Insulator detection method for power transmission line based on adaptive double-weighted feature network

Abstract

This invention proposes a method for detecting transmission line insulators based on an adaptive dual-weighted feature network. The steps are as follows: Targets are divided into three levels based on their size; insulator images are selected according to the level, and the number of images is expanded using data augmentation strategies to obtain a dataset, which is then proportionally divided into training, testing, and validation sets; an insulator detection model is constructed: a context-aware feature extraction module and a cross-scale residual perception network are added to the neck network, and feature fusion is performed using a dual-weighted feature fusion network; the constructed insulator detection model is trained using the training set to obtain the weight parameters of the insulator detection model; the detection results of the trained insulator detection model are tested and validated using the testing and validation sets, respectively. This invention can effectively detect small insulator targets, while effectively preserving the channel and spatial features of the images, effectively solving the problem of poor detection performance for small insulator targets.

Description

Technical Field

[0001] This invention relates to the field of target detection based on deep learning, specifically to the technical field of insulator detection for transmission lines. In particular, it relates to a method for detecting insulators for transmission lines based on an adaptive dual-weighted feature network, which achieves high-precision detection of insulators, especially for the detection of small insulators in transmission lines. Background Technology

[0002] In power applications, high-voltage insulators require regular inspection to ensure the safe and stable operation of the power system. Accurate insulator location is crucial for effective subsequent defect detection. Insulators are essential components in power transmission and distribution systems, preventing current from flowing from conductors to supporting structures. They maintain the integrity and stability of the power grid by providing electrical isolation and mechanical support. Insulator inspection plays a vital role in ensuring the safe and reliable operation of the power system. With the growth of electricity demand and the development of high-voltage transmission lines, the normal operation of insulators has become increasingly important. However, due to environmental conditions such as pollution, humidity, and temperature variations, insulators deteriorate over time. Insulator defects can severely affect the reliable operation of transmission lines and the stability of the power system. Therefore, insulator inspection, as a preliminary step in fault detection and line inspection, is a critical task.

[0003] Convolutional Neural Networks (CNNs) have achieved great success in computer vision, with wide applications in areas such as image classification, object detection, and image generation. Through the design of shared weights and local receptive fields, CNNs effectively capture local features and spatial relationships in images, extracting more discriminative features and thus performing exceptionally well in image processing and analysis tasks. Beyond computer vision, CNNs also play a crucial role in natural language processing, medical image processing, speech recognition, video analysis, and AI-assisted processing. Their powerful feature extraction and generalization capabilities make them an important tool for processing and analyzing grid-structured data. In recent years, researchers have continuously improved and optimized the structure of CNNs to enhance their performance. For example, new architectures such as ResNet, Inception, and EfficientNet have achieved significant results in various image recognition competitions, achieving higher accuracy and efficiency. Through continuous improvement of network structure, optimization of training strategies, and the introduction of new technologies, CNNs have made breakthroughs in object detection, object recognition, image generation, and transfer learning, providing more powerful tools and methods for solving computer vision tasks. How to effectively protect shallow feature information while preserving deep feature information is an important problem that needs to be solved in object detection.

[0004] Patent application number 202310117643.3 discloses an aerial insulator multi-defect detection method integrating CAT-BiFPN and attention mechanisms. The method involves preprocessing aerial insulator defect images, training and evaluating a YOLOv7 network model, improving the YOLOv7 network based on the evaluation, and training the improved YOLOv7 network. To address the shortcomings of the YOLOv7 network model, such as insufficient feature fusion due to the lack of weight adjustment for the multi-scale feature fusion module, difficulty in accurately detecting targets with significantly different scales within the same image, and the absence of a detection layer for small objects, the method replaces the bidirectional path fusion network (PANet) in YOLOv7 with a weighted bidirectional feature pyramid (CAT-BiFPN) with an intra-unit skip structure. This reduces redundancy, improves the fusion degree of multi-scale target features, and forms a fourth detection layer for small target detection. Furthermore, by adding an attention and convolutional hybrid module (ACmix), the method focuses more on the details in the features, enabling the detection model to capture key information more efficiently, improving the model's detection capability, and achieving high detection accuracy and speed. The aforementioned patent application integrates CAT-BiFPN and attention mechanism models for insulator defect detection, but the number of parameters is relatively large, and the detection accuracy still needs to be improved. Summary of the Invention

[0005] To address the technical problem of low accuracy in detecting small insulators using existing insulator detection methods, this invention proposes a transmission line insulator detection method based on Adaptive Dual Weighted Net (ADFNet). This method can quickly detect targets and has particularly high accuracy in detecting insulators (mAP of 98.19%, 3.92% higher than the source code YOLOv7). It also solves the problem that existing target detection methods, such as YOLOv7, cannot extract more detailed feature information from images, resulting in low accuracy in detecting small targets.

[0006] To achieve the above objectives, the technical solution of the present invention is as follows: a method for detecting transmission line insulators based on an adaptive double-weighted feature network, comprising the following steps:

[0007] Step 1: Based on the size of the target, divide the target into three levels;

[0008] Step 2: Select insulator images according to the three levels, and then expand the number of images through data augmentation strategies to obtain a dataset, which is then divided into training set, test set and validation set according to the proportion;

[0009] Step 3: Constructing an insulator detection model: Add a context-aware feature extraction module and a cross-scale residual sensing network to the neck network, and use a dual-weighted feature fusion network for feature fusion;

[0010] Step 4: Train the constructed insulator detection model using the training set to obtain the weight parameters of the insulator detection model;

[0011] Step 5: Test and validate the detection results of the trained insulator detection model using the test set and validation set respectively.

[0012] Preferably, the target in step one is divided into three levels based on the pixel size occupied by the target in the image, as follows:

[0013] Size range 0×0～32×32 32×32～96×96 ≥96×96 .

[0014] Preferably, the data augmentation strategies in step two include random flipping, image cropping, adjusting brightness, adjusting contrast and saturation, and adding Gaussian noise; the dataset is divided into training set, test set and validation set according to the proportions of 81%, 10% and 9%.

[0015] Preferably, the insulator detection model includes a backbone network, a neck network, and a prediction network. A context-aware feature extraction module and a cross-scale residual perception network are added to the neck network, and the ordinary fusion module in the neck is replaced with a dual-weighted feature fusion network. The detection head of the prediction network contains a series of convolutional layers for integrating spatial and channel information of high-level features.

[0016] Preferably, the effective feature map obtained by feature extraction through the backbone network is input into the context-aware feature extraction module and then sequentially passed to the average pooling layer and the channel attention module. The context-aware feature extraction module includes a context feature enhancement module and an average pooling layer. The context feature enhancement module takes the feature map output by the backbone network as input and contains three 3*3 convolutional layers with different dilation rates and one 1*1 convolutional layer. Each convolutional layer has 32 output channels. The average pooling layer moves the pooling window on the input feature map, sums the values ​​of all pixels within the pooling window, and divides the sum by the number of pixels to obtain the average value as the pooling result. The channel attention module refers to introducing an attention mechanism in the channel dimension, and weighting the features of different channels through the weight matrix generated by the channel attention.

[0017] Preferably, the cross-scale residual perception network is designed behind the context-aware feature extraction module, with a skip connection structure added at both ends of the convolution. The feature map before convolution is passed to the feature map after convolution through the skip connection and concatenated with the feature map before convolution, so that the two feature maps of different depths are fused, thereby passing shallow information to deep layers.

[0018] The dual-weighted feature fusion module fully represents the features of the target through effective feature extraction and feature weighting. The implementation method is as follows: the original feature matrix of the effective feature map is multiplied with the attention matrix generated by the channel attention module to obtain the weighted feature map.

[0019] Preferably, the weighting method is as follows:

[0020] M(E) = CA(Avgpool(CFE(E)))

[0021] E′=M(E)×E;

[0022] E″ = E′cat E;

[0023] Where M(E) is the semantic weight matrix generated by the average pooling layer and the channel attention module, E is the effective feature map, E′ is the selected feature map, and E″ is the robust feature map; CFE represents the context feature enhancement module, CA is the channel attention module, and cat is the feature concatenation method.

[0024] Preferably, the input image is processed by a backbone network to extract features, resulting in three feature maps of different depths. These three depth feature maps are a large-scale feature matrix, a medium-scale feature matrix, and a small-scale feature matrix, respectively. The convolution operation in the backbone network is performed by cross-correlation operations from left to right and from top to bottom using convolution kernels.

[0025] The detection head divides the high-level feature map into grid cells of different scales. Each grid cell is responsible for predicting one or more bounding boxes. Each bounding box includes coordinate information, target confidence, and class probability. Finally, it outputs three target detection result parameters: the target object's localization box, detection confidence, and class from the fused features at large, medium, and small scales, respectively.

[0026] Preferably, the backbone network extracts features from the input image to obtain three effective feature maps of different depths; the context-aware feature extraction module enhances the features of the three effective feature maps, and the output feature layer is input into the cross-scale residual perception network. The cross-scale residual perception network adds a skip connection structure at both ends of the convolution to complete the transfer of information from shallow to deep layers. The output features are input into the dual-weighted feature fusion module, which multiplies the effective feature maps of different depths with a specific attention matrix to obtain weighted feature maps. These are then input into the three detection head networks of the prediction network—the large-scale detection layer, the medium-scale detection layer, and the small-scale detection layer—for detection, respectively detecting insulators of different sizes and obtaining the detection results.

[0027] Preferably, the insulator detection model uses the Adam optimizer during training, with an initial learning rate of 0.001 and a training batch size of 8.

[0028] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0029] Building upon YOLOv7, this invention adds a Context-aware Feature Extraction with Channel Attention (CFECA) module, a cross-scale residual perception network, and a Dual Weighted Feature Fusion (DWFF) module, which significantly improves the low accuracy of small insulator detection. This invention significantly enhances the target representation ability of convolutional neural networks through an adaptive dual-weighted feature network, extracting more effective feature information, especially for small targets, thereby improving the detection accuracy of the ADFNet. This invention effectively detects small targets, is relatively easy to train, and effectively preserves the channel and spatial features of the image. Experiments show that on the insulator dataset, the ADFNet of this invention achieves excellent detection performance for both large and small targets, effectively solving the problem of poor small target detection. Attached Figure Description

[0030] 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.

[0031] Figure 1 This is a schematic diagram of the process of the present invention.

[0032] Figure 2 This is a structural diagram of the ADFNet of the present invention.

[0033] Figure 3 The structural diagram of the context-aware feature extraction module provided by the present invention.

[0034] Figure 4 This is a structural diagram of the dual-weighted feature fusion module provided by the present invention.

[0035] Figure 5 The results show the mAP comparison between the present invention and the original YOLOv7 target detection algorithm on the insulator dataset, where (a) represents the present invention and (b) represents the original YOLOv7 target detection algorithm.

[0036] Figure 6 This is a comparison chart of the detection performance of the present invention and YOLOv7 on a small target insulator dataset.

[0037] Figure 7 This is a comparison chart of the detection performance of the present invention and the YOLOv8 detection model on a small target insulator dataset. Detailed Implementation

[0038] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0039] The core ideas of this invention are: (1) A context-aware feature extraction module is designed to improve the performance of small insulator detection and facilitate practical application. (2) A cross-scale residual sensing network is introduced, which can better integrate the key features of shallow and deep layers, making the reconstructed image features more accurate. (3) A dual-weighted feature fusion network is proposed to replace the ordinary feature fusion module in the neck, which can better integrate the feature maps of three different levels, improve the insulator detection accuracy, and thus improve the detection performance and robustness of the model.

[0040] The hardware environment used to implement this invention is as follows: the ADFNet detection model is implemented using CUDA 10.0 under the PyTorch 1.2 deep learning framework. All improved models were trained and tested on an Intel NVIDIA TITAN RTX with 256GB of memory and an 8163 processor.

[0041] like Figure 1As shown, a transmission line insulator detection method based on an adaptive dual-weighted feature network is presented. In deep learning target detection, firstly, a context-aware feature extraction with channel attention (CFECA) module is designed and applied to all effective feature layers obtained from the YOLOv7 backbone network. Semantic weights are calculated through deeper feature mapping, and then the obtained weights are used to guide features. Secondly, a cross-scale residual perception network is combined with the effective feature layers to collect more information from different levels, which not only helps improve the convergence speed but also solves the problem of accuracy degradation caused by gradient descent as the number of network layers increases. Finally, a dual-weighted feature fusion (DWFF) network is proposed to replace the common feature fusion mode in the neck area, fully representing the target's features. This invention improves the accuracy and performance of the model through effective feature extraction and feature weighting. The specific implementation steps of this invention are as follows:

[0042] Step 1: Based on the size of the target, divide the target into three levels.

[0043] As shown in Table 1, the targets are first classified in detail, and the size of the targets is used as the basis for classification, and the target size is divided into three levels.

[0044] Table 1 Target Size Classification

[0045] Size range 0×0～32×32 32×32～96×96 ≥96×96

[0046] Dividing the image into three levels based on the pixel size occupied by the target can enhance the matching degree between the detection model and the input image containing the target, thereby enabling the detection model to achieve better detection results for the actual input image.

[0047] Step 2: Create a training set: Select insulator images according to the grade, expand the number of images through data augmentation strategies to obtain a dataset, and divide it into training set, test set and validation set according to the proportion.

[0048] Data augmentation strategies, such as expanding the number of images, can prevent overfitting and improve the model's robustness. The data augmentations used include random flipping, image cropping, adjusting brightness, contrast and saturation, and adding Gaussian noise. Originally containing 1000 images, the dataset was expanded to 4000 images to increase data diversity, alleviate overfitting, and increase training samples. The training, test, and validation sets consisted of 3240, 400, and 360 images, respectively. Due to the large number of insulators and the complex and diverse backgrounds in the dataset, the images were divided into training, test, and validation sets at a ratio of 81%, 10%, and 9%, respectively. It is important to ensure that the training set is large enough to allow the model to learn patterns from the data, while the validation and test sets are relatively large to provide a reliable assessment of the model's generalization performance. This dataset partitioning not only helps ensure that the model performs well on the training data but also enables it to generalize to new data.

[0049] Step 3: Construct the ADFNet insulator detection model: Add a context-aware feature extraction module and a cross-scale residual sensing network to the neck network, and use a dual-weighted feature fusion network for feature fusion.

[0050] like Figure 2 As shown, the insulator detection model includes a backbone network, a neck network, and a prediction network. After processing, the image data is input into the ADFNet-based insulator detection model for insulator detection. The ADFNet insulator detection model adds a designed context-aware feature extraction module and introduces a cross-scale residual sensing network to the neck network, and replaces the ordinary fusion module in the neck network with a proposed dual-weighted feature fusion network.

[0051] The input image is processed by a backbone network to extract features, resulting in three feature maps at different depths. The backbone network generates three layers of depth features at different scales: large-scale, medium-scale, and small-scale depth feature matrices. The convolutional operations in the backbone network perform cross-correlation operations from left to right and top to bottom using convolution kernels, preserving positional information. The values ​​of the red, green, and blue channels of the image are then summed to obtain the feature information at different depths.

[0052] The Context-aware Feature Extraction with ChannelAttention (CFECA) module plays a crucial role in focusing on the features surrounding the target. For example... Figure 3As shown, the three effective feature maps output after feature extraction by the backbone network are input into the context-aware feature extraction module, and then passed to the average pooling layer and the channel attention module to obtain selected feature maps. The context-aware feature extraction module plays a crucial role with the context feature enhancement module and the average pooling layer, used to better capture different levels of contextual information from shallow features. The context feature enhancement module takes the feature map output from the backbone network as input and contains three 3x3 convolutional layers with different dilation rates and one 1x1 convolutional layer, with each convolutional layer having 32 output channels. The average pooling layer moves a pooling window across the input feature map; the window size is W*H, where W and H represent the width and height, respectively. Within the pooling window, the values ​​of all pixels are summed and then divided by the number of pixels to obtain the average value as the pooling result. This process of moving the pooling window and repeating the above operation continues until the entire feature map has been processed. The channel attention module mainly introduces an attention mechanism along the channel dimension. By weighting the features of different channels, it enhances the model's understanding of the correlation and importance between different channels, thereby improving model performance. Figure 3 The Weights in the algorithm are weight matrices generated through channel attention, which are then multiplied by the feature matrix of the third branch to obtain the weighted feature map, i.e., the Selected feature.

[0053] To significantly enhance the feature extraction capability of the detection module, the context feature enhancement module, utilizing the semantic matrix generated by the second branch, can clearly distinguish the importance of features, thus effectively guiding the selection of key features and enhancing the saliency of the target. Furthermore, fusing shallow and deep feature information helps enhance feature representation and improve the model's detection accuracy. The weighting method is as follows:

[0054] M(E) = CA(Avgpool(CFE(E)))

[0055] E′=M(E)×E;

[0056] E″ = E′cat E;

[0057] Where M(E) is the semantic weight matrix generated by the average pooling layer and the channel attention module, E is the effective feature map, E′ is the selected feature map, and E″ is the robust feature map. CFE represents the context feature enhancement module, CA refers to the channel attention module, and cat indicates the feature concatenation method.

[0058] A cross-scale residual sensing network was designed in the neck, specifically in Figure 2Within the dashed box. The cross-scale residual sensing network can collect more information from different levels, which not only helps improve convergence speed but also solves the problem of accuracy degradation due to gradient descent as the number of network layers increases. The cross-scale residual sensing network of this invention is designed behind the context-aware feature extraction module. Specifically, this network adds a skip connection structure at both ends of the convolution, passing the feature map before convolution to the feature map after convolution and concatenating them, thus fusing the two feature maps of different depths and completing the transfer of information from shallow to deep layers.

[0059] like Figure 4 As shown, the dual-weighted feature fusion module replaces the common feature fusion mode in the neck area. Through effective feature extraction and weighting, it fully represents the target's features, selects more important feature details, and more accurately calculates channel and spatial attention levels. It also provides directional information of shallow and deep features in a novel way. Specifically, the original feature matrix is ​​multiplied by the attention matrix to obtain a weighted feature map. Weighting is performed using spatial and channel attention, emphasizing key parts and suppressing relatively unimportant information. This helps the model focus more intently on task-beneficial information in deep features and also improves the model's efficiency in utilizing inter-channel correlation information.

[0060] The detection head is like Figure 2 The Repconv module in the structure diagram contains a series of convolutional layers used to integrate spatial and channel information from high-level features. It divides the high-level feature map into grid cells of different scales, with each grid cell responsible for predicting one or more bounding boxes. Each bounding box typically includes coordinate information, object confidence, and class probability. Finally, it outputs three object detection result parameters: the object's localization box, detection confidence, and class from the fused features at large, medium, and small scales, respectively.

[0061] Combining the overall structure Figure 2For example, given an input image, the backbone network extracts features from the image, resulting in three effective feature layers of different depths. Then, a context-aware feature extraction module enhances these three effective feature layers, playing a crucial role in focusing on features surrounding the target. Next, the output feature layers are fed into a cross-scale residual perception network, designed after the context-aware feature extraction module. This network adds a skip connection at both ends of the convolution to facilitate the transfer of information from shallow to deep layers. This not only improves convergence speed but also addresses the accuracy degradation caused by gradient descent as the number of network layers increases. Finally, the output features are input into a proposed dual-weighted feature fusion network. The Dual-Weighted Feature Fusion (DWFF) module replaces the ordinary fusion module in the neck region. This module selects more important feature details by multiplying the original feature matrices at different depths with specific attention matrices to obtain weighted feature maps. This allows for more accurate calculation of channel and spatial attention. It also fuses feature information in a novel weighted manner. Finally, the data is fed into the large-scale, medium-scale, and small-scale detection layers of the prediction network for detection, respectively, to detect insulators of different sizes, thus obtaining the final detection results.

[0062] Step 4: Train the ADFNet detection model using the training set to obtain the weight parameters of the ADFNet insulator detection model.

[0063] First, three effective feature maps are obtained through the backbone network. Then, three feature maps with enhanced features are obtained through the context-aware module and the dual-weighted feature fusion module in the neck area. Finally, these are input into three detection head networks to obtain target detection results of different sizes, thus obtaining the weight parameters of the ADFNet insulator detection model. The weight parameters are generated during model training. During training, the ADFNet insulator detection model uses the Adam optimizer, with an initial learning rate set to 0.001 and a training batch size set to 8. These parameter settings are to ensure optimal performance while maintaining stable model training.

[0064] Step 5: After training the ADFNet insulator detection model using the training set, test and validation sets are used to test and validate the insulators.

[0065] Based on the above method and specific implementation steps, the effectiveness of the invention was verified through experiments. The experimental parameters and dataset used in the experiments of this invention are as shown in the specific steps above. Recall, precision, F1 score, and mean average precision (mAP) were used to evaluate the model's performance. The IOU threshold during model training was set to 0.5. The F1 score is the harmonic mean of recall and precision, which helps to more accurately measure the model's performance. mAP represents the average precision value of multiple targets under different recall conditions.

[0066] In this embodiment, to verify the effectiveness of the improved insulator target detection model proposed in this invention, the network input image resolution was 640*640*3, and the training was performed for 120 epochs (one "epoch" represents the process of the entire training dataset being completely forward-propagated and backward-propagated once during model training). The detection accuracy of YOLOv8, YOLOv7, DETR (DEtection Transformer, a novel detection model in recent years), and ADFNet networks were compared respectively. The experimental results are shown in Table 2.

[0067] Table 2 Simulation Results

[0068]

[0069]

[0070] As shown in Table 2, except for accuracy, the recall, F1 score, and mAP of this invention are all higher than other algorithms on the insulator dataset. Although the ratio of precisely predicted positive samples to negative samples is not as good as YOLOv8, it achieves a very high mAP value. Comparing the detection results of ADFNet with those of YOLOv7, DETR, and YOLOv8, the mAP is improved by 4.04%, 7.16%, and 0.47%, respectively. In the ADFNet insulator detection model of this invention, although the P-value of ADFNet is 2.06% lower than the highest DETR model but 12.35% higher than the lowest DETR model, the R-value is 6.55% higher than the lowest YOLOv7 model, and the F1 score is 9% higher than the lowest DETR model. Overall, it can be seen that the model of this invention performs well in target detection, especially in terms of recall, F1 score, and mAP. This experiment demonstrates the advantages of the ADFNet network, showing good robustness and excellent performance in target detection, especially for small targets.

[0071] Figure 5The graph shows the mAP comparison between the present invention and the original YOLOv7 target detection algorithm on the insulator dataset. This accuracy graph was generated during the evaluation of the two different models. It clearly demonstrates that the present invention has significantly higher detection accuracy compared to the original model, showcasing its superior performance. Figure 6 and Figure 7 The visualized detection maps were obtained by inputting the same original image into both the ADFNet model and the contrastive model for detection. Figure 6 As can be seen from the comparison chart in the first row, this model not only maintains high accuracy in detecting large targets but also improves performance in detecting small targets. The left image in this row only detected two large targets, while the right image, in addition to detecting these, also detected a small target, an insulator, at the bottom of the image. The comparison chart in the second row shows that, compared to the original model, the model of this invention improves the accuracy of detecting each insulator and also detects the rightmost insulator in the image. Figure 7 The comparison chart in the first row shows that the accuracy of the model in this invention in detecting each insulator is improved. YOLOv8's detection accuracy for insulators is 79%, 93%, 80%, 65%, and 89%, while the accuracy of this invention is 91%, 95%, 91%, 86%, and 95%. Furthermore, it detected a small insulator on the left side of the chart that YOLOv8 failed to detect with 67% accuracy. The comparison chart in the second row shows that the accuracy of the model in this invention in detecting each insulator is improved. In addition, it detected a very small insulator with 83% accuracy. In conclusion, these visualization results demonstrate that this invention has excellent detection performance in insulator detection.

[0072] This invention proposes an adaptive dual-weighted feature network (ADFNet) detection method for transmission line insulators, which effectively improves the low accuracy of small insulator detection. The ADFNet insulator detection model of this invention significantly enhances the target representation ability of convolutional neural networks through the adaptive dual-weighted feature network, extracting more effective feature information, especially for small targets, thereby improving the detection accuracy of ADFNet. This invention can effectively detect small targets, is relatively easy to train, and effectively preserves the channel and spatial features of the image. Experiments show that on insulator datasets, this invention achieves excellent detection performance for both large and small targets, effectively solving the problem of poor small target detection.

[0073] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for detecting transmission line insulators based on an adaptive dual-weighted feature network, characterized in that, The steps are as follows: Step 1: Based on the size of the target, divide the target into three levels; Step 2: Select insulator images according to the three levels, and then expand the number of images through data augmentation strategies to obtain a dataset, which is then divided into training set, test set and validation set according to the proportion; Step 3: Constructing an insulator detection model: Add a context-aware feature extraction module and a cross-scale residual sensing network to the neck network, and use a dual-weighted feature fusion network for feature fusion; Step 4: Train the constructed insulator detection model using the training set to obtain the weight parameters of the insulator detection model; Step 5: Test and validate the detection results of the trained insulator detection model using the test set and validation set respectively; The insulator detection model includes a backbone network, a neck network, and a prediction network. A context-aware feature extraction module and a cross-scale residual perception network are added to the neck network, and the ordinary fusion module in the neck is replaced with a dual-weighted feature fusion network. The detection head of the prediction network contains a series of convolutional layers for integrating spatial and channel information of high-level features. The effective feature maps obtained by feature extraction through the backbone network are input into the context-aware feature extraction module and then passed to the average pooling layer and the channel attention module in sequence. The context-aware feature extraction module includes a context feature enhancement module and an average pooling layer. The context feature enhancement module takes the feature map output by the backbone network as input and contains three 3*3 convolutional layers with different dilation rates and one 1*1 convolutional layer. Each convolutional layer has 32 output channels. The average pooling layer moves the pooling window on the input feature map, sums the values ​​of all pixels within the pooling window, and divides the sum by the number of pixels to obtain the average value as the pooling result. The channel attention module introduces an attention mechanism in the channel dimension, and uses the weight matrix generated by the channel attention to weight the features of different channels. The cross-scale residual perception network is designed behind the context-aware feature extraction module. A skip connection structure is added at both ends of the convolution to pass the feature map before convolution to the feature map after convolution and concatenate it with the feature map before convolution. This allows the two feature maps of different depths to be fused, thereby passing shallow information to deep layers. The dual-weighted feature fusion network fully represents the features of the target through effective feature extraction and feature weighting. The implementation method is as follows: the original feature matrix of the effective feature map is multiplied with the attention matrix generated by the channel attention module to obtain the weighted feature map.

2. The method for detecting transmission line insulators based on an adaptive dual-weighted feature network according to claim 1, characterized in that, The targets in step one are divided into three levels based on the pixel size occupied by the targets in the image: small targets have a size range of 0×0~32×32; medium targets have a size range of 32×32~96×96; and large targets have a size range of ≥96×96.

3. The method for detecting transmission line insulators based on an adaptive dual-weighted feature network according to claim 2, characterized in that, The data augmentation strategies in step two include random flipping, image cropping, adjusting brightness, adjusting contrast and saturation, and adding Gaussian noise; the dataset is divided into training set, test set and validation set according to the proportions of 81%, 10% and 9%.

4. The method for detecting transmission line insulators based on an adaptive dual-weighted feature network according to any one of claims 1-3, characterized in that, The weighting method is as follows: ； ； ； Where M(E) is the semantic weight matrix generated by the average pooling layer and the channel attention module, E is the effective feature map, E' is the selected feature map, and E'' is the robust feature map; CFE represents the context feature enhancement module, CA is the channel attention module, and cat is the feature concatenation method.

5. The method for detecting transmission line insulators based on an adaptive dual-weighted feature network according to claim 4, characterized in that, The input image is processed by the backbone network to extract features, resulting in three feature maps of different depths: a large-scale feature matrix, a medium-scale feature matrix, and a small-scale feature matrix. The convolution operation in the backbone network is performed by cross-correlation operations from left to right and from top to bottom using convolution kernels. The detection head divides the high-level feature map into grid cells of different scales. Each grid cell is responsible for predicting one or more bounding boxes. Each bounding box includes coordinate information, target confidence, and class probability. Finally, it outputs three target detection result parameters: the target object's localization box, detection confidence, and class from the fused features at large, medium, and small scales, respectively.

6. The method for detecting transmission line insulators based on an adaptive dual-weighted feature network according to claim 5, characterized in that, The backbone network extracts features from the input image, obtaining three effective feature maps at different depths. The context-aware feature extraction module enhances the features of the three effective feature maps, and the output feature layers are input into a cross-scale residual perception network. The cross-scale residual perception network adds a skip connection structure at both ends of the convolution to complete the transfer of information from shallow to deep layers. The output features are input into a dual-weighted feature fusion network, where effective feature maps at different depths are multiplied by specific attention matrices to obtain weighted feature maps. These are then input into three detection head networks of the prediction network—large-scale detection layer, medium-scale detection layer, and small-scale detection layer—for detection, detecting insulators of different sizes and obtaining the detection results.

7. The method for detecting transmission line insulators based on an adaptive dual-weighted feature network according to claim 1 or 6, characterized in that, During training, the insulator detection model uses the Adam optimizer with an initial learning rate of 0.001 and a training batch size of 8.

Images