An improved YOLOv8-based insulator defect detection method under severe weather conditions

Abstract

This invention relates to the field of artificial intelligence for detecting and identifying defect targets in power systems, and particularly to an improved YOLOv8-based method for insulator defect detection under severe weather conditions. The method includes: acquiring a dataset of insulator defects; using bounding boxes to label the locations of insulators and defects in sample images of the dataset, and indicating the image category as either a normal insulator or a defect; training an improved YOLOv8 model using the dataset; using the trained improved YOLOv8 model to detect insulator defects in the insulator images; and improving the effectiveness and superiority of the algorithm in insulator defect detection tasks.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence for detecting and identifying defect targets in power systems, and in particular to an improved YOLOv8 method for detecting insulator defects under severe weather conditions. Background Technology

[0002] The stable operation of the power system is a key factor in ensuring the stable development of people's lives and production. Insulators are indispensable components of the power system, not only supporting high-voltage transmission lines but also maintaining safe isolation between current and the ground or other conductors. However, insulators exposed to harsh environments for extended periods may suffer from flashover, damage, and other faults. Therefore, rapid and accurate detection is crucial for identifying insulators and their defects, directly impacting the safe and stable operation of the power grid. Traditional insulator inspections typically rely on manual patrols, which are often inefficient and susceptible to human error, making it difficult to ensure the accuracy and consistency of the results.

[0003] Currently, in the field of insulator defect identification, researchers have proposed various innovative methods and technologies to address the issues of detection accuracy and efficiency. The mainstream methods can be divided into two categories: two-stage algorithms and one-stage algorithms. Two-stage algorithms first extract the object region and then classify and identify it. Traditional algorithms include the Fast R-CNN and Faster R-CNN series. These algorithms have high accuracy but require two stages of computation, thus resulting in slower speed. For insulator defect detection, timely elimination of hidden dangers is crucial to ensure the stable operation of the power system; therefore, faster detection speed is essential. One-stage algorithms directly generate candidate bounding boxes on the input image and classify and identify these boxes. Representative algorithms include the YOLO series. This algorithm has a fast detection speed, but its accuracy is not high for blurred images and small targets. Power system inspections must accurately identify insulator defects to achieve the inspection objectives; therefore, further optimization of these algorithms is necessary. Summary of the Invention

[0004] The purpose of this invention is to provide an improved YOLOv8-based insulator defect detection method under severe weather conditions, thereby enhancing the effectiveness and superiority of the algorithm in insulator defect detection tasks.

[0005] The present invention provides a solution using the following technical method:

[0006] In a first aspect, the present invention provides an improved YOLOv8-based method for detecting insulator defects under severe weather conditions, comprising:

[0007] A dataset of insulator defects is obtained; the sample images in the dataset are labeled with bounding boxes indicating the location of insulators and defects, and the image category is indicated as normal insulator or defect;

[0008] The improved YOLOv8 model was trained using the dataset.

[0009] The trained improved YOLOv8 model was used to detect insulator defects in insulator images.

[0010] The improvements of the improved YOLOv8 model are as follows:

[0011] Add a space-to-depth convolutional module (SPDConv) between the convolutional blocks and c2f modules in the backbone network;

[0012] A lightweight 3D attention module (SimAM) is added to the connection between the backbone network and the neck network;

[0013] Replace the CIOU loss function of the baseline YOLOv8 model with an efficient loss function for accurate bounding boxes (Focal and Efficient IOU, Focal_EIOU).

[0014] As a preferred embodiment, the process of acquiring the dataset includes:

[0015] A dataset of insulator defects is obtained, which includes sample images of normal insulators, damaged insulators, and insulator flashovers;

[0016] The sample images are subjected to physical operations according to preset rules, and weather effects are added on the basis of the physical operations to expand the number of sample images in the dataset.

[0017] The sample images in the dataset are preprocessed and labeled, and the image categories are normal insulators or defects.

[0018] As a preferred embodiment, the training process of the improved YOLOv8 model includes:

[0019] The dataset is normalized and divided into training set, validation set and test set according to a preset ratio;

[0020] Set the model parameters;

[0021] The training set is input into the model for forward propagation to output prediction results and model parameters;

[0022] The loss function of the model is calculated based on the prediction results and model parameters;

[0023] The training set is input into the model for backpropagation. The gradient of the error with respect to each model parameter is calculated using the chain rule. The model parameters are then updated using the gradient for optimization.

[0024] Repeat the forward and backward propagation until the loss function converges to obtain the trained improved YOLOv8 model.

[0025] As a preferred embodiment, the step of inputting the training set into the model for forward propagation to output prediction results and model parameters includes:

[0026] The input sample images have features extracted through the backbone network of the improved YOLOv8 model;

[0027] Then, multi-scale features are fused from the neck area to generate an enhanced feature map;

[0028] The enhanced feature map is input into the decoupled head detection branch of the head network, which is divided into a regression branch and a classification branch. The regression branch is used to perform regression prediction on the bounding box, and the classification branch is used to classify the defects to obtain the probability of the type of defect in the bounding box.

[0029] In a preferred embodiment, the spatial-to-depth convolutional module (SPDConv) includes spatial-to-depth (SPD) layers and non-strided-conv layers;

[0030] The scaling factor is set to 2, and the original feature map (S, S, C1), where S represents the height and width and C1 represents the number of channels, is divided along the x and y directions to obtain four sub-feature maps of size (S / 2, S / 2, C1), and a downsampling operation is performed.

[0031] The formula is expressed as follows:

[0032]

[0033] In the formula, the sub-feature map fx,y consists of all i+x and i+y in X(i,j) that are divisible by the scaling factor scale.

[0034] All sub-feature maps are concatenated along the channel dimension to obtain a feature map X'(S / 2, S / 2, 4C1) that is half the size of the original feature map in the spatial dimension and four times larger in the channel dimension. The feature map X' is then mapped and input into a non-staggered convolutional layer with C2 filters to output the feature map X'(S / 2, S / 2, C2).

[0035] As a preferred embodiment, the lightweight 3D attention module (SimAM) includes:

[0036] For the input feature map X, an energy function is defined. Assess the importance of each neuron. The smaller the value, the greater the difference between the target neuron and the surrounding neurons;

[0037] By calculating the mean and variance of neurons, the central tendency and dispersion of their features are quantified, providing a basis for weight generation.

[0038] Using energy function The reciprocal of the formula generates dynamic three-dimensional weights;

[0039] The 3D weights are adapted to the input resolution and fused with the original input features using a weighted method.

[0040] As a preferred embodiment, the energy function The expression is:

[0041]

[0042] In the formula, t represents the target neuron; μ represents the average value of all neurons except t; σ 2 λ represents the variance of all neurons except t; M represents the number of neurons; λ is the regularization term.

[0043] As a preferred embodiment, the expression for the lightweight 3D attention module (SimAM) is:

[0044]

[0045] Where E represents all neurons in the input feature map. A set of values, with a sigmoid function added to limit values ​​in E that are too large.

[0046] As a preferred embodiment, the efficient loss function (Focal and Efficient IOU, Focal_EIOU) is defined as follows:

[0047]

[0048] L Focal-EIOU =IOU γ L EIOU

[0049]

[0050] Where, ρ 2 (b,b gt ) represents the distance between the center point of the predicted bounding box and the center point of the ground truth bounding box, c is the diagonal distance between the minimum bounding enclosed region of the predicted bounding box and the ground truth bounding box, and the parameter w gth gt w and h represent the width and height of the ground truth bounding box and the width and height of the predicted bounding box of the detected object, respectively. A and B represent the volumes of the two arbitrary shapes being measured, and γ is a parameter that controls the degree of outlier suppression.

[0051] Secondly, the present invention provides an insulator defect detection system based on an improved YOLOv8 for use in severe weather conditions, comprising:

[0052] The acquisition module is used to acquire a dataset of insulator defects; the sample images in the dataset are marked with bounding boxes indicating the location of insulators and defects, and the image category is indicated as normal insulator or defect;

[0053] The training module is used to train the improved YOLOv8 model using the dataset.

[0054] The improvements of the improved YOLOv8 model are as follows:

[0055] Add a space-to-depth convolutional module (SPDConv) between the convolutional blocks and c2f modules in the backbone network;

[0056] A lightweight 3D attention module (SimAM) is added after the backbone network;

[0057] Replace the CIOU loss function of the baseline YOLOv8 model with an efficient loss function for accurate bounding boxes (Focal and Efficient IOU, Focal_EIOU).

[0058] The detection module is used to detect insulator defects in insulator images using a trained improved YOLOv8 model.

[0059] Compared with the prior art, the present invention has the following beneficial effects:

[0060] This invention utilizes deep learning networks for object detection, addressing the shortcomings of traditional detection methods such as lack of real-time performance, low detection accuracy, and low efficiency. The invention adds a SimAM attention mechanism to the connection parts of the backbone and neck region to extract more accurate and discriminative multi-scale spatial information features, thereby improving the overall performance of the model. Furthermore, this invention adds a spatial-to-depth convolutional module (SPDConv) to the backbone network to address the lack of fine-grained information, aiming to improve the detection capability for small targets and low-resolution images. Finally, this invention employs an efficient loss function with accurate bounding boxes to effectively measure the difference between the target box and the ground truth box, improving the detection performance for small targets. Attached Figure Description

[0061] Figure 1 This is a flowchart of an insulator defect detection method according to an embodiment of the present invention;

[0062] Figure 2 This is a model structure diagram of the improved YOLOv8 according to an embodiment of the present invention;

[0063] Figure 3 This is a structural diagram of the SimAM attention mechanism according to an embodiment of the present invention;

[0064] Figure 4 This is a structural diagram of the SPDConv convolutional module according to an embodiment of the present invention;

[0065] Figure 5 The image shows the detection results of the improved YOLOv8 model in this embodiment of the invention. Detailed Implementation

[0066] It should be noted that the technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments and specific features described herein are detailed descriptions of the technical solution of the present invention, not limitations thereof. Where there is no conflict, the embodiments and technical features described herein can be combined with each other. The term "and / or" merely describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0067] To make the purpose, technical solution, and advantages of this invention patent clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0068] A method for detecting insulator defects under severe weather conditions based on an improved YOLOv8, such as... Figure 1 As shown, it includes:

[0069] S1. Obtain the insulator defect dataset;

[0070] In this invention, the insulator defect dataset is obtained from the "Insulator Defect Image Dataset (IDID)", which contains 487 original images, including three categories: normal insulator, damaged insulator, and flashover insulator.

[0071] S2. Perform data augmentation on the insulator defect dataset;

[0072] This invention addresses the problem of insufficient data in the insulator defect dataset by expanding the dataset through operations such as brightness enhancement, cropping, and rotation. Based on the expanded dataset, an algorithm is used to add severe weather conditions such as rain, snow, and fog, and corresponding label samples are determined.

[0073] S3. Perform data preprocessing and annotation on the insulator defect image;

[0074] The images were labeled using the Labellim image annotation tool, and an XML file was generated after annotation. The corresponding XML file was then converted into a TXT file. The dataset was then divided into training, testing, and validation sets in an 8:1:1 ratio.

[0075] S4. Improve the YOLOv8 model;

[0076] The improved YOLOv8 model described in this invention specifically makes the following three improvements to the model, and the network structure diagram based on the improved YOLOv8 is as follows: Figure 2 As shown, a lightweight 3D attention module (SimAM) requiring no additional parameters is added to the connection between the backbone and neck network of the baseline YOLOv8 model; a spatial-to-depth convolutional module (SPDConv) is added to the backbone of the baseline YOLOv8 model to address the problem of missing fine-grained information; and the CIOU loss function of the baseline YOLOv8 model is replaced with an efficient loss function for accurate bounding boxes. Specific improvements are as follows:

[0077] (1) Add a lightweight 3D attention module (Simple, Parameter-Free Attention, SimAM) that requires no additional parameters after the backbone of the baseline model YOLOv8:

[0078] SimAM's attention mechanism enables the collaborative work of channel attention and spatial attention, making the model more coordinated and efficient in task processing without adding extra task parameters. By dynamically adjusting the weights of pixels in the feature map, it performs fine-grained weight allocation on deep semantic features, highlighting the target region and reducing background noise interference.

[0079] In this invention, the network structure of the SimAM attention mechanism is as follows: Figure 3 As shown, H, W, and C represent the height, width, and number of channels of the image, respectively; X is the input feature map. The SimAM attention mechanism defines an energy function... The importance of each neuron is assessed using the following mathematical expression:

[0080]

[0081] In the formula, t represents the target neuron; μ represents the average value of all neurons except t; σ 2 λ represents the variance of all neurons except t; M represents the number of neurons; λ is the regularization term.

[0082] when The smaller the value, the greater the difference between the neuron and its surrounding neurons, and the more important it is for visual processing. Therefore, the importance of each neuron is represented by the reciprocal of the minimum energy; the larger the reciprocal, the higher the importance. The final SimAM attention module expression is as follows:

[0083]

[0084] E represents all neurons in the input feature map. The set of values, and the sigmoid is added to limit the value in E to be too large.

[0085] Furthermore, by calculating the mean and variance of neurons to quantify the central tendency and dispersion of features, a basis for weight generation is provided. Dynamic three-dimensional weights are generated using the reciprocal of the energy function, corresponding to... Figure 3 The generation stage in the middle; then the weights are adapted to the input resolution and fused with the original features in a weighted manner, corresponding to Figure 3 The integration phase in the process.

[0086] (2) Add a spatial-to-depth convolutional module (SPDConv) to the backbone of the baseline model YOLOv8:

[0087] SPDConv consists of spatial-to-depth (SPD) layers and non-strided-Conv layers. By decomposing the standard convolution operation, it breaks down the spatial convolution into multiple convolutions with smaller kernels, thereby reducing computational complexity. This decomposition method preserves the spatial receptive field of the original convolution and enhances the model's ability to extract features from small objects.

[0088] In this invention, the SPDConv structure diagram is as follows: Figure 4 As shown:

[0089] With scale=2, the original feature map (S, S, C1), where S represents height and width, and C1 represents the number of channels, is segmented along the x and y directions to obtain four sub-feature maps of size (S / 2, S / 2, C1), and a downsampling operation is performed. All sub-feature maps are then concatenated along the channel dimension to obtain a feature map X'(S / 2, S / 2, 4C1), which is half the size of the original feature map in spatial dimension and four times the size in channel dimension. The Conca operation is performed on this feature map, and after inputting this feature map into a non-staggered convolutional layer with C2 filters, the output feature map X'(S / 2, S / 2, C2) is subjected to the Conv operation. The basic formula is as follows:

[0090]

[0091] In the formula, the subgraph fx,y consists of all i+x and i+y in X(i,j) that are divisible by the scale. The SPDConv convolutional module with scale=2 is introduced into the original network model to enhance the feature extraction capability for low-resolution images and small targets under severe weather conditions, and improve the recognition rate of insulator defects.

[0092] (3) Replace the CIOU loss function of the baseline YOLOv8 model with an efficient loss function for accurate bounding boxes (Focal and Efficient IOU, Focal_EIOU):

[0093] The Focal_EIOU loss function is defined as follows:

[0094]

[0095] L Focal-EIOU =IOU γ L EIOU

[0096]

[0097] Where ρ 2 (b,b gt ) represents the distance between the center point of the predicted bounding box and the center point of the ground truth bounding box, c is the diagonal distance between the predicted bounding box and the minimum bounding enclosed region of the ground truth bounding box, and the parameter w gt ,h gt w and h represent the width and height of the ground truth bounding box and the width and height of the predicted bounding box, respectively; A and B represent the volumes of the two arbitrary shapes being measured; and γ is a parameter controlling the degree of outlier suppression. Furthermore, EIOU explicitly measures the difference in three geometric factors within the bounding box: the overlap area, the center point, and the side length. This solves the gradient vanishing problem in IoU loss when the predicted and ground truth bounding boxes do not overlap, which slows down convergence and leads to detector inaccuracies.

[0098] Furthermore, the training set of the pre-divided insulator defect dataset is input into the improved YOLOv8 network model for training. Image data features are extracted through the backbone network, and then multi-scale features are fused through the neck layer to generate enhanced feature maps. These feature maps are input into the decoupled head detection branch of the head network, which is divided into regression and classification branches. The regression branch is used to predict the bounding boxes, and the classification branch is used to classify the defects, obtaining the probability of the defect type within the bounding boxes. The loss value between the predicted and true values ​​is calculated based on the loss function, and then the loss value is reduced through backpropagation and optimization algorithms, i.e., the loss function converges, at which point the model training ends.

[0099] S5. Training in the improved YOLOv8 model, the training process includes;

[0100] (1) Configure the model training environment: The experimental environment of this invention is configured on the Linux Ubuntu operating system, where the GPU version is NVIDIA GeForce RTX 4090 and PyTorch is used as the deep learning framework.

[0101] (2) Set model parameters: The input image resolution is 640×640 pixels, the batch_size of the samples selected in each round is set to 16; the SGD optimizer is selected, the initial learning rate is 0.001, and the training rounds are 200.

[0102] (3) Model training: Input the constructed insulator defect dataset into the deep network. The data includes defect images, defect labels and model configuration files.

[0103] (4) Evaluation metrics: The weight file trained by the improved YOLOv8 model is used to validate the divided validation set. Average precision (AP), recall (R), precision (P), and mean average precision (mAP) are used as evaluation metrics for the network model performance. The evaluation metrics are defined as follows:

[0104]

[0105] Where TP is the number of correctly detected samples; FP is the number of incorrectly detected samples; FN is the number of missed samples; AP is the area enclosed by the PR curve and the coordinate axes; mAP is the average AP of all categories, with mAP@0.5 (average precision calculated when the IoU threshold is 0.5) and mAP@0.5:0.95 (average precision calculated when the IoU threshold is between 0.5 and 0.95, with a value interval of 0.05); N is the number of categories;

[0106] (5) Prediction Results: Finally, the trained weight file is used to test the divided test set to obtain the actual insulator defect prediction results, such as... Figure 5 As shown:

[0107] Subgraphs (a), (b), (c), and (d) represent YOLOv8 detection under normal, rain, snow, and fog conditions, respectively, while (a1), (b1), (c1), and (d1) represent the detection results of the improved YOLOv8 model under these conditions. The figures show that the original YOLOv8 model has a problem of missing detections under both normal and severe weather conditions, while the improved YOLOv8 model can detect all insulator defects. Furthermore, it is clearly evident that the improved YOLOv8 model has a significantly higher accuracy in detecting insulator defects than the original YOLOv8 model.

[0108] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0109] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0110] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0111] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0112] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A method for detecting insulator defects under severe weather conditions based on an improved YOLOv8, characterized in that, include: A dataset of insulator defects is obtained; the sample images in the dataset are labeled with bounding boxes indicating the location of insulators and defects, and the image category is indicated as normal insulator or defect; The improved YOLOv8 model was trained using the dataset. The trained improved YOLOv8 model was used to detect insulator defects in insulator images. The improvements of the improved YOLOv8 model are as follows: Add a spatial-to-depth convolutional module (SPDConv) between the convolutional blocks and c2f modules in the backbone network. A lightweight 3D attention module (SimAM) is added to the connection between the backbone network and the neck network; Replace the CIOU loss function of the baseline YOLOv8 model with an efficient loss function for accurate bounding boxes (Focal and Efficient IOU, Focal_EIOU); The spatial-to-depth convolutional module (SPDConv) includes spatial-to-depth (SPD) layers and non-strided-conv layers; The scaling factor is set to 2, and the original feature map (S, S, C1), where S represents the height and width and C1 represents the number of channels, is divided along the x and y directions to obtain four sub-feature maps of size (S / 2, S / 2, C1), and a downsampling operation is performed. The formula is expressed as follows: In the formula, the sub-feature maps fx,y are derived from... It consists of all i+x and i+y that are divisible by the scaling factor; All sub-feature maps are concatenated along the channel dimension to obtain a feature map X'(S / 2, S / 2, 4C1) that is half the size of the original feature map in the spatial dimension and four times larger in the channel dimension. The feature map X' is then mapped and input into a non-staggered convolutional layer with C2 filters to output the feature map X''(S / 2, S / 2, C2). The lightweight 3D attention module (SimAM) includes: For the input feature map X, an energy function is defined. Assess the importance of each neuron. The smaller the value, the greater the difference between the target neuron and the surrounding neurons; By calculating the mean and variance of neurons, the central tendency and dispersion of their features are quantified, providing a basis for weight generation. Using energy function The reciprocal of the formula generates dynamic three-dimensional weights; The three-dimensional weights are adapted to the input resolution and fused with the original input features using a weighted method. The energy function The expression is: In the formula, t represents the target neuron; This represents the average value of all neurons except t; represents the variance of all neurons except t; M represents the number of neurons; It is a regular term; The expression for the lightweight 3D attention module (SimAM) is: sigmoid( ) Where E represents all neurons in the input feature map. A set of values; The efficient loss function (Focal and Efficient IOU, Focal_EIOU) is defined as follows: in, The parameter represents the distance between the center point of the predicted bounding box and the center point of the ground truth bounding box; c is the diagonal distance between the minimum bounding enclosed region of the predicted bounding box and the ground truth bounding box; , w and h represent the width and height of the ground truth bounding box and the width and height of the predicted bounding box of the detected object, respectively; A and B represent the volumes of the two arbitrary shapes being measured, respectively. Parameters used to control the degree of outlier suppression.

2. The insulator defect detection method based on improved YOLOv8 under severe weather conditions according to claim 1, characterized in that, The process of acquiring the dataset includes: A dataset of insulator defects is obtained, which includes sample images of normal insulators, damaged insulators, and insulator flashovers; The sample images are subjected to physical operations according to preset rules, and weather effects are added on the basis of the physical operations to expand the number of sample images in the dataset. The sample images in the dataset are preprocessed and labeled, and the image categories are normal insulators or defects.

3. The insulator defect detection method based on improved YOLOv8 under severe weather conditions according to claim 1, characterized in that, The training process of the improved YOLOv8 model includes: The dataset is normalized and divided into training set, validation set and test set according to a preset ratio; Set the model parameters; The training set is input into the model for forward propagation to output prediction results and model parameters; The loss function of the model is calculated based on the prediction results and model parameters; The training set is input into the model for backpropagation. The gradient of the error with respect to each model parameter is calculated using the chain rule. The model parameters are then updated using the gradient for optimization. Repeat the forward and backward propagation until the loss function converges to obtain the trained improved YOLOv8 model.

4. The insulator defect detection method based on improved YOLOv8 under severe weather conditions according to claim 3, characterized in that, The step of inputting the training set into the model for forward propagation to output prediction results and model parameters includes: The input sample images have features extracted through the backbone network of the improved YOLOv8 model; Then, multi-scale features are fused from the neck area to generate an enhanced feature map; The enhanced feature map is input into the decoupled head detection branch of the head network, which is divided into a regression branch and a classification branch. The regression branch is used to perform regression prediction on the bounding box, and the classification branch is used to classify the defects to obtain the probability of the type of defect in the bounding box.

5. An improved YOLOv8-based insulator defect detection system under severe weather conditions, characterized in that it is used to perform the improved YOLOv8-based insulator defect detection method under severe weather conditions according to any one of claims 1-4, the system comprising: The acquisition module is used to acquire a dataset of insulator defects; the sample images in the dataset are marked with bounding boxes indicating the location of insulators and defects, and the image category is indicated as normal insulator or defect; The training module is used to train the improved YOLOv8 model using the dataset. The improvements of the improved YOLOv8 model are as follows: Add a spatial-to-depth convolutional module (SPDConv) between the convolutional blocks and c2f modules in the backbone network. A lightweight 3D attention module (SimAM) is added after the backbone network; Replace the CIOU loss function of the baseline YOLOv8 model with an efficient loss function for accurate bounding boxes (Focal and Efficient IOU, Focal_EIOU); The detection module is used to detect insulator defects in insulator images using a trained improved YOLOv8 model.

Images