Aerial photography insulator multi-defect detection method, system and equipment integrating CAT-BiFPN and attention mechanism

Abstract

A method, system, and device for detecting multiple defects in aerial insulators that integrate a CAT-BiFPN (Cat-BiFPN) and an attention mechanism are disclosed. The method preprocesses aerial insulator defect images, trains and evaluates a YOLOv7 network model, improves the YOLOv7 network based on the evaluation, and trains the improved YOLOv7 network. The system addresses issues in the YOLOv7 network model, such as the lack of adjustment of the weights of the multi-scale feature fusion module, insufficient feature fusion, difficulty in accurately detecting objects with greatly different scales in the same image, the lack of a detection layer for small objects, and poor detection of small-scale insulator defects. A weighted bidirectional feature pyramid (CAT-BiFPN) with an intra-unit skip structure replaces the bidirectional path fusion network (PANet) in YOLOv7, reducing redundancy, improving the fusion of multi-scale target features, and forming a fourth detection layer for small target detection. The system also adds an attention and convolution hybrid module (ACmix) to focus more on details in the features, enabling the detection model to more efficiently capture key information, improving the model's detection capability, and achieving high detection accuracy and speed.

Description

Technical Field

[0001] The present invention relates to the technical field of aerial photography insulator defect detection, and in particular to an aerial photography insulator multi-defect detection method, system, and device that integrates CAT-BiFPN and an attention mechanism. Background Art

[0002] The safe and stable operation of transmission lines is crucial for reliable power supply in power systems. Insulators support conductors and prevent current backflow in transmission and distribution circuits. Insulators are exposed to the elements for long periods of time, and are susceptible to damage, spontaneous explosions, and contamination from various natural hazards, such as temperature and humidity, lightning, strong electric fields, and pollution. Insulator defects and bird nests on high-voltage towers seriously endanger the safe and stable operation of transmission lines. Therefore, timely detection of insulator defects is crucial. Given the scale of today's power grids, manual field surveys or traditional methods are no longer able to adapt to the demands of grid development and transformation. Furthermore, detection accuracy is poor due to factors such as the natural environment and visual angle.

[0003] Using target detection algorithms in computer vision to realize automatic detection of drone inspection images has become a research hotspot. However, the target scales of insulators in aerial photography vary greatly, and insulator defects have the characteristics of small target scales and diverse angles, which makes it extremely difficult for the algorithm to achieve better detection accuracy.

[0004] Patent application CN113298789A discloses an insulator defect detection method, system, electronic device, and readable storage medium. The insulator defect detection method comprises the following steps: inputting an aerial image of the power grid being inspected into a pretrained insulator defect detection model, and outputting an insulator defect detection result for the aerial image of the power grid being inspected using the pretrained insulator defect detection model; the insulator defect detection model comprises: an insulator detection and foreground segmentation module for acquiring a regional image containing only insulators based on the input aerial image of the power grid; and an insulator defect detection module for determining whether the insulator is defective based on the acquired regional image containing only insulators using an image classification network based on multi-head self-attention, and outputting an insulator defect detection result. This patent application utilizes a multi-head attention encoding and decoding model for feature capture and employs foreground segmentation plus defect detection, resulting in insufficient detection speed. Summary of the Invention

[0005] In order to overcome the defects of the above-mentioned prior art, the purpose of the present invention is to provide a method, system and equipment for multi-defect detection of aerial insulators that integrates CAT-BiFPN and attention mechanism. On the basis of the standard YOLOv7 algorithm, the original backbone network composed of four efficient aggregation modules (ELAN) is retained, and the self-attention and convolution mixing module (ACmix) is connected after the output of the first and second feature layers to pay more attention to the details in the key areas and features; the four feature layers are input into a weighted bidirectional feature pyramid (CAT-BiFPN) with an intra-unit skip structure, and the different feature layers obtained by the up- and down-sampling combination module are spliced ​​to achieve feature fusion and obtain a new feature map. Finally, the four groups of outputs obtained by CAT-BiFPN form a corresponding detection head for small target detection, with high detection accuracy and fast detection speed.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is:

[0007] A method for detecting multiple defects in insulators using aerial photography that integrates CAT-BiFPN and an attention mechanism includes the following steps:

[0008] Step 1: Preprocess the aerial insulator defect images, including image annotation, dataset image amplification, and dataset division.

[0009] 1) Image annotation: Use image annotation tools to annotate the insulator defect detection dataset, frame the defective parts in each image with rectangular frames, and indicate the category to which they belong;

[0010] 2) Dataset image augmentation: random noise, color jitter, and flipping are performed on the original images to expand the dataset;

[0011] 3) Dataset division: The dataset expanded in step 2) is divided into a training set, a validation set, and a test set. The original images are compressed to the default size of the YOLOv7 detection network.

[0012] Step 2: Train and evaluate the YOLOv7 network model:

[0013] The image processed in step 3) of step 1 is used as the input image of the YOLOv7 network model, and the operating parameters of the YOLOv7 network model are set. The YOLOv7 network model is trained on the experimental operation platform, and then the evaluation indicators commonly used in the field of target detection are used to evaluate the effect of the trained YOLOv7 network model;

[0014] Step 3: Improve the YOLOv7 network based on the evaluation in step 2;

[0015] Step 4: Train and improve the YOLOv7 network:

[0016] The image processed in step 3) of step 1 is used as the input image of the improved YOLOv7 network in step 3, and the operating parameters of the YOLOv7 network model are set. The improved YOLOv7 network model in step 3 is trained on the experimental operation platform. Then, the evaluation indicators commonly used in the field of target detection are used to evaluate the effect of the trained YOLOv7 network model, and finally the detection results of insulator defects are output.

[0017] The specific method of step three is as follows:

[0018] 1) Replace the feature fusion module and add a fourth detection layer:

[0019] 1.1) The four Efficient LAN modules (ELANs) in the backbone network of the YOLOv7 model are retained. The input image first passes through the backbone network, that is, undergoes four rounds of pooling and ELAN operations of the standard YOLOv7 network model. Feature extraction is performed on the middle, mid-high, and high-level feature layers of the backbone network, and the input is fed into a BiFPN with an intra-unit skip structure (CAT-BiFPN). The extended Efficient LAN module (E-ELAN) is combined with upsampling and downsampling to fuse features from different layers, remove redundant nodes, and use a method of adding weights to the input to distinguish the importance of different features during the feature fusion process, thereby improving the fusion of multi-scale target features.

[0020] 1.2) Add a prediction layer for small-scale object detection to the YOLOv7 model. The first efficient aggregation module in the backbone network generates a low-level, high-resolution feature map as the fourth input layer to the CAT-BiFPN:

[0021] The input image first passes through the backbone network, that is, it undergoes four pooling and efficient aggregation module (ELAN) operations; after the first pooling and efficient aggregation module operation, it undergoes a 1×1 convolution and enters the ACmix module to form the fourth input layer of the CAT-BiFPN module; after the second pooling and efficient aggregation module operation, it undergoes a 1×1 convolution and enters the ACmix module to form the third input layer of the CAT-BiFPN module; after the third pooling and efficient aggregation module operation, it undergoes a 1×1 convolution and forms the second input layer of the CAT-BiFPN module; after the fourth pooling and efficient aggregation module operation, it undergoes a 1×1 convolution and forms the first input layer of the CAT-BiFPN module; after the fourth input layer passes through the CAT-BiFPN feature fusion module, it forms the fourth detection layer of the algorithm;

[0022] 2) Design a self-attention and convolution hybrid module:

[0023] The self-attention and convolution hybrid module (ACmix) is embedded in the YOLOv7 network model using CAT-BiFPN. That is, the large-scale feature maps output by the low and middle layers of the backbone network are learned, and more features are captured through the convolution channel and self-attention channel within the ACmix mechanism. Its output becomes the input of the third and fourth input layers of the CAT-BiFPN feature fusion module.

[0024] The operating parameters of the network model in step 2 and step 4 include the initial learning rate, learning momentum, and weight decay rate, and the stochastic gradient descent method SGD is used to update the network parameters.

[0025] The commonly used evaluation indicators in the field of target detection in Step 2 and Step 4 are Average Precision (AP) and Mean Average Precision (mAP). Average Precision is determined by Recall and Precision, and is an intuitive standard for the performance results of single-category detection.

[0026] The recall calculation formula (1) is:

[0027]

[0028] The precision calculation formula (2) is:

[0029]

[0030] Where TP is the number of correctly detected target objects, FN is the number of undetected target objects, and FP is the number of falsely detected target objects.

[0031] With the recall rate as the horizontal axis and the maximum precision rate corresponding to each recall rate as the vertical axis, the Precision-Recall curve is drawn. The area under the curve is the AP value after the curve is integrated. After obtaining multiple single-category values, the average value is calculated to obtain the mAP value. The mAP value is the average precision rate of all categories. The detection effect of the model on all categories is measured by the mAP value. Its calculation formula (3) is:

[0032]

[0033] Where N is the number of categories of targets to be detected in the dataset.

[0034] A multi-defect detection system for aerial insulators that integrates CAT-BiFPN and attention mechanism includes:

[0035] Data preprocessing module, used for aerial insulator defect image annotation, dataset image amplification, and dataset partitioning;

[0036] The backbone network module includes four pooling and four efficient aggregation modules, which are used to extract features from the input image.

[0037] The self-attention and convolution hybrid module is used to obtain more feature information from the feature map, distinguish different defects, distinguish between background and defects, and reduce the false detection rate and missed detection rate. Ultimately, the features of large-scale and small-scale objects enter the feature fusion network in a relatively balanced state, thereby improving the model's sensitivity to smaller insulator defects.

[0038] The feature fusion network module includes a BiFPN (CAT-BiFPN) with an intra-unit skip structure, which is used to effectively fuse input feature maps of different scales and improve the fusion of multi-scale target features;

[0039] The detection layer module is used to detect features and restore them to the original anchor box.

[0040] An aerial insulator multi-defect detection device that integrates CAT-BiFPN and attention mechanism, including:

[0041] Image collector, used to collect aerial images of insulator defects;

[0042] A program processor, a Jetson nano platform, is used to store a computer program and implement the aerial insulator multi-defect detection method integrating CAT-BiFPN and an attention mechanism described in steps 1 to 4 when executing the computer program;

[0043] The display is used to display the detection result of the insulator defect in the image.

[0044] Compared with the prior art, the present invention has the following beneficial effects:

[0045] 1. Aiming at the shortcomings of the existing model, the present invention first replaces the feature fusion module with the CAT-BiFPN module, deletes the redundant nodes of the original feature fusion module, and uses the method of adding weights to the input to distinguish the importance of different features in the feature fusion process, thereby improving the fusion degree of multi-scale targets.

[0046] 2. To address the problem of low detection accuracy of small targets of insulator defects, the present invention adds a fourth detection layer for small target detection. Although the four prediction layers increase certain model parameters and calculation amount, the structure of the four prediction layers can fully alleviate the negative impact of large target scale differences on model detection accuracy, and greatly improve the detection performance of small target objects.

[0047] 3. The present invention uses a self-attention and convolution hybrid module (ACmix) to help the feature extraction network focus on the target area, that is, using a hybrid attention mechanism to combine multi-head attention (self-attention) and traditional convolution. This flexible attention module can focus on different areas more adaptively and capture more features; learn the distribution rules from the features, recalibrate them, and focus on the position, so that the detection model can capture key information more efficiently and improve the detection ability of the model.

[0048] In summary, this paper builds on the standard YOLOv7 model by modifying the model's feature fusion module to improve the fusion of multi-scale target features. It then utilizes a self-attention and convolutional hybrid module (ACmix) to focus on detail within the features, better distinguishing between defects and background, more accurately differentiating between defects and reducing false positives and missed detections. Furthermore, the addition of a fourth detection layer further improves the accuracy of small target detection. This demonstrates that our model can accurately detect insulator defects and provides support for this field. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 This is an aerial insulator multi-defect detection model that integrates CAT-BiFPN and attention mechanism.

[0050] Figure 2 This is the LabelImg annotation diagram of the present invention.

[0051] Figure 3 This is the CAT-BiFPN feature fusion network structure diagram of the present invention.

[0052] Figure 4 This is the ACmix module structure diagram of the present invention.

[0053] Figure 5 This is the ACmix module embedding design diagram of the present invention. DETAILED DESCRIPTION

[0054] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments:

[0055] See also Figure 1 , a method for detecting multiple defects in aerial insulators by integrating CAT-BiFPN and attention mechanism, specifically including the following steps:

[0056] Step 1: Preprocess the aerial insulator defect images: including image annotation, dataset image amplification, and dataset division

[0057] 1) Image annotation: Use image annotation tools to annotate the insulator defect detection dataset, frame the defective parts in each image with rectangular frames, and indicate the category to which they belong, see Figure 2 ;

[0058] 2) Dataset image augmentation: random noise, color jitter, and flipping are performed on the original images to expand the dataset;

[0059] 3) Dataset division: The dataset expanded in step 2) is divided into a training set, a validation set, and a test set. The original images are compressed to the default size of the YOLOv7 detection network.

[0060] Step 2: Train and evaluate the YOLOv7 network model

[0061] The image processed in step 3) of step 1 is used as the input image of the YOLOv7 network model, and the operating parameters of the YOLOv7 network model are set. The YOLOv7 network model is trained on the experimental operation platform, and then the evaluation indicators commonly used in the field of target detection are used to evaluate the effect of the trained YOLOv7 network model;

[0062] Step 3: Improve the YOLOv7 network based on the evaluation in step 2

[0063] 1) Replace the feature fusion module and add a fourth detection layer

[0064] PANet, as a feature fusion network of the YOLOv7 network model, has nodes that have only one-side input but no feature fusion. Redundant nodes have little effect on feature fusion, but increase additional parameters and computational complexity.

[0065] The weighted bidirectional feature pyramid CAT-BiFPN with an intra-unit skip structure is used to replace the bidirectional path fusion network PANet in YOLOv7.

[0066] Since the detection targets such as self-explosion, contamination, and damage in insulators are small in scale, a prediction layer for small-scale target detection is added to the YOLOv7 model. The first efficient aggregation module in the backbone network generates a low-level, high-resolution feature map as the fourth feature layer input to CAT-BiFPN.

[0067] See also Figure 2 , the four input layers of CAT-BiFPN are and Each node in the CAT-BiFPN network adds weights to the input features and trains these weights using a fast normalization method. Taking the second layer as an example, the first layer input and the second layer input are fused with weights to obtain Feature fusion is shown in formula (4):

[0068]

[0069] Conv is a convolution operation, and Resize performs an upsampling or downsampling operation on the input. When a node has an input smaller than the feature map of this layer, it is upsampled. When it has an input larger than the feature map of this layer, it is downsampled. Formula (4) represents an upsampling operation, ω1 is the learnable weight of the first layer output, and ω2 is the learnable weight of the second layer output.

[0070] Take a connection node as a small unit and use the jump connection structure within the unit to connect and Use concat to connect, and then use 1×1 convolution to reorganize the existing features to form new features, and add nonlinear excitation to the learning of the previous layer to obtain the output of this node This intra-unit skip connection structure not only reconstructs features, but also increases the number of channels, maintains the balance of the original number of channels in the YOLOv7 model, and further improves the feature fusion capability of the feature pyramid. The connection of the intra-unit skip connection structure is shown in formula (5):

[0071]

[0072] Among them, concat is a concatenation operation.

[0073] is the final output of the third layer of the CAT-BiFPN network. Similarly, when three nodes are input, weighted fusion is performed to obtain And get the final output of the second layer As shown in formulas (6) and (7):

[0074]

[0075]

[0076] Among them, Resize is the downsampling operation, ω3, ω4 and ω5 are The learnable weights corresponding to the input.

[0077] The feature map sizes of the four output layers of CAT-BiFPN are and The four output layers undergo reparameterized convolution and guided label assignment from coarse-grained to fine-grained to detect insulator defects.

[0078] Compared with the original three prediction layers, although the four prediction layers increase certain model parameters and computational complexity, the structure of the four prediction layers can fully alleviate the negative impact of large target scale differences on model detection accuracy and significantly improve the detection performance of small target objects.

[0079] 2) Design a self-attention and convolution hybrid module

[0080] Self-attention and convolution hybrid module (ACmix), model structure see Figure 4 It effectively combines the advantages of traditional convolution and self-attention modules. The former uses an aggregation function on the local receptive field of the feature map according to the convolution weights. These weights are shared in the entire feature map, which brings an indispensable inductive bias to image processing; the latter adopts a weighted average operation based on the input feature context, and dynamically calculates the attention weights through the similarity function between related pixel pairs. This flexibility enables the attention module to adaptively focus on different areas, obtain a larger receptive field and contextual information, capture more features, and distinguish between background and detection targets.

[0081] The large-scale feature maps output by the lower and middle layers of the backbone network contain rich information about small objects. The addition of the attention mechanism will help the feature extraction network pay more attention to the characteristics of small objects. Figure 5 That is, the large-scale feature maps output by the low and middle layers of the backbone network are learned, and more features are captured through the convolution channel and self-attention channel inside the ACmix mechanism. The output becomes the input of the third and fourth input layers of the CAT-BiFPN feature fusion module.

[0082] The ACmix mechanism enables the network to more flexibly obtain more feature information through feature maps, more accurately distinguish different defects and distinguish between background and defects, and reduce false detection and missed detection rates. Ultimately, the features of large-scale and small-scale targets enter the feature fusion network in a relatively balanced state, thereby improving the model's sensitivity to smaller insulator defects.

[0083] Step 4: Train and improve the YOLOv7 network

[0084] The processed image from step 1, step 3) was used as the input for the improved YOLOv7 network in step 3. The network model's operating parameters were set and the improved YOLOv7 network model from step 3 was trained on an experimental platform. The trained YOLOv7 network model was then evaluated using commonly used evaluation metrics in the field of object detection, and the final insulator defect detection results were output. The improved YOLOv7 algorithm achieved 3.3%, 16.8%, and 9.6% higher precision, recall, and mean average precision (mAP) than the standard YOLOv7 algorithm, respectively, demonstrating the effectiveness of this method.

[0085] The operating parameters of the network model in step 2 and step 4 include the initial learning rate, learning momentum, and weight decay rate, and the stochastic gradient descent method SGD is used to update the network parameters.

[0086] The operating platform for steps 2 and 4 is Ubuntu 2.4, using the Pytorch 1.1 deep learning framework. The CPU model is AMD EPYC 7542 32-Core Processor, the graphics card (GPU) model is NVIDIA GeForce RTX A50, with 24GB of graphics card memory, and the programming language is Python.

[0087] The commonly used evaluation indicators in the field of target detection in Step 2 and Step 4 are Average Precision (AP) and Mean Average Precision (mAP). Average Precision is determined by Recall and Precision, and is an intuitive standard for the performance results of single-category detection.

[0088] The recall calculation formula (1) is:

[0089]

[0090] The precision calculation formula (2) is:

[0091]

[0092] Where TP is the number of correctly detected target objects, FN is the number of undetected target objects, and FP is the number of falsely detected target objects.

[0093] With the recall rate as the horizontal axis and the maximum precision rate corresponding to each recall rate as the vertical axis, the Precision-Recall curve is drawn. The area under the curve is the AP value after the curve is integrated. After obtaining multiple single-category values, the average value is calculated to obtain the mAP value. The mAP value is the average precision rate of all categories. The detection effect of the model on all categories is measured by the mAP value. Its calculation formula (3) is:

[0094]

[0095] Where N is the number of categories of targets to be detected in the dataset. In this experiment, N(5), namely normal insulator, insulator self-explosion, insulator contamination, insulator breakage, and bird's nest foreign matter.

[0096] The number of images detected within 1 second is measured in frames per second (FPS), which reflects the detection speed of the model. The higher the model's FPS, the more it can meet the real-time detection requirements in engineering applications.

[0097] A multi-defect detection system for aerial insulators that integrates CAT-BiFPN and attention mechanism includes:

[0098] Data preprocessing module, used for aerial insulator defect image annotation, dataset image amplification, and dataset partitioning;

[0099] The backbone network module includes four pooling and four efficient aggregation modules, which are used to extract features from the input image.

[0100] The self-attention and convolution hybrid module is used to more flexibly obtain more feature information from the feature map, more accurately distinguish different defects and distinguish between background and defects, and reduce the false detection rate and missed detection rate. Ultimately, the features of large-scale and small-scale objects enter the feature fusion network in a relatively balanced state, thereby improving the model's sensitivity to smaller insulator defects.

[0101] The feature fusion network module includes a BiFPN (CAT-BiFPN) with an intra-unit skip structure, which is used to effectively fuse input feature maps of different scales and improve the fusion of multi-scale target features;

[0102] The detection layer module is used to detect features and restore them to the original anchor box.

[0103] An aerial insulator multi-defect detection device that integrates CAT-BiFPN and attention mechanism, including:

[0104] Image collector, used to collect aerial images of insulator defects;

[0105] A program processor, a Jetson nano platform, is used to store a computer program and implement the aerial insulator multi-defect detection method integrating CAT-BiFPN and an attention mechanism described in steps 1 to 4 when executing the computer program;

[0106] The display is used to display the detection result of the insulator defect in the image.

[0107] The proposed aerial insulator multi-defect detection method integrating CAT-BiFPN and attention mechanism is compared and analyzed with standard YOLOv7, TPH-YOLOv5l, YOLOR-p6, SSD-VGG and Centernet algorithms, and verified on an insulator multi-defect dataset. The results show that the proposed method has significant improvements in precision, recall and mAP compared with other methods, proving that the model of the present invention can effectively detect various defects on insulators and bird nests on poles and towers.

[0108] The experimental results are shown in the following table.

[0109]

Claims

1. A method for detecting multiple defects in insulators using aerial photography that integrates CAT-BiFPN and an attention mechanism, characterized by: The following steps are involved: Step 1: Preprocess the aerial insulator defect images, including image annotation, dataset image amplification, and dataset division. 1) Image annotation: Use image annotation tools to annotate the insulator defect detection dataset. Use rectangular frames to outline the defective parts in each image and indicate the category to which they belong. 2) Dataset image augmentation: random noise, color jitter, and flipping are performed on the original images to expand the dataset; 3) Dataset division: Divide the dataset expanded in step 2) into training, validation, and test sets, and compress the original images to the default size of the YOLOv7 detection network. Step 2: Train and evaluate the YOLOv7 network model: Use the image processed in step 3 of step 1 as the input image of the YOLOv7 network model, set the operating parameters of the YOLOv7 network model, train the YOLOv7 network model on the experimental operation platform, and then use the evaluation indicators commonly used in the field of object detection to evaluate the effectiveness of the trained YOLOv7 network model; Step 3: Improve the YOLOv7 network based on the evaluation in step 2; The specific method of step three is as follows: 1) Replace the feature fusion module and add a fourth detection layer: 1.1) The four E-Lan modules in the backbone network of the YOLOv7 model are retained. The input image first passes through the backbone network, undergoing four rounds of pooling and E-Lan operations using the standard YOLOv7 network model. Feature extraction is performed on the middle, mid-high, and high-level feature layers of the backbone network, and the resulting input is fed into a BiFPN (CAT-BiFPN) with an intra-unit skipping structure. The extended E-Lan module is combined with upsampling and downsampling to fuse features from different layers, removing redundant nodes. Furthermore, by adding weights to the input, the importance of different features in the feature fusion process is distinguished, improving the fusion of multi-scale object features. 1.2) Add a prediction layer for small-scale target detection to the YOLOv7 model. The first efficient aggregation module in the backbone network generates a low-level, high-resolution feature map as the fourth input layer to CAT-BiFPN: The input image first passes through the backbone network, that is, it undergoes four operations of the pooling and efficient aggregation module ELAN; after the first pooling and efficient aggregation module operation, it undergoes a 1×1 convolution and enters the ACmix module to form the fourth input layer of the CAT-BiFPN module; after the second pooling and efficient aggregation module operation, it undergoes a 1×1 convolution and enters the ACmix module to form the third input layer of the CAT-BiFPN module; after the third pooling and efficient aggregation module operation, it undergoes a 1×1 convolution to form the second input layer of the CAT-BiFPN module; after the fourth pooling and efficient aggregation module operation, it undergoes a 1×1 convolution to form the first input layer of the CAT-BiFPN module; after the fourth input layer passes through the CAT-BiFPN feature fusion module, it forms the fourth detection layer of the algorithm; 2) Design a self-attention and convolution hybrid module: The self-attention and convolution hybrid module ACmix is ​​embedded in the YOLOv7 network model using CAT-BiFPN. This module learns the large-scale feature maps output by the low- and middle-layers of the backbone network, and captures more features through the convolutional and self-attention channels within the ACmix mechanism. The output of ACmix becomes the input of the third and fourth input layers of the CAT-BiFPN feature fusion module. Step 4: Train and improve the YOLOv7 network: The processed image from step 3 of step 1 is used as the input image for the improved YOLOv7 network in step 3. The operating parameters of the YOLOv7 network model are set. The improved YOLOv7 network model from step 3 is trained on the experimental operation platform. The trained YOLOv7 network model is then evaluated using evaluation indicators commonly used in the field of target detection. Finally, the detection results of insulator defects are output.

2. The method for detecting multiple defects in insulators using aerial photography by integrating CAT-BiFPN and attention mechanism according to claim 1 is characterized by: The operating parameters of the network model in step 2 and step 4 include the initial learning rate, learning momentum, and weight decay rate, and the stochastic gradient descent method SGD is used to update the network parameters.

3. The method for detecting multiple defects in insulators using aerial photography that integrates CAT-BiFPN and attention mechanism according to claim 1 is characterized by: The commonly used evaluation indicators in the field of target detection in steps 2 and 4 are the average precision AP and the mean average precision mAP. The average precision is determined by the recall rate Recall and the precision rate Precision, and is an intuitive standard for the performance results of single-category detection. The recall calculation formula (1) is: (1) The precision calculation formula (2) is: (2) Where TP is the number of correctly detected target objects, FN is the number of undetected target objects, and FP is the number of falsely detected target objects. With the recall rate as the horizontal axis and the maximum precision rate corresponding to each recall rate as the vertical axis, the Precision-Recall curve is drawn. The area under the curve is the AP value after the curve is integrated. After obtaining multiple single-category values, the average value is calculated to obtain the mAP value. The mAP value is the average precision rate of all categories. The detection effect of the model on all categories is measured by the mAP value. Its calculation formula (3) is: (3) Where, N is the number of categories of targets to be detected in the dataset.

4. A detection system based on the detection method according to any one of claims 1 to 3, comprising: Data preprocessing module, used for aerial insulator defect image annotation, dataset image amplification, and dataset partitioning; The backbone network module includes four pooling and four efficient aggregation modules, which are used to extract features from the input image. The self-attention and convolution hybrid module is used to obtain more feature information from the feature map, distinguish different defects, distinguish between background and defects, and reduce the false detection rate and missed detection rate. Ultimately, the features of large-scale and small-scale objects enter the feature fusion network in a balanced state, thereby improving the model's sensitivity to small insulator defects. The feature fusion network module includes a BiFPN with an intra-unit skip structure, namely CAT-BiFPN, which is used to effectively fuse input feature maps of different scales and improve the fusion degree of multi-scale target features; The detection layer module is used to detect features and restore them to the original anchor box.

5. An aerial insulator multi-defect detection device that integrates CAT-BiFPN and an attention mechanism, comprising: Image collector, used to collect aerial images of insulator defects; A program processor, a Jetson nano platform, is used to store a computer program and implement the aerial insulator multi-defect detection method integrating CAT-BiFPN and an attention mechanism as described in any one of claims 1 to 3 when executing the computer program; A display is used to display the detection results of insulator defects.

Images