Power transmission line defect detection method based on improved YOLOv5 and fuzzy image enhancement

Abstract

The application provides a power transmission line defect detection method based on improved YOLOv5 and fuzzy image enhancement, comprising the following steps: constructing a defect detection model based on improved YOLOv5; constructing a fuzzy image enhancement algorithm based on a generative adversarial network; collecting power transmission line image data in multiple scenes and performing pretreatment, and using the pretreated image data to train the generative adversarial network in the fuzzy image enhancement algorithm and the defect detection model; using the trained fuzzy image enhancement algorithm and defect detection model to perform frame-by-frame detection on the collected power transmission line inspection video, and first reconstructing the image through the fuzzy image enhancement algorithm and then performing defect detection through the defect detection model. The application combines the improved YOLOv5 model and the fuzzy image enhancement technology to perform power transmission line defect detection, improves the defect detection precision of the model on small-size components of the power transmission line, and can perform multi-scale target detection and fuzzy target detection in the field of power transmission line defect detection.

Description

Technical Field

[0001] This invention belongs to the field of power transmission line inspection technology, specifically relating to a method for detecting defects in power transmission lines based on improved YOLOv5 and fuzzy image enhancement. Background Technology

[0002] Transmission line inspection is a core task in preventing power outages and ensuring the safe and stable operation of the power grid. Through inspections, hidden dangers are eliminated promptly, ensuring electricity supply for people's lives and production. In recent years, traditional manual power line inspection methods have become insufficient to meet the ever-changing and growing demands of power line inspection operations due to low efficiency, inconsistent inspection quality, limited scope, and inherent risks. Against this backdrop, drone-based power line inspection can be conducted 24 / 7, offers diverse inspection methods, and overcomes many shortcomings of manual inspection, ensuring the safe and stable operation of transmission lines and other critical power equipment and facilities with higher inspection efficiency. During drone-based power line inspections, the images of the transmission line scene captured by the drone provide input for algorithms that identify key components and detect defects. These algorithms then identify key power components and fault locations within the images.

[0003] With the continuous iteration of deep learning methods, algorithms for identifying key components and detecting defects in power transmission lines using UAVs for power line inspection are gradually shifting from traditional image processing methods to target detection models based on convolutional neural networks. Currently, target detection models have achieved significant defect detection results on large components such as insulators and vibration dampers, as well as foreign objects on towers. However, the difficulty in locating and identifying defects in small components such as pins and screws remains high. This is mainly due to the small size of the targets and the fuzziness of defect features: on the one hand, in many commonly used target detection models, small-sized defects are easily eroded in the convolution process; on the other hand, during image acquisition by the UAV, factors such as motion posture and camera blurring cause blurring and distortion in the acquired inspection images, resulting in blurred defect features and leading to false positives or false negatives. This phenomenon is even more severe in small-sized components.

[0004] To address the two major technical challenges mentioned above, researchers in the field of power transmission line inspection often employ data augmentation for model optimization. This primarily involves random image cropping and scaling, graphical deformation, filtering and blurring, and adding noise to the training data during the model training phase. This effectively amplifies the location and semantic information of defect features in small-sized components within the dataset without affecting the labels, and also improves robustness for detecting blurred features and noisy targets. However, due to the randomness of data augmentation methods, the optimization effect on the model cannot be guaranteed, and blurred features are difficult to directly simulate using simple filtering and blurring operators. Summary of the Invention

[0005] In view of this, the present invention aims to solve the problem that existing transmission line inspection methods are ineffective in detecting defects in small-sized components.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] A transmission line defect detection method based on improved YOLOv5 and fuzzy image enhancement includes the following steps:

[0008] A defect detection model based on an improved YOLOv5 is constructed. The defect detection model is based on the YOLOv5 network structure. A shallow prediction branch for detecting small targets is added to the detection head network. The shallow prediction branch is combined with the original three detection branches to form a four-branch detection model.

[0009] Construct a blurred image enhancement algorithm based on generative adversarial networks;

[0010] Image data of power transmission lines in multiple scenarios are collected and preprocessed. The preprocessed image data is then used to train the generative adversarial network and defect detection model in the blurred image enhancement algorithm.

[0011] The trained fuzzy image enhancement algorithm and defect detection model are used to perform frame-by-frame detection on the collected power transmission line inspection video. Each frame of the image is first reconstructed by the fuzzy image enhancement algorithm and then the defect detection model is used to detect defects.

[0012] Furthermore, the defect detection model based on the improved YOLOv5 also includes replacing the Trans module at the end of the backbone network of the original YOLOv5 model and the Trans module after the Concat module in the feature fusion network with a Transformer encoder.

[0013] Furthermore, the defect detection model based on the improved YOLOv5 also includes adding a CBAM attention module before the Conv module in the feature fusion network of the original YOLOv5 model. The CBAM attention module is used to perform attention mapping inference sequentially along two independent dimensions, the channel dimension and the spatial dimension, and multiply the two attention mapping maps with the backbone feature map respectively.

[0014] Furthermore, the target confidence loss of the defect detection model is calculated according to the following formula:

[0015]

[0016] In the formula, L obj For target confidence loss, β1, β2, β3, and β4 are the losses for feature maps of different sizes mapped to the four detection branches, respectively, and are the weight factors of the four detection branches.

[0017] Furthermore, β1, β2, β3 and β4 are set to 4.0, 1.0, 0.4 and 0.2, respectively.

[0018] Furthermore, in the blurry image enhancement algorithm based on generative adversarial networks, a generative adversarial network is constructed using the RRDB network and the U-NET discriminant network.

[0019] Furthermore, in the blurry image enhancement algorithm based on generative adversarial networks, a second-order image degradation method is used to simulate the generation of blurry images. The second-order image degradation method models the image through two repeated degradation processes to generate blurry images.

[0020] Further, the degradation process specifically includes:

[0021] The image is convolved using a fuzzy kernel function;

[0022] Image resizing is performed through downsampling;

[0023] Different types of noise are randomly added to the image;

[0024] Perform JPEG compression on the image.

[0025] Furthermore, in the blurred image enhancement algorithm, the loss function includes adversarial loss function, content loss function, and edge loss function. The calculation formulas for each loss function are as follows:

[0026] Adversarial loss function:

[0027]

[0028] In the formula, l adv Indicating resistance to loss, and These are the discrimination results of the generated image and the real image under the U-NET discriminant network, respectively, where N is the number of images;

[0029] Content loss function:

[0030]

[0031] In the formula, l MSE The content loss is represented by r, W, and H, which are the number of images, width, and height, respectively, and x and y are the pixel coordinates. Represents the original image, I LR The image represented by G(I) is a blurred image generated using a second-order image degradation method. LR ) x，yRepresenting I LR The generated image after inputting into the RRDB network;

[0032] Marginal loss function:

[0033]

[0034] In the formula, l Edge The edge loss is represented by W and H, which represent the width and height of the image, and E represents the image edge features calculated by the Canny operator.

[0035] Furthermore, the training of the blurred image enhancement algorithm adopts a stochastic gradient descent optimizer as the network training optimization strategy, while the defect detection model adopts an adaptive moment estimation optimizer as the network training optimization strategy.

[0036] In summary, this invention provides a method for detecting transmission line defects based on improved YOLOv5 and fuzzy image enhancement. The method includes constructing a defect detection model based on improved YOLOv5; constructing a fuzzy image enhancement algorithm based on generative adversarial networks (GANs); acquiring transmission line image data from multiple scenarios and preprocessing it; using the preprocessed image data to train the GAN and defect detection model in the fuzzy image enhancement algorithm; and using the trained fuzzy image enhancement algorithm and defect detection model to perform frame-by-frame detection on the acquired transmission line inspection video. Each frame is first reconstructed using the fuzzy image enhancement algorithm and then subjected to defect detection by the defect detection model. This invention combines an improved YOLOv5 model and fuzzy image enhancement technology for transmission line defect detection, which not only improves the model's accuracy in detecting defects in small-sized components of transmission lines but also reduces false positives and false negatives caused by image blurring. The proposed method can perform multi-scale target detection and fuzzy target detection in the field of transmission line defect detection, exhibiting high robustness and practicality. Attached Figure Description

[0037] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0038] Figure 1 A flowchart illustrating a transmission line defect detection method based on improved YOLOv5 and blurred image enhancement provided in an embodiment of the present invention;

[0039] Figure 2 This is a schematic diagram of the structure of the defect detection model provided in an embodiment of the present invention;

[0040] Figure 3 This is a schematic diagram of the RRDB generation network provided in an embodiment of the present invention;

[0041] Figure 4 This is a flowchart illustrating the second-order image degradation method provided in an embodiment of the present invention. Detailed Implementation

[0042] To make the objectives, features, and advantages of this invention more apparent and understandable, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described below are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0043] This embodiment provides a transmission line defect detection method based on improved YOLOv5 and blurred image enhancement. In terms of model structure optimization, based on the YOLOv5 model, a prediction branch specifically for detecting small targets is added to mitigate the information erosion effect of interpolation on small-sized features during convolution. The Transformer encoder and CBAM attention module are integrated into the model to improve overall detection accuracy. Regarding blurred image enhancement, a blurred image enhancement algorithm based on generative adversarial networks (GANs) is proposed to reconstruct the image to be detected into a high-resolution, clear image, reducing false positives and false negatives caused by blurred images. The GAN is constructed using an RRDB generator network and a U-NET discriminator network. A second-order image degradation method is designed to simulate and generate more realistic blurred images, providing high-quality training data for network training and enhancing algorithm performance. Figure 1 As shown, the method specifically includes the following steps:

[0044] S1: Build an improved YOLOv5 model.

[0045] The YOLOv5 model is a highly efficient single-stage target detection model. In order to improve the model's ability to detect defects in small-sized components in the context of power transmission lines, this embodiment improves upon the YOLOv5 model by adding a shallow prediction branch to the detection head network.

[0046] Improve the overall structure of the YOLOv5 model as follows: Figure 2As shown, the system consists of three main parts: a backbone network, a feature fusion network, and a detection head network. The backbone network is based on the CSPDarknet53 network structure and employs the Cross-Stage Local Network (CSP) approach to generate feature layers of different scales from the input image, extracting rich information features. The feature fusion network combines a Feature Pyramid Network (FPN) and a Path Aggregation Network (PANet). For the multi-scale feature layers output by the backbone network, strong semantic features are first conveyed through the deep-to-shallow path of the FPN, and then strong localization features are conveyed through the shallow-to-deep network of PANet, thus achieving the fusion of shallow localization information and deep semantic information at different scales. The detection head network is responsible for feature mapping of the fused feature map and outputting the location and category information of the defect target in the image.

[0047] Reducing the number of network layers means that shallow features are better preserved, which is precisely the focus of small target detection. Furthermore, shallow feature maps have shorter feature strides and denser detection bounding boxes, enabling better localization of small targets. Based on this, this embodiment adds a dedicated shallow prediction branch for detecting small targets (branch 1 in the detection head network region of the figure) to the original YOLOv5 model's detection head network. Combined with the original three detection branches, this forms a four-branch detection model. This effectively mitigates the negative impact of interpolation factors during brute-force image scaling and scale changes, while also enabling more precise localization of small targets in the image.

[0048] S2: Design a blurry image enhancement algorithm based on generative adversarial networks.

[0049] In recent years, generative adversarial networks have been widely used in the field of image enhancement. By training the network with images of the same scene (low-quality and high-quality), the network learns the mapping relationship between low-quality and high-quality images.

[0050] S3: Train the improved YOLOv5 model and generative adversarial network.

[0051] This step involves training the generative adversarial network in the blurred image enhancement algorithm and improving the YOLOv5 model.

[0052] First, image data from multiple scenarios is collected by inspecting power transmission lines using drones. Data with abnormal brightness, high noise, and blurry images is then manually sorted and cleaned. The image data is then divided into training and testing sets in a 4:1 ratio.

[0053] Image data from multiple scenarios is collected by inspecting power transmission lines using drones. Data with abnormal brightness, high noise, and blurry images is cleaned by manual sorting. The image data is then divided into training and test sets in a 4:1 ratio.

[0054] Then, image data from multiple scenarios was collected by inspecting power transmission lines using drones. Data with abnormal brightness, high noise, and blurry images was manually sorted and cleaned. Annotation tools were used to label power components and defects in the images, and the data was saved as XML tag files in VOC data format. Power components mainly include insulators, vibration dampers, suspension clamps, bolts, and other components of various sizes. Defects include insulator explosions, vibration damper detachment, vibration damper displacement, missing pins, and missing nuts. The image data and corresponding annotation files were divided into training and testing sets at a 4:1 ratio. Multiple training iterations were conducted to obtain weight parameter files for several models. The accuracy of models using different weight parameter files was compared on the testing set, and the model with the highest accuracy was selected as the final model.

[0055] S4: Enables defect detection in power transmission lines.

[0056] The trained improved YOLOv5 model and blurred image enhancement algorithm were deployed on the computing platform.

[0057] When detecting defects in power transmission lines, the inspection videos of power transmission lines collected by drones are transmitted to the computing platform via the network. The platform then starts an improved YOLOv5 model to detect defects frame by frame. Each frame of the video is first processed by a blur image enhancement algorithm to generate a high-resolution and clear image, which is then input into the improved YOLOv5 model to detect defects. If a defect target is detected in the image, the location and category information of the defect in the image are marked. After the video detection is completed, a visualization report is generated so that power industry professionals can analyze and process the line.

[0058] This embodiment provides a method for detecting transmission line defects based on improved YOLOv5 and fuzzy image enhancement. This method combines an improved YOLOv5 model and fuzzy image enhancement technology to detect transmission line defects. It not only improves the model's accuracy in detecting defects in small-sized components of transmission lines, but also reduces false detections and missed detections caused by image fuzziness. The proposed method can perform multi-scale target detection and fuzzy target detection in the field of transmission line defect detection, and has high robustness and high practicality.

[0059] In an optional embodiment, the improved YOLOv5 model further includes model optimization based on a Transformer encoder. In this embodiment, some convolutional modules at the ends of the backbone network and in the feature fusion network of the original YOLOv5 model are replaced with Transformer encoder modules. Figure 2The Trans encoder module in the backbone network has a low resolution feature map at the end. Applying the Transformer encoder module to the low-resolution feature map can reduce computational cost, improve computational efficiency, and compress the model size. Applying the Transformer encoder module in the feature fusion network is beneficial for capturing global information during feature fusion, obtaining richer contextual information, and improving the overall detection accuracy of the model.

[0060] In an optional embodiment, the improved YOLOv5 model also includes model optimization based on the CBAM attention module. The CBAM attention module infers attention maps sequentially along two independent dimensions—channel and spatial—using feature mapping, and multiplies the two attention map spectra separately with the backbone feature map, achieving adaptive feature refinement. This embodiment introduces the CBAM attention module into the feature fusion network to extract attention regions, helping the model distinguish cluttered background elements and making the fused feature map more focused on small targets that need to be detected.

[0061] In an optional embodiment, since a new prediction branch is added to the improved YOLOv5 model, the loss function needs to be redesigned so that the model can use the information from the new branch's prediction results during training, thus improving model training. This embodiment modifies the target confidence loss L in the original YOLOv5 model's loss function. obj Redesigned and improved L obj The specific calculation formula is as follows. The losses correspond to the small, medium-small, medium, and large-sized feature maps mapped by branches 1, 2, 3, and 4, respectively; β1, β2, β3, and β4 are weighting factors. In order to increase the importance of small and medium-sized targets during model training, in this embodiment, the four factors are set to 4.0, 1.0, 0.4, and 0.2, respectively.

[0062]

[0063] In an optional embodiment, the blurry image enhancement algorithm based on generative adversarial networks (GANs) first constructs a GAN using an RRDB generative network and a U-NET discriminative network, and defines the loss function for the overall network. Then, a second-order image degradation method is used to simulate and generate more realistic blurry images, providing high-quality training data for network training and improving the network's convergence performance. The blurry image enhancement algorithm is used in the defect detection process to reconstruct the image to be detected into a high-resolution, clear image, reducing false positives and false negatives caused by blurry images.

[0064] Traditional generative adversarial networks (GANs) use ResNet as the generator network to produce high-resolution images from input images, and U-NET as the discriminator network to calculate the differences between the original high-resolution image and the generated image at both the global and local pixel levels, thus providing direction for the generator network to produce higher-quality images.

[0065] The performance of the generative network is related to whether it can correctly learn the mapping relationship between low-quality and high-quality images. In order to improve the ability of the generative network to generate detailed features, this invention uses the higher-performance RRDB network to replace the ResNet network and constructs a generative adversarial network based on the RRDB generative network and the U-NET discriminative network.

[0066] RRDB generates a network structure as follows: Figure 3 As shown, the RRDB generative network consists of three parts. The first part includes a 7×7 and two 3×3 convolutional blocks, where the input image passes through the convolutional layers to obtain shallow features. The middle part uses nine RRDB network units connected to each other via skip connections, allowing the model to fully learn the deep features of the image. The third part consists of two 3×3 transposed convolutional blocks and a 7×7 convolutional layer, which recovers the image and increases its detailed features through convolution, activation, and upsampling. Furthermore, global skip connections directly connect the input image to the output, reducing redundant computations in the network.

[0067] During the process of collecting images of power transmission lines by drones and inputting the final images into the target detection model for defect detection, the images usually undergo two stages of image degradation, resulting in blurry images. The first degradation is due to motion blur, sensor noise, and errors in the camera imaging system when the gimbal camera collects the images; the second degradation is due to quality degradation and JPEG compression that occur during image transmission.

[0068] To address the real-world phenomenon of image degradation, this invention designs a second-order image degradation method to simulate and generate more realistic blurred images. The image degradation process is as follows: Figure 4 As shown, image modeling is performed through two repeated degradation processes. Each degradation process includes four operations: (1) convolving the image with a fuzzy kernel function; (2) resizing the image by downsampling; (3) randomly adding different types of noise to the image; and (4) JPEG compression of the image.

[0069] Before training a generative adversarial network (GAN), a loss function needs to be designed to calculate the loss value, reflecting the difference between the model's generated images and real high-resolution images. The network parameters are then updated through backpropagation. The overall network loss function consists of three parts: an adversarial loss function, a content loss function, and an edge loss function.

[0070] Combat loss adv Calculated using the following formula. Where, and The images are the discrimination results of the generated image and the real high-resolution image under the U-NET discriminant network, respectively, where N is the number of images:

[0071]

[0072] Content loss MSE The design is based on the mean squared error function, and the formula is as follows. Where r, W, and H describe the number, width, and height of the image, respectively, and x and y are the pixel coordinates. Represents the original high-resolution image, I LR The image represented by G(I) is a blurred image generated using a second-order image degradation method. LR ) x，y Representing I LR The generated image after inputting into the RRDB network:

[0073]

[0074] The edge loss function assigns extra weight to edge loss to enhance the detail information of the generated image. By optimizing the edge pixels of the image, it drives the trained network to generate high-quality, salient edges, resulting in a sharper image. Edge loss l Edge First, the Canny operator is used to detect the edge features of the original high-resolution image and the generated image. Then, the difference between the image edge features is calculated using the following formula. Where W and H represent the width and height of the image, and E represents the image edge features calculated by the Canny operator:

[0075]

[0076] Based on the above design, when training the generative adversarial network (GAN) in the image enhancement algorithm, a second-order image degradation method is used to simulate the generation of blurred images. These blurred images are then combined with the original high-resolution images to form a [low-quality-high-quality] image for training the GAN. The model training employs a stochastic gradient descent optimizer as the network's training optimization strategy. The RRDB generative network and the U-NET discriminative network are cross-trained for 100 epochs; that is, the RRDB network and the U-NET network are trained separately in each epoch. The weight parameters of the U-NET network are frozen during RRDB network training, and vice versa. The learning rate is dynamically adjusted starting from 0.01, and multiplied by 0.1 after the 40th, 60th, and 80th epochs. The network loss is calculated after each epoch, and training stops when the loss value does not decrease for five consecutive epochs.

[0077] The improved YOLOv5 model training employs an adaptive moment estimation optimizer as the network's training optimization strategy, with a total of 50 training epochs. The learning rate is dynamically adjusted starting from 0.0003, and multiplied by 0.9 after each epoch update. For the first 25 epochs, the weight parameters of the backbone network and feature fusion network are frozen, and only the unfrozen network layers are fine-tuned for gradient updates during backpropagation; the last 25 epochs train the entire model. The model loss is calculated after each training epoch, and training stops when the loss value does not decrease for five consecutive epochs.

[0078] Compared with the prior art, the present invention has the following advantages:

[0079] 1. Combining the improved YOLOv5 model and fuzzy image enhancement technology for transmission line defect detection not only improves the model's accuracy in detecting defects in small-sized components of transmission lines, but also reduces false detections and missed detections caused by image blurring. The proposed method can perform multi-scale target detection and fuzzy target detection in the field of transmission line defect detection, and has high robustness and high practicality.

[0080] 2. Using the YOLOv5 model as the baseline defect detection model, by adding a small target detection branch, integrating the Transformer encoder and CBAM attention module into the model, and optimizing the loss function, the model's ability to detect defects in small-sized parts is effectively improved.

[0081] 3. To address the issues of blurred edge information and limited semantic information for small targets in blurred images, a blurred image enhancement algorithm based on generative adversarial networks (GANs) is proposed to reconstruct high-resolution, clear images from the images to be detected. Specifically, the generative network of the GAN is replaced with an RRDB network to improve the ability to generate detailed features. Furthermore, a second-order image degradation method is designed based on real-world image degradation phenomena to simulate more realistic blurred images as training data for the GAN, thereby improving the network's convergence performance.

[0082] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A power transmission line defect detection method based on improved YOLOv5 and fuzzy image enhancement, characterized in that, Comprising the following steps: Constructing a defect detection model based on improved YOLOv5, the defect detection model is based on the network structure of YOLOv5, and a shallow prediction branch for detecting small targets is added to the detection head network, and the shallow prediction branch combines the original three detection branches to form a four-branch detection model; Constructing a fuzzy image enhancement algorithm based on a generative adversarial network; Collecting power transmission line image data in multiple scenes and preprocessing, using the preprocessed image data to train the generative adversarial network in the fuzzy image enhancement algorithm and the defect detection model; Using the trained fuzzy image enhancement algorithm and defect detection model to detect the collected power transmission line inspection video frame by frame, each frame of image is first reconstructed by the fuzzy image enhancement algorithm, and then defect detection is performed by the defect detection model.

2. The power line defect detection method based on improved YOLOv5 and fuzzy image enhancement according to claim 1, characterized in that, In the defect detection model based on improved YOLOv5, the Trans module at the end of the backbone network of the original YOLOv5 model and the Trans module after the Concat module in the feature fusion network are replaced by a Transformer encoder.

3. The power transmission line defect detection method based on improved YOLOv5 and fuzzy image enhancement according to claim 1, characterized in that, In the defect detection model based on improved YOLOv5, a CBAM attention module is added before the Conv module in the feature fusion network of the original YOLOv5 model, the CBAM attention module is used to sequentially perform attention mapping inference along the channel dimension and the spatial dimension, and the two attention mapping maps are multiplied with the backbone feature map respectively.

4. The power line defect detection method based on improved YOLOv5 and fuzzy image enhancement according to claim 1, characterized in that, The target confidence loss of the defect detection model is calculated according to the following formula: In the formula, L obj is the target confidence loss, are the losses of different size feature maps mapped by the four detection branches respectively, and β1, β2, β3 and β4 are weight factors of the four detection branches.

5. The power line defect detection method based on improved YOLOv5 and fuzzy image enhancement according to claim 4, characterized in that, β1, β2, β3 and β4 are set to 4.0, 1.0, 0.4 and 0.2 respectively.

6. The power transmission line defect detection method based on improved YOLOv5 and fuzzy image enhancement according to claim 1, characterized in that, In the fuzzy image enhancement algorithm based on the generative adversarial network, a RRDB network and a U-NET discriminant network are used to construct a generative adversarial network.

7. The power transmission line defect detection method based on improved YOLOv5 and fuzzy image enhancement according to claim 1, characterized in that, In the fuzzy image enhancement algorithm based on the generative adversarial network, a second-order image degradation method is used to simulate and generate fuzzy images, and the second-order image degradation method models the image through two repeated degradation processes to generate fuzzy images.

8. The power transmission line defect detection method based on improved YOLOv5 and fuzzy image enhancement according to claim 7, characterized in that, The degradation process specifically includes: Convolve the image with a blur kernel function; Change the size of the image through downsampling operation; Randomly add different types of noise in the image; JPEG compress the image.

9. The power line defect detection method based on improved YOLOv5 and fuzzy image enhancement according to claim 7, characterized in that, In the fuzzy image enhancement algorithm, the loss function includes an adversarial loss function, a content loss function and an edge loss function, and the calculation formulas of each loss function are as follows: Adversarial loss function: In the formula, l adv represents the adversarial loss, and respectively, the discrimination results of the generated image and the real image under the U-NET discrimination network, and N is the number of images. Content loss function: wherein l MSE represents the content loss, r, W, H are the number of images, width and height respectively, and x and y are the coordinate information of the pixels, represents the original image, I LR represents the blurred image simulated by using the second-order image degradation method, G(I LR ) x，y represents the I LR generated image after inputting the RRDB network Edge loss function: In the formula, l Edge represents the edge loss, W and H represent the width and height of the image, and E represents the image edge features calculated by the Canny operator.

10. The power line defect detection method based on improved YOLOv5 and fuzzy image enhancement according to claim 7, characterized in that, The training of the fuzzy image enhancement algorithm uses a stochastic gradient descent optimizer as the training optimization strategy of the network, and the defect detection model uses an adaptive moment estimation optimizer as the training optimization strategy of the network.

Images