box (i.e., it has the highest IOU among all predictors in that grid cell). The grid design enforces spatial diversity in bounding box prediction. Often, it's clear which grid cell an object falls into, and the network predicts only one box for each object. However, some large objects or objects near multiple cell boundaries can be well localized by multiple cells. Non-maximum suppression can be used to address these multiple detections.

[0112] Deformable convolution is quite different from regular convolution. In regular convolution, the receptive field is fixed.

[0113] R={(x, y)|x, y∈{-1, 0, 1}};

[0114] However, the features to be collected often have varying shapes, which traditional convolution methods do not handle well. This paper proposes incorporating deformable convolution, modifying the convolution method for the receptive field in YOLO. The specific formula is as follows:

[0115] f (i，j) =∫ p∈R heat((i,j)+p)dp;

[0116] Here, P represents the deformed receptive field, (i, j) represents the specific coordinates, and f(i, j) represents the value of (i, j) in the feature map. However, the representation of R is modified here to make the expansion of the receptive field more random.

[0117] R={(x, y)|x, y∈(-1, 1)};

[0118] By expanding the receptive field, deformable convolution can adapt to different hotspot images, providing greater flexibility for sampling.

[0119] (VI) Hardware Implementation:

[0120] like Figure 4As shown, the system architecture consists of drones, base stations, and a server system. The drones typically fly along a pre-deployed photovoltaic panel platform, collecting data along a pre-set flight path. All data is transmitted to the server via the pre-deployed base stations. The server is equipped with a dedicated model for subsequent predictions.

[0121] Due to the instability of network transmission, especially the potential electromagnetic interference in the high-voltage environment of power plants, image compression to reduce transmission time and the addition of checksum verification for image recognition are essential. The relevant flowchart is as follows: Figure 5 As shown. Figure 5 The diagram illustrates the timing sequence of communication between the unmanned aerial vehicle (UAV) and the server. On the UAV side, there are three phases: Preparation, Photography, and Transmission. On the server side, there are two phases: Preparation and Receiving. During the Preparation phase, the UAV sends a SYN signal to the server via the base station, notifying the server to prepare to receive images. Upon receiving the SYN signal, the server returns an ACK signal to acknowledge receipt of the UAV's SYN signal. Simultaneously, the UAV replies with an ACK signal to indicate that it has received the server's signal indicating readiness to receive images. At this point, mutual confirmation and the activation of the receiving end are complete.

[0122] During the photography phase, the UAV takes pictures of the photovoltaic panels. After shooting, transmission begins. During transmission, the UAV first transmits the compressed image to the server via the base station. Once transmission is complete, the server sends an ACK signal to the UAV to indicate that the image reception process is finished. After the UAV receives the ACK signal from the server, the transmission service shuts down, and subsequent judgment processes are performed on the server side.

Claims

1. A photovoltaic panel anomaly detection method based on a YOLO V8 model, characterized in that Includes the following steps: Step 1: Acquire images of photovoltaic panels using a drone; Step 2: Collect and preprocess images of photovoltaic panels; Step 3: Transmit the photovoltaic panel image data after the preprocessing in Step 2 to the server; Step 4: Set up a prediction model on the server based on the YOLO V8 model to predict the photovoltaic panel images transmitted from the server in Step 3; Step 5: Set up the training model on the server side based on the YOLO V8 model. By using the paths of the positive and negative samples in the photovoltaic panel image dataset, the training model will normalize the positive and negative samples in the photovoltaic panel image dataset and train it through the built-in objective function. The above steps enable the detection of photovoltaic panel anomalies. Step 5 includes the following steps: S5.1: Configure the validation set, training set, and learning rate in the configuration file, and then read the configuration file; S5.2: Set termination conditions; S5.3: Determine if the termination condition has been met. If yes, proceed to S5.4; otherwise, terminate the training. S5.4: Image Segmentation: Divide the image into multiple segmentation units; S5.5: Calculate partition confidence: The formula for confidence prediction is based on the following formula: ； The formula calculates the confidence level of each cell as a certain type of object, and the category of the object with the highest confidence level is taken as the type of the cell. This indicates whether an object exists in the detection box. This indicates that the object is The probability; if an image is the background, then the probability should be 0; This represents the ratio of the interaction area between the true predicted bounding box and the predicted bounding box made using the neural network; ； wherein, denotes the area of intersection of the predicted and the ground truth box, denotes the area of union of the predicted and the ground truth box; For each prediction box, the classification with the highest confidence is calculated, and the final prediction result is obtained based on the classification. Since the hotspot area captured by the drone during detection is relatively small, the IOU coefficient is increased when calculating the detection confidence, as follows: * ； wherein the coefficients adjustment from images acquired on site; S5.6: Calculate the objective function and return to S5.3; The objective function considers the current prediction quality from two aspects: ①. Whether the target exists within the current bounding box; ②. The difference between the center coordinates of the identified box and the coordinates of the data center; the squared error function is used to measure the center error of the bounding box. The objective function is as follows: in: A coefficient representing the bounding box error; This refers to the center point position of the i-th predicted bounding box obtained through neural network calculation; This refers to the position of the center point of the i-th predicted bounding box in the dataset. Let represent the width and height of the i-th predicted bounding box, respectively, predicted by the neural network. These represent the width and height of the predicted bounding boxes labeled in the dataset, respectively; In step 5, the convolution method of the receptive field in the YOLO V8 model is modified. The specific formula is as follows: ； Where: P represents the receptive field of the deformation; (i,j) represents the specific coordinates; This represents the value of (i,j) in the feature map; Receptive fields of sensation The representation is as follows: 。 2. The photovoltaic panel anomaly detection method based on the YOLO V8 model according to claim 1, characterized in that: Step 2, the photovoltaic panel image preprocessing includes the following steps: S2.1: Determine if the photovoltaic panel image is qualified. If yes, proceed to S2.2; otherwise, discard it directly. S2.2: Determine if the image is an abnormal hotspot. If so, mark the abnormal location; otherwise, mark it as a normal image and proceed to S2.

3. S2.3: Place all abnormal images in one path folder and all normal images in another path folder.

Images