Double-backbone network insulator defect detection method based on YOLOv5

Abstract

The invention discloses a double-backbone network insulator defect detection method based on YOLOv5. The method comprises the following steps: constructing a GELAN-Ghost double-backbone network as a new backbone network of a YOLOv5 model to extract feature information of an input image; a plug-and-play high-efficiency multi-scale attention module iEMA is designed; based on an attention module iEMA, specifically adding an iEMA module in a small target detection layer and a medium target detection layer of the model, and constructing an iFPN-PAN structure; according to the method, a novel DynamicHeadv3 detection head is constructed, and a deformable convolution and attention mechanism is utilized to extract richer features and enhance the perception ability of a model; and 4) taking a YOLOv5 model as a reference, introducing the GELAN-Ghost double-backbone network proposed in the step 1), the step 2) and the step 4), an iEMA attention mechanism and a DynamicHeadv3 detection head to improve the YOLOv5 model, and constructing the double-backbone network insulator defect detection method based on the YOLOv5.

Description

Technical Field

[0001] The present invention belongs to the technical field of visual computing and processing of electric power systems, and in particular relates to a dual-trunk network insulator defect detection method based on YOLOv5. Background Art

[0002] In the power system, insulators, as one of the core equipment in the transmission line, play a vital role in the safe transmission of electricity. However, due to long-term environmental exposure and power load, insulators will be damaged, flashover and other defects, which brings great challenges to power transmission safety.

[0003] Early manual inspections were costly and inefficient. In recent years, drone inspections and real-time monitoring of edge devices have gradually replaced traditional manual inspections. The YOLO series of algorithms are more widely used in edge devices such as drones and cameras because their detection speed is faster than two-stage methods such as R-CNN and Fast R-CNN, while also taking into account higher detection accuracy. In the field of insulator defect detection research, researchers have actively explored and improved the YOLO series of algorithms to meet the needs of different application scenarios. Scholars mainly add new research results to the backbone network, neck network and other structures of the YOLO model, or replace the original loss function, activation function, etc. to improve the performance of the model. Although the above research methods have improved the overall performance of the model to a certain extent, there are still problems such as insufficient feature extraction, poor small target detection, and unbalanced overall algorithm performance.

[0004] Therefore, in the insulator defect detection scenario, it is urgent to design a detection model with balanced performance in all aspects to solve the problems of insufficient feature extraction, poor small target detection effect, and overall unbalanced algorithm performance in the target detection algorithm currently deployed on edge devices. This model needs to maintain high accuracy while overcoming the performance degradation that may be caused by the lightweight process, and the detection speed must also be guaranteed to ensure that insulator defects can be effectively detected in practical applications, thereby ensuring the safety of transmission lines.

[0005] The information disclosed in this background technology section is only intended to enhance the understanding of the overall background of the invention and should not be regarded as an acknowledgment or any form of suggestion that the information constitutes the prior art already known to a person skilled in the art. Summary of the invention

[0006] The purpose of the present invention is to provide a dual-trunk network insulator defect detection method based on YOLOv5, aiming to solve the problems of insufficient feature extraction, poor small target detection effect, and unbalanced overall algorithm performance in the insulator defect detection algorithm currently deployed on edge devices, so as to achieve better performance in the insulator defect detection scenario.

[0007] To achieve the above object, the present invention provides a dual-trunk network insulator defect detection method based on YOLOv5, comprising the following steps: Step 1) Construct a GELAN-Ghost dual backbone network as the new backbone network of the YOLOv5 model to extract feature information of the input image.

[0008] Step 2) A plug-and-play inverted efficient multi-scale attention module iEMA (inverted efficient multi-scale attention) is designed, which enables the model to capture the global dependencies between different parts of the input data and thus process long-distance information and feature information of small targets more effectively.

[0009] Step 3) Based on the attention module iEMA in step 2), iEMA modules are added to the small target detection layer and medium target detection layer of the model to construct the iFPN-PAN structure.

[0010] Step 4) A novel DynamicHeadv3 detection head is constructed, which utilizes deformable convolution and attention mechanisms to extract richer features and enhance the perception ability of the model.

[0011] Step 5) Taking the YOLOv5 model as the benchmark, the GELAN-Ghost dual-trunk network, iEMA attention mechanism, and DynamicHeadv3 detection head proposed in steps 1), 2), and 4) are introduced to improve the YOLOv5 model, and a dual-trunk network insulator defect detection method based on YOLOv5 is constructed.

[0012] The network structure diagram mainly includes three parts, namely Backbone, Neck and Head. Backbone is the backbone network of the model, responsible for feature extraction of the input image, which is completed by the dual backbone network GELAN-Ghost; Neck is the neck feature fusion network of the model, responsible for combining feature maps of different levels for top-down and bottom-up joint feature fusion, obtaining rich contextual information, and enhancing semantic expression and positioning capabilities at multiple scales, which is completed by the iFPN-PAN structure; Head is the detection layer network of the model, consisting of three DynamicHeadv3 detection heads, which uses deformable convolution and attention mechanisms to extract richer features and enhance the perception ability of the model.

[0013] Preferably, in the above technical solution, step 1) constructs a GELAN-Ghost dual-trunk network, which mainly addresses the problem that the existing lightweight algorithm is not sufficient for feature extraction and loses a lot of information in the feature extraction feedforward process. The network structure includes three parts, in which the main branch is used for the reasoning process, the auxiliary reversible branch is used to solve the problems caused by the depth of the neural network, and the multi-level auxiliary information is used to deal with the error accumulation problem caused by deep supervision, especially for the architecture and lightweight model of multiple prediction branches.

[0014] Preferably, in the above technical solution, step 2) designs a new plug-and-play efficient multi-scale attention module iEMA, which uses an inverted residual structure to fuse dynamic global modeling and static local information, and cascades EMA and Conv modules to increase the receptive field of the model and improve its ability for downstream tasks. This enables the model to capture the global dependencies between different parts of the input data and thus more effectively process long-distance information and feature information of small targets.

[0015] Preferably, in the above technical solution, step 3) is based on the efficient multi-scale attention module iEMA in step 2), and the iEMA module is added to the small target detection layer and the medium target detection layer of the model in a targeted manner, so as to construct a new feature fusion module iFPN-PAN structure, which can improve the ability of small target detection while better ensuring the lightweight of the model.

[0016] Preferably, in the above technical solution, step 4) improves the DynamicHead detection head based on the DCNv3 module, and proposes a new DynamicHeadv3 detection head, which uses deformable convolution and attention mechanisms to extract richer features and enhance the perception ability of the model. By introducing new features such as the shared weight mechanism, multi-grouping mechanism, and normalized modulation scalar of the DCNv3 operator, the model can achieve better performance during training.

[0017] Preferably, in the above technical solution, step 5) takes the YOLOv5 model as a benchmark, introduces the GELAN-Ghost dual-trunk network, iEMA attention mechanism, and DynamicHeadv3 detection head proposed in steps 1), 2), and 4) to improve the YOLOv5 model, and constructs a dual-trunk network insulator defect detection method based on YOLOv5.

[0018] Preferably, in the above technical solution, step 1) constructs a dual-trunk feature extraction network GLEAN-Ghost, and introduces the C3_Ghost module while modifying the GLEAN structure to form a dual-trunk feature extraction network suitable for YOLOv5; the dual-trunk network structure mainly includes three parts, wherein the main branch is used for the reasoning process, the auxiliary reversible branch is used to solve the problems caused by the depth of the neural network, and the multi-level auxiliary information is used to deal with the error accumulation problem caused by deep supervision; the C3_Ghost module is an improvement on the C3 module, combining the characteristics of the C3 module and the advantages of the Ghost module; the C3 module contains three convolutional layers, which can extract rich features; Step 2) Design a plug-and-play efficient multi-scale attention module iEMA; the iEMA module structure adopts an inverted residual structure to fuse dynamic global modeling and static local information, cascading EMA and Conv modules so that the model can capture the global dependencies between different parts of the input data and more effectively process the feature information of long-distance information and small targets; the EMA module first performs feature grouping. For any given input feature map EMA divides X into G sub-features in the cross-channel dimension direction to learn different semantics; then it starts parallel subnetwork operation, and EMA uses three parallel routes to extract the attention weights of the grouped feature maps, two of which are located in the 1x1 branch and the third is located in the 3x3 branch; it models the cross-channel information interaction in the channel direction; finally, it performs cross-space learning, and EMA uses the cross-space information aggregation method in different spatial dimensions to achieve richer feature aggregation; Step 3) Based on the YOLOv5 model, the iEMA module is added to the small target detection layer and the medium target detection layer to construct the iFPN-PAN structure; Step 4) Improve the DynamicHead detection head based on the DCNv3 module. The structure contains three key attention functions, which are used for scale self-attention, spatial self-attention and multi-task self-attention. These attention functions are combined by stacking. The calculation formula of DCNv3 is as follows:

[0019] For the g-th group, Represents the position-independent projection weight of the group, where Indicates the dimension of the group. represents the modulation scalar of the k-th sampling point, normalized by the softmax function along the k-dimension; represents the input feature map after slicing; Δpgk is the offset corresponding to the sampling position pk of the g-th group of grids; Step 5) Taking the YOLOv5 model as the benchmark, the YOLOv5 model is optimized based on the GLEAN-Ghost backbone network, iFPN-PAN structure, and DynamicHeadv3 detection head to construct a dual-backbone network insulator defect detection method based on YOLOv5.

[0020] Compared with the prior art, the present invention has the following beneficial effects: In the research of insulator defect detection model, scholars mainly improve the performance of the model by adding new research results to the backbone network, neck network and other structures of the model, or replacing the original loss function, activation function, etc. Although the defect detection accuracy of the model is improved to a certain extent, the performance of feature extraction, small target detection effect and overall algorithm performance is still not good enough. The present invention constructs a dual-trunk feature extraction network GELAN-Ghost suitable for YOLOv5 to solve the problem of the model losing a lot of information in the feature extraction feedforward process to enhance the feature extraction ability of the model. In the feature fusion stage of the model, the inverted residual efficient multi-scale attention module iEMA is introduced in a targeted manner. The iFPN-PAN structure is constructed based on the iEMA module to better ensure the lightweight of the model while improving the ability of small target detection. The DynamicHead detection head is improved based on the DCNv3 module, and a new DynamicHeadv3 detection head is proposed. The DynamicHeadv3 structure has stronger expression ability and better generalization performance, and can more accurately detect targets of various sizes and categories, improves the accuracy and robustness of j insulator defect detection, and brings better effects to practical application scenarios. Experiments on professional and real insulator defect data sets show that the present invention has achieved an effective balance in the three aspects of detection accuracy, inference speed and model scale, which is more in line with the requirements of insulator defect detection and edge device deployment. The experimental results show that the accuracy of the improved model has increased by 1.4%, the number of parameters and the amount of calculation have decreased by 46% and 33% respectively, the detection speed has also been improved to a certain extent, and the performance has achieved a good balance, which is more in line with the requirements of drone and edge insulator defect detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 The figure is a flowchart of a dual-trunk network insulator defect detection method based on YOLOv5.

[0022] Figure 2 This is a network structure diagram of a dual-trunk network insulator defect detection method based on YOLOv5. The trunk network part is the GELAN-Ghost dual-trunk network diagram, and the neck network part is the iFPN-PAN structure diagram. The upper right part is the Conv_BN_SiLU and C3_Ghost structure diagram.

[0023] Figure 3 It is the GLEAN structure of the GELAN-Ghost dual-trunk network graph in a dual-trunk network insulator defect detection method based on YOLOv5.

[0024] Figure 4 This is a structural diagram of the iEMA module that incorporates an efficient attention mechanism in a dual-trunk network insulator defect detection method based on YOLOv5.

[0025] Figure 5 This is a diagram of the EMA module structure in the iEMA module structure in a dual-trunk network insulator defect detection method based on YOLOv5.

[0026] Figure 6 This is a structural diagram of the DynamicHeadv3 detection head in a dual-trunk network insulator defect detection method based on YOLOv5. DETAILED DESCRIPTION The specific embodiments of the present invention are described in detail below, but it should be understood that the protection scope of the present invention is not limited by the specific embodiments.

[0028] Unless explicitly stated otherwise, throughout the specification and claims, the term “comprise” or variations such as “include” or “comprising”, etc., will be understood to include the stated elements or components but not to exclude other elements or components.

[0029] like Figure 1 As shown, it is a step flow chart of a dual-trunk network insulator defect detection method based on YOLOv5 provided by the present invention, including building a GELAN-Ghost dual-trunk network, designing an efficient multi-scale attention module iEMA, building an iFPN-PAN structure based on the iEMA module, building a DynamicHeadv3 detection head, and building an insulator defect detection algorithm based on GELAN-Ghost, iFPN-PAN, and DynamicHeadv3 based on the YOLOv5 algorithm.

[0030] like Figure 2As shown, it is a network structure diagram of a dual-backbone network insulator defect detection method based on YOLOv5 provided by the present invention. It mainly includes three parts, namely Backbone, Neck and Head. Backbone is the backbone network of the model, which is responsible for feature extraction of the input image, and is completed by the lightweight dual-backbone network structure GELAN-Ghost; Neck is the neck feature fusion network of the model, which is responsible for combining feature maps of different levels to perform top-down and bottom-up joint feature fusion, obtain rich contextual information, and enhance the semantic expression and positioning capabilities of multiple scales, which is completed by the iFPN-PAN structure; Head is the detection layer network of the model, which is composed of three DynamicHeadv3 detection heads, and uses deformable convolution and attention mechanisms to extract richer features and enhance the perception ability of the model. Figure 2 The upper right part is a detailed description of the Conv_BN_SiLU and C3_Ghost structures. The Conv_BN_SiLU structure is a basic structure consisting of a convolutional layer, a batch normalization layer, and a SiLU activation function; C3_Ghost is an improved CSPNet structure that combines the lightweight convolution technology of GhostNet. It not only maintains the diversity and integrity of features when the network depth increases, but also significantly reduces the computational complexity and parameter amount of the model.

[0031] The specific construction process of a dual-trunk network insulator defect detection method based on YOLOv5 is as follows: Step 1) Construct a dual-backbone feature extraction network GELAN-Ghost. Figure 2 As shown in the backbone network, the GLEAN structure is modified and the C3_Ghost module is introduced to ensure the lightweight structure, and a lightweight dual-backbone feature extraction network suitable for YOLOv5 is designed. Figure 3 As shown in the figure, in the process of extracting features from the feature extraction network, as the network level deepens, there may be information loss in the input data in the feedforward stage. As shown in the following formula, I represents mutual information, f and g are transformation functions, and are the parameters of f and g respectively. In deep neural networks, and Represent the operations of two consecutive layers in a deep neural network. From the above formula 1, it can be seen that as the number of network layers increases, the possibility of original data loss will also increase. This information loss may cause the deep network to establish incorrect associations between the target and the input, resulting in incorrect predictions in the training model. The GLEAN structure can effectively solve this problem by generating reliable gradients through auxiliary reversible branches, so that deep features can still maintain key features to perform target tasks. Figure 3 As shown in , the dual-trunk network structure mainly includes three parts, among which the main branch is used for the reasoning process, the auxiliary reversible branch is used to solve the problems caused by the depth of the neural network, and the multi-level auxiliary information is used to deal with the error accumulation problem caused by deep supervision, especially for the architecture of multiple prediction branches and lightweight models. The C3_Ghost module is an improvement on the C3 module, combining the characteristics of the C3 module and the advantages of the Ghost module. The C3 module usually contains three convolutional layers and can extract rich features. C3_Ghost generates lightweight feature maps using the Ghost module to reduce the amount of calculation. Its design makes the model more efficient while maintaining performance, and is particularly suitable for use in resource-constrained environments. Among them, Ghost convolution is a convolution module proposed by Huawei Noah Laboratory in 2020. The main idea is that the features on different channels obtained by convolution are relatively similar, which can reduce the number of channels of the convolution output. The feature channels are expanded by linearly transforming the features on the channels obtained by convolution, and the same results can be obtained while reducing FLOPS. The GELAN-Ghost dual backbone network was used as the backbone network of YOLOv5 for experiments. The number of model parameters and computational complexity were greatly reduced. Compared with the original YOLOv5 model, the number of parameters was reduced from 7.2×10 6 Down to 3.0×10 6 , the computational complexity is reduced by 15.8×10 9 Down to 8.2×10 9 .

[0032] Step 2) A plug-and-play efficient multi-scale attention module iEMA is designed. The iEMA module structure is as follows Figure 4 As shown in the figure, the inverted residual structure can be used to fuse dynamic global modeling and static local information, and the EMA and Conv modules can be cascaded to increase the receptive field of the model and improve the ability of downstream tasks. This enables the model to capture the global dependencies between different parts of the input data and process long-distance information and small target feature information more effectively. The EMA structure is as follows Figure 5 As shown, the EMA module first performs feature grouping. For any given input feature map EMA divides X into G sub-features in the cross-channel dimension direction to learn different semantics. Then the parallel subnet operation begins. Because the huge local receptive field of neurons enables neurons to collect multi-scale spatial information, EMA uses three parallel routes to extract the attention weights of the grouped feature maps, two of which are located in the 1x1 branch and the third in the 3x3 branch. In order to capture dependencies across all channels and reduce the computational budget, the cross-channel information interaction in the channel direction is modeled. Finally, for cross-space learning, EMA uses a cross-spatial information aggregation method in different spatial dimension directions to achieve richer feature aggregation.

[0033] Step 3) If Figure 2 As shown in the middle neck network part, based on the YOLOv5 model, the iEMA module is added to the small target detection layer and the medium target detection layer of the model, and the iFPN-PAN structure is constructed, which improves the ability of small target detection while better ensuring the lightweight of the model. After using the iFPN-PAN structure, the accuracy and average precision of the model exceed the original YOLOv5s model, reaching 94.4% and 92.1% respectively.

[0034] Step 4) Based on the DCNv3 module, the DynamicHead detection head is improved, and a new DynamicHeadv3 detection head is proposed. This detection head uses deformable convolution and attention mechanism to extract richer features and enhance the perception ability of the model. Compared with the traditional detection head, the DynamicHeadv3 structure has stronger expression ability and better generalization performance, can more accurately detect objects of various sizes and categories, improves the accuracy and robustness of object detection, and brings better results for practical application scenarios. Figure 6 As shown in the figure, the structure contains three key attention functions, which are used for scale self-attention, spatial self-attention and multi-task self-attention. These attention functions are stacked and combined to construct an efficient multi-dimensional attention learning framework, providing DynamicHeadv3 with powerful feature expression capabilities. Among them, DCNv3 is the latest research result in the direction of deformable convolution. By introducing new features such as shared weight mechanism, multi-grouping mechanism and normalized modulation scalar, the model can achieve better performance during training. The DCNv3 operator makes up for the shortcomings of standard convolution in long-distance dependence and adaptive spatial aggregation, making the model more efficient and with higher computational and memory efficiency when there is less training data and shorter training time. The calculation formula of DCNv3 is shown below, For the g-th group, Represents the position-independent projection weight of the group, where Indicates the dimension of the group. represents the modulation scalar of the k-th sample point, normalized by the softmax function along the k-th dimension. represents the input feature map after slicing. Δpgk is the sampling position of the g-th group of grids The corresponding offset.

[0035] Step 5) Taking the YOLOv5 model as the benchmark, the YOLOv5 model is optimized based on the GLEAN-Ghost backbone network, iFPN-PAN structure, and DynamicHeadv3 detection head to construct a dual-backbone network insulator defect detection method based on YOLOv5. In general, compared with the YOLOv5 algorithm, this method has improved the accuracy by 1.4% and the average precision by 2.1% while reducing the number of parameters and the amount of calculation by 46% and 33% respectively. The detection speed has been improved to a certain extent, so that the overall performance of the model has been well balanced.

[0036] The foregoing description of specific exemplary embodiments of the present invention is for the purpose of illustration and demonstration. These descriptions are not intended to limit the present invention to the precise form disclosed, and it is clear that many changes and variations can be made based on the above teachings. The purpose of selecting and describing the exemplary embodiments is to explain the specific principles of the present invention and its practical application, so that those skilled in the art can realize and utilize various different exemplary embodiments of the present invention and various different selections and changes. The scope of the present invention is intended to be limited by the claims and their equivalents.

Claims

1. A dual-trunk network insulator defect detection method based on YOLOv5, characterized in that: The following steps are involved: Step 1) Build a GELAN-Ghost dual-backbone network as the new backbone network of the YOLOv5 model to extract feature information from the input image; the network structure consists of three parts: the main branch is used for the reasoning process, the auxiliary reversible branch is used to solve the problem of neural network depth, and the multi-level auxiliary information is used to deal with the error accumulation problem caused by deep supervision; Step 2) Design a plug-and-play efficient multi-scale attention module iEMA. The iEMA module combines dynamic global modeling and static local information through an inverted residual structure. By cascading EMA and Conv modules, the model can capture the global dependencies between various parts of the input data, thereby processing the characteristics of long-distance information and small targets. Step 3) Add iEMA modules to the small target detection layer and medium target detection layer of the model to build an improved iFPN-PAN structure to improve the detection capabilities of different types of targets; Step 4) A new DynamicHeadv3 detection head is constructed, which uses deformable convolution and attention mechanisms to extract richer features and enhance the perception ability of the model; Step 5) Taking the YOLOv5 model as the benchmark, the GELAN-Ghost dual-trunk network, iEMA attention mechanism, and DynamicHeadv3 detection head proposed in steps 1), 2), and 4) are introduced to improve the YOLOv5 model, and a dual-trunk network insulator defect detection method based on YOLOv5 is constructed; The network structure diagram mainly includes three parts, namely Backbone, Neck and Head. The Backbone part adopts the GELAN-Ghost dual backbone network, which is responsible for extracting features from the input image; The Neck part uses the iFPN-PAN structure to achieve the joint fusion of feature maps at different levels to obtain richer contextual information and enhance multi-scale semantic expression and positioning accuracy; the Head part consists of three DynamicHeadv3 detection heads, which use deformable convolution and attention mechanisms to improve the quality of feature extraction and the perception ability of the model.

2. The dual-trunk network insulator defect detection method based on YOLOv5 according to claim 1 is characterized in that: Step 2) A new plug-and-play efficient multi-scale attention module iEMA is designed. The inverted residual structure can fuse dynamic global modeling and static local information. The EMA and Conv modules are cascaded to increase the receptive field of the model. The model can capture the global dependencies between different parts of the input data and thus process long-distance information and feature information of small targets more effectively.

3. The dual-trunk network insulator defect detection method based on YOLOv5 according to claim 1 is characterized in that: Step 3) Based on the efficient multi-scale attention module iEMA in step 2), the iEMA module is added to the small target detection layer and the medium target detection layer of the model to construct a new feature fusion module iFPN-PAN structure to improve the ability of small target detection while better ensuring the lightweight of the model.

4. The dual-trunk network insulator defect detection method based on YOLOv5 according to claim 1 is characterized in that: Step 4) Based on the DCNv3 module, the DynamicHead detection head is improved, and a new DynamicHeadv3 detection head is proposed, which utilizes deformable convolution and attention mechanisms to extract richer features and enhance the perception ability of the model. By introducing new features such as the shared weight mechanism, multi-grouping mechanism, and normalized modulation scalar of the DCNv3 operator, the model can achieve better performance during training.

5. The dual-trunk network insulator defect detection method based on YOLOv5 according to claim 1 is characterized in that: Step 5) Taking the YOLOv5 model as the benchmark, the GELAN-Ghost dual-trunk network, iEMA attention mechanism, and DynamicHeadv3 detection head proposed in steps 1), 2), and 4) are introduced to improve the YOLOv5 model, and a dual-trunk network insulator defect detection method based on YOLOv5 is constructed.

6. The dual-trunk network insulator defect detection method based on YOLOv5 according to claim 1 is characterized in that: Step 1) Construct a dual-trunk feature extraction network GLEAN-Ghost, and introduce the C3_Ghost module while modifying the GLEAN structure to form a dual-trunk feature extraction network suitable for YOLOv5; the dual-trunk network structure mainly includes three parts: the main branch is used for the reasoning process, the auxiliary reversible branch is used to solve the problems caused by the depth of the neural network, and the multi-level auxiliary information is used to deal with the error accumulation problem caused by deep supervision; the C3_Ghost module is a combination of the C3 module and the Ghost module; the C3 module contains three convolutional layers, which can extract rich features; Step 2) Design a plug-and-play efficient multi-scale attention module iEMA; the iEMA module structure adopts an inverted residual structure to fuse dynamic global modeling and static local information, cascading the EMA and Conv modules so that the model can capture the global dependencies between different parts of the input data and process long-distance information and small target feature information more effectively; The EMA module first performs feature grouping. For any given input feature map EMA will divide X into G sub-features in the cross-channel dimension direction to learn different semantics; then the parallel subnetwork operation begins, and EMA uses three parallel routes to extract the attention weights of the grouped feature maps, two of which are located in the 1x1 branch and the third is located in the 3x3 branch; Model the cross-channel information interaction in the channel direction; finally, perform cross-space learning. EMA uses the cross-space information aggregation method in different spatial dimensions to achieve richer feature aggregation; Step 3) Based on the YOLOv5 model, the iEMA module is added to the small target detection layer and the medium target detection layer to construct the iFPN-PAN structure; Step 4) Improve the DynamicHead detection head based on the DCNv3 module. The structure contains three key attention functions, which are used for scale self-attention, spatial self-attention and multi-task self-attention. These attention functions are combined by stacking. The calculation formula of DCNv3 is as follows:

7. For the g-th group, Represents the position-independent projection weight of the group, where Indicates the dimension of the group. represents the modulation scalar of the k-th sampling point, normalized by the softmax function along the k-dimension; represents the input feature map after slicing; Δpgk is the offset corresponding to the sampling position pk of the g-th group of grids; Step 5) Taking the YOLOv5 model as the benchmark, the YOLOv5 model is optimized based on the GLEAN-Ghost backbone network, iFPN-PAN structure, and DynamicHeadv3 detection head to construct a dual-backbone network insulator defect detection method based on YOLOv5.

Images