Insulator defect detection method based on improved RetinaNet

Abstract

The invention provides an insulator defect detection method based on improved RetinaNet, which is used for solving the problem of low detection precision caused by complex background, various target scales and small detection target of aerial insulator images of a high-voltage transmission line. The method comprises the following specific steps: constructing a data set containing a large number of insulator defect images, and accurately marking and reasonably dividing the data set; a baseline RetinaNet model is improved, an adaptive convolution module is designed to optimize a backbone network, and the small target detection capability is enhanced; a coordinate enhanced BiFPN is adopted to carry out feature fusion, so that the multi-scale target detection performance is improved; a brand new F-CIoU Loss loss function is introduced, so that the defect detection precision in a complex scene is improved; utilizing the training set to train the improved model, and adjusting and optimizing parameters through the verification set; and finally evaluating the model by using the test set. According to the method, the insulator defect detection precision and efficiency can be effectively improved, powerful support is provided for electric power inspection work, and the method has remarkable practical application value.

Description

Technical Field

[0001] The present invention relates to the field of computer vision technology, and in particular to an insulator defect detection method based on an improved RetinaNet. Background Art

[0002] Electricity, as an indispensable form of energy in social production and daily life, supports the normal operation of modern society. In high-voltage transmission lines, insulators are key components that ensure the stable operation of power systems. Their performance is directly related to the safety and reliability of power transmission. However, due to long-term exposure to complex and changing natural environments, such as strong and weak light, and the presence of cluttered trees and grass, insulators are prone to various defects such as cracks and breakage. Once these defects occur, they can not only affect the electrical performance of the insulators but also cause large-scale power outages, resulting in significant economic losses and social impacts.

[0003] Traditional insulator defect detection methods rely primarily on the experience of professional line inspectors to record and judge defects. This approach suffers from low efficiency, high cost, and significant safety risks. With the rapid development and maturity of drone technology, combining it with efficient, high-performance, and intelligent target detection models to automate insulator defect detection has become a key development direction in transmission line inspection.

[0004] In insulator defect detection, single-stage detection models are widely used in tasks requiring high real-time performance due to their simple structure and fast inference speed. The RetinaNet algorithm, a single-stage detection model with excellent detection accuracy and speed, excels in detecting small and medium-sized objects. However, in actual detection, many complex and uncertain factors, such as varying lighting conditions, complex backgrounds, and the diversity of defect types, place even higher demands on the RetinaNet algorithm's detection accuracy.

[0005] Although some research has been devoted to improving target detection models to enhance the accuracy of insulator defect detection, these methods often have the following defects and deficiencies: 1. Optimization for specific scenarios or defect types: For example, the transmission line defect recognition method based on an improved RetinaNet network proposed in CN110705374A improves recognition accuracy by optimizing the feature extraction network using DenseNet (densely connected convolutional network). However, this method primarily targets overall defect recognition for transmission lines and does not fully consider defect detection in insulators, a specific component, and in complex backgrounds. Similarly, other improved methods in the prior art often optimize for specific scenarios or defect types, lacking comprehensive consideration of complex backgrounds, multi-scale targets, and small target detection capabilities.

[0006] 2. Inadequate detection accuracy in complex backgrounds: Aerial images of insulators feature complex and changing backgrounds, such as strong and weak light, cluttered trees and grass. These background factors can interfere with the model's detection accuracy. Existing technologies struggle to effectively distinguish insulator defects from complex backgrounds, leading to missed or false detections.

[0007] 3. Challenges in multi-scale object detection: Insulators and their defects appear at various scales in images, ranging from tiny cracks to large damage. Existing models perform poorly when handling multi-scale objects, making it difficult to simultaneously maintain accurate detection for both large and small objects.

[0008] 4. Limited detection of small objects: Small objects occupy a relatively small portion of an image, and information is easily lost. Existing models have limited detection capabilities for small objects, resulting in a high rate of missed detection of defects such as tiny cracks.

[0009] Therefore, how to further improve the detection accuracy of the RetinaNet algorithm in complex backgrounds, especially the detection capability for small and multi-scale targets, has become a technical problem that needs to be solved urgently in the current field of insulator defect detection. In summary, the present invention aims to provide an insulator defect detection method based on an improved RetinaNet. By designing an adaptive convolution module to optimize the backbone network, using coordinate-enhanced BiFPN for feature fusion, and introducing a new F-CIoU Loss (Focaland Complete IoU Loss) loss function, the detection accuracy of the model in complex backgrounds is improved, especially the detection capability for small and multi-scale targets, providing a more accurate and efficient solution for insulator defect detection in power inspection work. Summary of the Invention

[0010] The technical problem solved by the present invention is to provide an insulator defect detection method based on an improved RetinaNet, which solves the problem of low detection accuracy caused by complex background, diverse target scales and small detection targets during the detection of insulator defects in high-voltage transmission lines. Traditional manual inspection methods are not only inefficient and costly, but also pose serious safety hazards. However, the detection accuracy and robustness of existing automated detection technologies still need to be further improved when faced with complex and changing natural environments and multi-scale, small-target insulator defects. Therefore, the present invention proposes an insulator defect detection method based on an improved RetinaNet to overcome the limitations of the existing technology and improve detection accuracy and efficiency.

[0011] In order to achieve the above objectives, the present invention adopts the following technical solutions: S1. Dataset construction and annotation: First, we used drones to collect a large number of insulator images along high-voltage transmission lines. These images covered scenes at different times and under different weather conditions to ensure the diversity and representativeness of the dataset. Subsequently, the LabelImg annotation tool is used to accurately annotate the insulator strings and their various defects (such as cracks, breakage, and contamination) in the image. This automatically generates an XML (eXtensible Markup Language) annotation file containing information such as the defect category and the coordinates of the annotation box. Finally, the dataset is divided into training set, validation set, and test set in proportion to ensure the independence and effectiveness of model training, tuning, and testing.

[0012] S2. Model improvement and construction: Several improvements are made to the baseline RetinaNet model to build a model more suitable for insulator defect detection: Backbone network optimization: We designed an adaptive convolution module to optimize the ResNet (Residual Network) backbone network. Specifically, we replaced the standard 3×3 convolutions in all residual modules in ResNet with deformable convolutions. These deformable convolutions automatically adjust the shape and position of the convolution kernel based on the content of the input feature map, thereby enhancing the model's ability to detect small and complex objects. Neck network feature fusion: The original FPN (Feature Pyramid Network) network is replaced with a Bidirectional Feature Pyramid Network with CoordinateAttention (BiFPN) layer. BiFPN introduces a coordinate attention mechanism to sample and fuse bidirectional features from feature maps at different levels output by the multi-layer adaptive convolution module, enhancing the model's ability to detect multi-scale objects. Specifically, it uses the coordinate attention mechanism to capture deeper spatial information related to feature maps processed by depthwise separable convolution in BiFPN, improving the accuracy and efficiency of feature fusion. Improved prediction network loss function: In the prediction network, the Focal Loss (FocalLoss) loss function in the classification subnetwork is replaced with the newly designed F-CIoU Loss loss function. F-CIoU Loss combines the advantages of the Complete Intersection over Union (CIoU) Loss function and the Focal Loss loss function. CIoU Loss is introduced to constrain the regression subnetwork, and IoU (Intersection over Union) is used as a weighting coefficient in combination with Focal Loss to form the F-CIoU Loss loss function. This design enables the model to more accurately locate and classify defects in complex scenarios.

[0013] S3, model training and tuning: The improved model is trained using the training set, with parameters such as the number of iterations, batch size, optimizer, initial learning rate, and weight decay set. During training, the model parameters are continuously adjusted using the backpropagation algorithm to minimize the loss function. After training, the model parameters are fine-tuned using the validation set, specifically by adjusting hyperparameters such as the learning rate and batch size, and using techniques such as early stopping to prevent overfitting. Metrics such as recall, accuracy, and precision are measured to ensure that the model meets the expected performance standards.

[0014] S4 model testing and application: Based on the various weight parameters of training and verification, the model is tested using a test set. The test set of insulator defect images to be detected is input into the trained and verified detection model to obtain the defect detection results output by the prediction network, including information such as defect category, location coordinates, and confidence level. These results can be used to guide power inspection personnel to repair or replace insulators in a timely manner, thereby ensuring the stable operation of the power system.

[0015] The present invention provides an insulator defect detection method based on an improved RetinaNet, which has the following beneficial effects: 1. To address the problem of low detection accuracy caused by complex backgrounds, diverse target scales, and small detection targets in aerial insulator images, the present invention provides a more accurate defect detection solution.

[0016] 2. The present invention significantly improves the accuracy of insulator defect detection through model improvement and optimization (including adaptive convolution module, coordinate enhanced BiFPN layer, F-CIoULoss loss function, etc.). Experimental results show that the improved model has achieved significant improvements in various indicators, especially in recall rate and mAP value, which have increased to 91.41% and 91.46% respectively, which are 5.95% and 5.34% higher than those of the unoptimized model.

[0017] 3. The present invention improves the robustness of the model in complex scenes and can maintain high detection accuracy under different lighting conditions and background environments.

[0018] 4. The insulator defect dataset constructed by the present invention uses the LabelImg annotation tool to annotate the defects and then divides them into training set, validation set and test set to reduce the risk of model overfitting.

[0019] 5. The adaptive convolution module constructed by the present invention optimizes the ResNet network, enabling the convolution kernel to automatically adjust its shape and position according to the content of the input feature map, thereby enhancing the detection capability of small targets and targets with complex shapes.

[0020] 6. This paper designs a coordinate-enhanced BiFPN layer, which improves the accuracy and efficiency of feature fusion by introducing a coordinate attention mechanism, and enhances the model's ability to detect multi-scale targets.

[0021] 7. The present invention uses the F-CIoU Loss loss function in the prediction network, combining the advantages of CIoU Loss and Focal Loss to improve the ability of defect detection in complex scenarios.

[0022] 8. The present invention improves the efficiency of power inspection work and reduces the cost and safety hazards of manual inspection through automated detection methods.

[0023] 9. The present invention provides strong support for power inspection work, guiding power inspection personnel to repair or replace insulators in a timely manner to ensure the stable operation of the power system. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 Schematic diagram of the method flow of the present invention; Figure 2 This is a structural diagram of the adaptive convolution module of the present invention; Figure 3 This is the coordinate enhanced BiFPN structure diagram of the present invention; Figure 4 This is the coordinate attention principle diagram of the present invention; Figure 5This is the overall network structure diagram of the present invention. DETAILED DESCRIPTION

[0025] The technical solutions of the present invention are further described below with reference to the accompanying drawings and embodiments: Example 1 like Figures 1 to 5 As shown, this embodiment provides an insulator defect detection method based on the improved RetinaNet, comprising the following steps: 1. Dataset Construction Image acquisition: Using drones equipped with high-definition cameras, systematic aerial photography is conducted along high-voltage transmission lines. The aerial photography plan must cover different time periods (e.g., morning, noon, and evening), different seasons (to account for different lighting and weather conditions), and different geographical environments (e.g., forests, grasslands, and cities) to ensure the diversity and representativeness of the dataset. Image annotation: Use the LabelImg annotation tool to carefully annotate the captured images. The annotation content includes the outline and position of the insulator string, as well as the specific location and category of various defects (such as cracks, damage, dirt, bird nests, etc.). During the annotation process, the accuracy of the annotation box and the completeness of the annotation information must be ensured. An XML format annotation file containing information such as category and annotation box coordinates is automatically generated. Dataset division: The labeled dataset is divided into training set, validation set, and test set in a ratio of 8:1:1. During the division process, the independence and non-overlap between the sets must be ensured to ensure the fairness of model training and evaluation.

[0026] 2. Model improvement and construction Backbone network optimization: Based on the baseline RetinaNet model, the ResNet backbone network was optimized. Specifically, the standard 3×3 convolutions in all residual modules were replaced with deformable convolutions to form adaptive convolution modules. These deformable convolutions can dynamically adjust the shape and position of the convolution kernel based on the content of the input feature map, thereby better capturing the characteristics of insulators and their defects. Neck network feature fusion: The original FPN network is replaced with a coordinate-enhanced BiFPN layer. Coordinate-enhanced BiFPN introduces a coordinate attention mechanism to sample and bidirectionally fuse feature maps at different levels output by multi-layer adaptive convolution modules. This fusion approach not only considers the spatial information of the feature maps but also enhances the model's ability to detect multi-scale objects. Improved prediction network loss function: In the prediction network, the Focal Loss loss function in the classification subnetwork is replaced with the newly designed F-CIoU Loss loss function. F-CIoU Loss combines the advantages of CIoU Loss and Focal Loss. By introducing CIoU Loss to constrain the regression subnetwork and using IoU as a weighting coefficient in combination with Focal Loss, it improves the accuracy of defect detection in complex scenarios.

[0027] 3. Model training and tuning Training parameter settings: Use the training set to train the improved model, set the number of iterations to 300, the batch size to 32, the optimizer to Adam, the initial learning rate to 0.001, and the weight decay to 0.0005; Training process monitoring: During training, the model's performance is evaluated by monitoring changes in the loss function, accuracy, recall, and other metrics. Techniques such as early stopping are used to prevent overfitting. Training is terminated early when performance metrics on the validation set no longer significantly improve. Hyperparameter tuning: Use the validation set to tune the model parameters. Specifically, adjust hyperparameters such as learning rate, batch size, and weight decay to optimize model performance. At the same time, try different network structures and loss function combinations to find the optimal model that best suits the current task.

[0028] 4. Model testing and application Model testing: Input the insulator defect image test set to be detected into the trained and verified detection model to obtain the defect detection results output by the prediction network; the results include information such as defect category, location coordinates, and confidence level; Performance evaluation: Based on the test results, calculate the model's accuracy, recall, precision, and mAP (Mean Average Precision) to evaluate model performance. Practical Application Guidance: Based on the test results, power inspection personnel are guided to promptly repair or replace insulators. Simultaneously, test results are fed back into the model training process to continuously optimize model performance and improve detection accuracy and efficiency.

[0029] Example 2 In another preferred embodiment, based on embodiment 1 and taking into account illumination conditions, this embodiment provides an insulator defect detection method based on an improved RetinaNet, comprising the following steps: 1. Dataset Construction Lighting condition simulation: During aerial photography, special attention is paid to capturing images of insulators under different lighting conditions (such as strong light, weak light, backlight, and shadows). Different lighting conditions can be simulated by adjusting the drone's flight time, shooting angle, and using different filters. Image annotation and enhancement: The collected images are annotated with the same annotation content as in Example 1. At the same time, data enhancement techniques such as random lighting adjustment, contrast enhancement, and noise addition are introduced during the training process to simulate scenes under different lighting conditions and enhance the generalization ability of the model.

[0030] 2. Model improvement and construction Light-insensitive feature extraction: Add a light-insensitive feature extraction layer to the backbone network; for example, use a color constancy algorithm to preprocess the image to reduce the impact of light changes on feature extraction. At the same time, introduce a light-adaptive mechanism in the adaptive convolution module, allowing the convolution kernel to automatically adjust its receptive field and weight distribution according to lighting conditions. Multi-illumination condition fusion: In the neck network, a multi-illumination condition fusion layer is designed to fuse feature maps under different illumination conditions. By introducing an attention mechanism, the model can dynamically focus on important features under different illumination conditions, improving the model's detection capabilities under complex illumination conditions.

[0031] 3. Model training and tuning Training data partitioning: Divide the dataset containing different lighting conditions into training, validation, and test sets. Ensure that the lighting conditions are evenly distributed across the sets to evaluate the model's performance under different lighting conditions. Lighting Condition Adaptive Training: During the training process, we pay special attention to the model's adaptability under different lighting conditions. By adjusting hyperparameters and optimizing strategies, we enable the model to maintain high detection accuracy and stability under various lighting conditions. Performance evaluation and tuning: Use the validation set to evaluate the performance of the model. Specifically, the adaptability of the model is evaluated by calculating indicators such as accuracy and recall under different lighting conditions. Based on the evaluation results, the model is further tuned and optimized.

[0032] 4. Model testing and application Multi-lighting condition testing: The test set is divided into subsets under different lighting conditions, and each subset is input into the trained and verified detection model for testing. During the test, the detection results and performance indicators under different lighting conditions must be recorded; Practical application guidance: Based on the test results, the model's performance under different lighting conditions is evaluated and used to guide inspection work in actual applications. At the same time, the test results are fed back into the model training process to continuously optimize the model performance and improve its detection accuracy and stability under complex lighting conditions.

[0033] Example 3 In another preferred embodiment, based on embodiments 1 and 2, this embodiment provides an insulator defect detection method based on improved RetinaNet, such as Figure 1 As shown, the following steps are included: S1. Construct an insulator defect dataset, use the LabelImg annotation tool to annotate the defects, and then divide the dataset into training, validation, and test sets. S2. Improve the baseline RetinaNet by designing an adaptive convolution module to optimize the backbone network, use coordinate-enhanced BiFPN for feature fusion, and introduce a new F-CIoU Loss function to build the final defect detection model. The insulator defect detection model based on the improved RetinaNet includes a backbone network, a neck network, and a prediction network. The output end of the backbone network is connected to the input end of the neck network, and the output end of the neck module is connected to the input end of the prediction network. The backbone network includes an adaptive convolution module to optimize the ResNet backbone network; The neck network includes a coordinate-enhanced BiFPN layer; The prediction network includes a classification subnetwork and a regression subnetwork.

[0034] The adaptive convolution module optimizes the ResNet backbone network to extract features from insulator defect images. The specific steps are as follows: 1) Preliminary feature extraction through convolution and maximum pooling; 2) Use multi-layer adaptive convolution module to extract multi-layer features of defect features; 3) Output feature maps at different levels.

[0035] Among them, the adaptive convolution module replaces the ordinary 3×3 convolution in all residual modules with deformable convolution to form an adaptive convolution module, such as Figure 2 shown. Note: Figure 2 In the figure, Conv, convolution; DCN, deformable convolution module; DFConv, deformable convolution; In the neck network, the original FPN network is replaced with the coordinate-enhanced BiFPN, and the feature maps output by different layers of the multi-layer adaptive convolution module are sampled and fused. The specific steps are as follows: 1) Input the feature maps of different levels output by the backbone network; 2) Using the coordinate-enhanced BiFPN structure to sample the feature map and perform bidirectional feature fusion; 3) Output enhanced multi-scale feature maps.

[0036] Among them, the coordinate enhanced BiFPN structure uses the coordinate attention mechanism to capture deeper spatial position-related information on the feature map after depth-separable convolution processing in BiFPN, such as Figure 3 As shown, the specific feature fusion principle of BiFPN is as follows: (1) (2) Where, 、 、 Indicates the The intermediate, input, and output features of the layer; Indicates the The input features of the layer; Indicates the Output features of the layer; is an upsampling or downsampling operation; 、 are the learning weights 1 and 2 of the input features respectively; represents the bias term; 、 、 are the learning weights 1, 2, and 3 of the output features respectively; Note: Figure 3 In the figure, P3, P4, P5, P6, and P7 represent the feature maps of the 3rd, 4th, 5th, 6th, and 7th layers, respectively; BN, batch normalization operation; DSC, depthwise convolution; CA, coordinate attention; Output, output.

[0037] The coordinate attention mechanism captures deeper information about the features processed by the depthwise separable convolution in BiFPN as follows: 1) The feature map after depth-wise separable convolution in BiFPN is fed into a 1×1 convolution kernel to reduce the dimension and merge the feature maps in two different directions. The principle is as follows: (3) Where, is the merging operation of channel features; is a nonlinear activation function; 、 Represent the spatial features in the horizontal and vertical directions respectively, is the spatial feature generated by coordinate attention; 2) Segment the feature map into two independent feature tensors and process them through 1×1 convolution and sigmoid function so that the resulting horizontal and vertical attention vectors are consistent with the dimensions of the input feature map. The principle is as follows: (4) (5) 3) Output and The weighted result is the output of the coordinate attention block. The principle is as follows: (6) Where, is the attention vector in the vertical direction; is the attention vector in the horizontal direction; is the Sigmoid function; 、 They are feature maps in the vertical and horizontal directions respectively; 、 are the feature tensors in the vertical and horizontal directions respectively; Output feature map for coordinate attention; Input feature map for coordinate attention; is the attention weight in the vertical direction; is the attention weight in the horizontal direction; The coordinate attention mechanism structure is as follows Figure 4 shown.

[0038] The F-CIoU Loss loss function introduces the CIoULoss loss function to constrain the regression sub-network, and uses IoU as a weighted coefficient combined with Focal Loss to form the F-CIoU Loss loss function. The specific principles are as follows: (7) Where, is the IoU weight coefficient, is the CIoU loss function, Used to control the weights of classification loss and regression loss, the improved RetinaNet insulator defect detection model is as follows Figure 5 shown; note: Figure 5 In the figure, C1, C2, C3, C4, and C5 represent the feature extraction layers of the 1st, 2nd, 3rd, 4th, and 5th layers; MaxPool represents the maximum pooling.

[0039] S3. After training the improved model using the training set, use the validation set to tune the model parameters. The specific steps are as follows: During the training phase, the number of iterations in the experiment was set to 300, the batch size was set to 32, the optimizer was Adam, the initial learning rate was set to 0.001, the weight decay was set to 0.0005, and the improved model was trained using the training set. After the training, the model parameters were adjusted using the validation set, and the recall rate, accuracy rate, and precision rate were measured. If the standards were met, the model was determined to be a detection model.

[0040] S4. Based on the weight parameters of training and validation, use the test set to test the model. The specific steps are as follows: The test set of insulator defect images to be detected is input into the trained and verified detection model to obtain the defect detection results output by the prediction network.

[0041] To evaluate the model optimization capabilities of the proposed solution, ablation experiments were designed, and the data is shown in Table 1. The experimental results show that the baseline RetinaNet model achieves a mAP value of 81.12%. Furthermore, the introduction of the adaptive convolution module, coordinate-enhanced BiFPN, and Focal-CIoU alone increased mAP by 1.22%, 2.34%, and 1.42%, respectively. Combining the two optimization methods further improved mAP compared to introducing the improved methods individually. Ultimately, after combining all optimization methods, the model's recall and mAP increased to 91.41% and 91.46%, respectively, representing improvements of 5.95% and 5.34%, respectively, compared to the unoptimized state. This demonstrates that the proposed optimization method significantly enhances feature extraction, fusion, and localization capabilities, thereby significantly improving the model's overall detection performance.

[0042] Table 1 Ablation experiment data

[0043] To further verify the effectiveness of the proposed method, other mainstream models were used for training on the same dataset. These models include SSD, Faster RCNN, YOLOv5, YOLOv7, YOLO-X, EfficientDet, and CenterNet, which have been widely used in insulator defect detection tasks. The results are shown in Table 2.

[0044] Table 2 Comparison of the algorithm of the present invention with other algorithms

[0045] This invention addresses the problem of low detection accuracy in aerial insulator images, which is caused by complex backgrounds, diverse object scales, and small detection targets. Using a self-developed dataset, the model implemented in this invention achieved significant improvements across all metrics, validating the effectiveness and practicality of this method. This work provides an effective solution for insulator defect detection in power inspections, with promising applications and practical significance.

[0046] In the preferred solution, the detailed steps of step S1 are: using a drone to collect insulator images, using the LabelImg annotation tool to annotate the insulator strings and various defects in the data set, automatically generating an annotation file containing relevant information such as categories and annotation boxes, and dividing the data set into training set, validation set and test set at an 8:1:1 ratio; the above settings ensure that there is sufficient data support during the model training process. Then, in step S2, a deep learning framework is used to build an insulator defect detection model, the preprocessed image data is input, features are extracted through a convolutional neural network, and the defect type is identified through a classifier, laying a solid foundation for subsequent accurate detection.

[0047] In the preferred scheme, in step S2, the insulator defect detection model based on the improved RetinaNet includes a backbone network, a neck network and a prediction network. The output end of the backbone network is connected to the input end of the neck network, and the output end of the neck module is connected to the input end of the prediction network. The backbone network includes an adaptive convolution module to optimize the ResNet backbone network, the neck network includes a coordinate enhanced BiFPN layer, and the prediction network includes a classification subnetwork and a regression subnetwork. The above settings effectively improve the accuracy and efficiency of the insulator defect detection model, the adaptive convolution module enhances the feature extraction capability, the coordinate enhanced BiFPN layer optimizes feature fusion, and the classification subnetwork and regression subnetwork respectively improve the accuracy of defect classification and positioning.

[0048] In the preferred solution, the adaptive convolution module replaces the standard 3×3 convolutions in all residual modules with deformable convolutions to form an adaptive convolution module. This configuration aims to improve the model's adaptability to objects of different shapes and scales and enhance feature extraction capabilities. At the same time, by introducing an attention mechanism, feature weight distribution is further optimized to ensure that key information is effectively transmitted and processed.

[0049] In the preferred solution, the coordinate-enhanced BiFPN layer uses the coordinate attention mechanism to capture deeper spatial position-related information on the feature map after depth-separable convolution processing in BiFPN; the above settings enable the model to more accurately locate the target object in the target detection task, significantly improving the detection accuracy. At the same time, the solution optimizes the operating efficiency of the model by reducing computational redundancy, achieving a dual improvement in speed and accuracy.

[0050] In the preferred solution, the Focal Loss loss function in the prediction network is replaced by the F-CIoU Loss loss function. The F-CIoULoss loss function introduces the CIoULoss loss function to constrain the regression subnetwork, and uses IoU as a weighted coefficient to combine with Focal Loss to form the F-CIoULoss loss function. The above settings effectively improve the accuracy of the model's target bounding box regression, while enhancing the model's attention to difficult-to-classify samples, significantly improving the overall detection performance, especially in complex backgrounds and dense target scenes.

[0051] In the preferred solution, in step S3, the training model parameters include the number of iterations, batch size, optimizer, initial learning rate, and weight decay. The improved model is trained using the training set. After the training, the model parameters are adjusted using the validation set, and the recall rate, accuracy rate, and precision rate are measured. If the standards are met, the model is determined to be a detection model. The above settings ensure the stability and reliability of the model in identifying the target object. In addition, an early stopping strategy is introduced to avoid overfitting. When the performance of the validation set no longer improves, the training is automatically terminated, further improving the generalization ability of the model.

[0052] In the preferred solution, in step S4, the test set of insulator defect images to be detected is input into the detection model after training and verification to obtain the defect detection results output by the prediction network; the above settings can effectively identify defects such as cracks, stains, and missing parts on the surface of the insulator, thereby improving the accuracy and efficiency of detection. At the same time, the solution also has a high generalization ability and can adapt to the needs of insulator defect detection under different environmental conditions.

[0053] In summary, this paper proposes an insulator defect detection method based on an improved RetinaNet. This method systematically addresses the low detection accuracy issues associated with complex backgrounds, diverse target scales, and small detection targets in high-voltage transmission line insulator defect detection. Compared to the inefficiency, high costs, and safety hazards of traditional manual inspections, as well as the limitations of existing automated inspection technologies in complex environments, multi-scale detection, and small targets, this method successfully overcomes these deficiencies and significantly improves detection accuracy and efficiency.

[0054] This invention achieves significant technical improvements, including designing an adaptive convolutional module to optimize the backbone network, using coordinate-augmented BiFPN for feature fusion, and introducing a novel F-CIoU Loss function. These innovative measures, tailored to the specific needs of insulator defect detection and uncommon in existing technologies, effectively improve detection performance.

[0055] The present invention also constructs a dataset containing a large number of insulator defect images, accurately annotating and rationally segmenting them, providing strong support for model training and validation. This dataset construction approach may not be common in existing technologies, especially for the specific scenario of aerial photography of insulator images of high-voltage transmission lines.

[0056] This invention integrates these improvements into insulator defect detection, forming a comprehensive and efficient detection method. Its effectiveness and practicality have been verified experimentally. This invention addresses challenges such as complex backgrounds, diverse target sizes, and small detection targets by providing effective solutions, significantly improving detection accuracy and robustness, and representing a significant breakthrough in existing technologies.

[0057] In addition, the innovative technologies introduced in this invention, such as the adaptive convolution module, coordinate enhanced BiFPN and the new F-CIoU Loss loss function, are of groundbreaking significance in the field of insulator defect detection. They not only improve the detection performance, but also provide new ideas and directions for subsequent related technical research.

[0058] This invention has been experimentally validated for its effectiveness and practicality in insulator defect detection, providing strong support for power inspections. This close integration of technological innovation and practical application fully demonstrates the invention's creativity and practical value, giving it significant advantages and enormous potential in the field of insulator defect detection.

Claims

1. An insulator defect detection method based on improved RetinaNet, characterized in that: The following steps are involved: S1: Construct an insulator defect dataset, use the LabelImg annotation tool to annotate the defects, and then divide the dataset into training, validation, and test sets. S2: Improve the baseline RetinaNet by designing an adaptive convolution module to optimize the backbone network, use coordinate-enhanced BiFPN for feature fusion, and introduce a new F-CIoU Loss function to build the final defect detection model. S3: After training the improved model using the training set, the validation set is used to tune the model parameters; S4: Based on the weight parameters of training and validation, use the test set to test the model.

2. The insulator defect detection method based on the improved RetinaNet according to claim 1, characterized in that: The detailed steps of step S1 are: using a drone to collect insulator images, using the LabelImg annotation tool to annotate the insulator strings and various defects in the dataset, automatically generating an annotation file containing relevant information such as categories and annotation boxes, and dividing the dataset into a training set, a validation set, and a test set.

3. The insulator defect detection method based on the improved RetinaNet according to claim 2, characterized in that: In step S2, the insulator defect detection model based on the improved RetinaNet includes a backbone network, a neck network and a prediction network. The output end of the backbone network is connected to the input end of the neck network, and the output end of the neck module is connected to the input end of the prediction network. The backbone network includes an adaptive convolution module to optimize the ResNet backbone network, the neck network includes a coordinate enhanced BiFPN layer, and the prediction network includes a classification subnetwork and a regression subnetwork.

4. The insulator defect detection method based on the improved RetinaNet according to claim 3, characterized in that: The adaptive convolution module in the backbone network optimizes the ResNet backbone network to extract features from the insulator defect image. The specific steps are as follows: S2.1.1: Perform preliminary feature extraction through convolution and max pooling; S2.1.2: Use multi-layer adaptive convolution module to perform multi-layer feature extraction on defect features; S2.1.3: Output feature maps at different levels.

5. The insulator defect detection method based on improved RetinaNet according to claim 4, characterized in that: The adaptive convolution module replaces the ordinary convolutions in all residual modules with deformable convolutions to form an adaptive convolution module.

6. The insulator defect detection method based on improved RetinaNet according to claim 5, characterized in that: In the neck network, the original FPN network is replaced with the coordinate-enhanced BiFPN, and the feature maps output by different layers of the multi-layer adaptive convolution module are sampled and fused. The specific steps are as follows: S2.2.1: Input the feature maps of different levels output by the backbone network; S2.2.2: Use the coordinate-enhanced BiFPN structure to sample the feature map and perform bidirectional feature fusion; S2.2.3: Output enhanced multi-scale feature maps.

7. The insulator defect detection method based on improved RetinaNet according to claim 6, characterized in that: The coordinate enhanced BiFPN layer uses a coordinate attention mechanism to capture deeper spatial position-related information on the feature map after depthwise separable convolution processing in BiFPN.

8. The insulator defect detection method based on improved RetinaNet according to claim 7, characterized in that: The Focal Loss loss function in the prediction network is replaced by the F-CIoU Loss loss function. The F-CIoULoss loss function introduces the CIoULoss loss function to constrain the regression subnetwork, and uses IoU as a weighted coefficient combined with Focal Loss to form the F-CIoULoss loss function.

9. The insulator defect detection method based on improved RetinaNet according to claim 8, characterized in that: In step S3, the training model parameters include the number of iterations, batch size, optimizer, initial learning rate, and weight decay. The improved model is trained using the training set. After the training, the model parameters are adjusted using the validation set, and the recall rate, accuracy rate, and precision rate are measured. If the standards are met, the model is determined to be a detection model.

10. The insulator defect detection method based on improved RetinaNet according to claim 9, characterized in that: In step S4, a test set of insulator defect images to be detected is input into a trained and verified detection model to obtain a defect detection result output by a prediction network.

Images