Power transmission line small target defect identification method based on improved YOLOv8

Abstract

The invention relates to a power transmission line small target defect identification method based on improved YOLOv8, and belongs to the technical field of image identification. The method comprises the following steps: carrying out structure improvement on a Backbone network and a Neck network by taking YOLOv8n as a basic model; in a Backbone network, a C2fScConv module fused with an ScConv module is used for replacing an original C2f module, so that feature redundancy is reduced, and the parameter quantity is reduced; a BiFPN neural network architecture is adopted in the Neck network, and a bidirectional information transmission path is introduced, so that the performance of the feature pyramid network is improved; a P2 small target detection layer is introduced into the improved Neck network and is used for enhancing the multi-scale feature fusion capability and improving the small target recognition precision. The ScConv module enhances the expression ability of useful features by reducing redundant features, and in combination with a multi-scale bidirectional fusion mechanism of BiFPN, the robustness of the model in complex background, illumination variation and electromagnetic interference environments is improved, and the false drop rate is reduced.

Description

Technical Field

[0001] This invention belongs to the field of image recognition technology, specifically relating to a method for identifying small target defects in power transmission lines based on an improved YOLOv8. Background Technology

[0002] YOLOv8 is one of the latest versions in the YOLO (You Only Look Once) object detection series. Released by Ultralytics in January 2023, it inherits the core idea of ​​the YOLO series: treating object detection as a single regression problem and predicting both the object's class and bounding box simultaneously through a single forward propagation. There are five versions, from smallest to largest: v8n, v8s, v8m, v8l, and v8x. As the model size increases, the model accuracy continuously improves. Different network depths and widths can be selected according to the task requirements for object detection, image classification, instance segmentation, and keypoint detection tasks.

[0003] In power transmission line inspection scenarios, defects such as loose bolts, corrosion, and missing bolts are typically small in size, densely distributed, have complex backgrounds, and are subject to significant lighting variations. Traditional YOLOv8 models exhibit significant false negatives and missed detections when handling such small targets, especially in high-voltage transmission line environments where strong background interference and low target pixel ratios further reduce detection reliability. Furthermore, due to hardware resource limitations of embedded robot platforms, it is necessary to achieve model lightweighting while maintaining detection accuracy. Therefore, there is an urgent need for a high-precision, lightweight detection method for small target defects in power transmission lines to improve the accuracy and robustness of defect identification and meet the needs of practical engineering deployments. Summary of the Invention

[0004] The purpose of this invention is to solve the problems of high false negative rate, high false positive rate, large number of model parameters, high computational cost, and difficulty in deployment on embedded platforms in the detection of small target defects in transmission lines using the existing YOLOv8 model, and to provide a method for identifying small target defects in transmission lines based on an improved YOLOv8.

[0005] To achieve the above objectives, the technical solution of the present invention is: a method for identifying small target defects in transmission lines based on an improved YOLOv8, comprising:

[0006] Based on the YOLOv8n model, structural improvements were made to the Backbone and Neck networks.

[0007] In the Backbone network, the original C2f module is replaced by a C2f_ScConv module that incorporates the ScConv module, in order to reduce feature redundancy and the number of parameters.

[0008] BiFPN neural network architecture is adopted in the Neck network, and a bidirectional information transmission path is introduced to improve the performance of the feature pyramid network.

[0009] A P2 small target detection layer is introduced in the improved Neck network to enhance the multi-scale feature fusion capability and improve the small target recognition accuracy.

[0010] Further, the C2f_ScConv module is realized by embedding the ScConv module into the Bottleneck network of the C2f module, specifically including: inserting the ScConv module between the two convolution layers of the Bottleneck network, reconstructing the spatial and channel of the convolution feature map, and outputting the feature map after reducing the redundancy.

[0011] Further, the ScConv module is composed of a spatial reconstruction unit SRU and a channel reconstruction unit CRU, which are respectively used to reduce the spatial and channel dimension feature redundancy.

[0012] Further, the spatial reconstruction unit SRU includes two operations of Separate and Reconstruct; the Separate operation separates the input feature map From the feature map with less corresponding information in the spatial content by group normalization, Sigmoid activation function processing and threshold setting; each parameter of the feature map X Is the batch axis, channel axis, spatial height, and width axis, respectively, and the input feature X is standardized by group normalization processing; after group normalization processing, the related weight is mapped to the range of (0, 1) through the Sigmoid function, and the threshold is used for gating; the weight higher than the threshold is set to 1, and the weight lower than the threshold is set to 0, and the obtained weight is respectively And ; the feature X is multiplied element by element with And , to obtain the feature With less information ; then, the two features with different information are added element by element to obtain the feature With more information And save space; finally, the reconstructed feature And are connected to obtain the spatial reconstruction feature map .

[0013] Further, in the group normalization processing, the variance of the batch and channel spatial pixels is measured by using the trainable parameter γ, and the related weight W is calculated by the following formulaγ :

[0014] ;

[0015] In the formula, Indicates the number of channels. No. One trainable parameter, Indicates all channels The sum of trainable parameters.

[0016] Furthermore, the channel reconstruction unit (CRU) includes three operations: Split, Transform, and Fuse; the Split operation reconstructs the spatial feature map. Perform a segmentation operation to divide the C channels of the feature map into... and Two parts; the Transform operation compresses the features of both parts using 1×1 convolutions, and the compressed channel features are divided into... and Two parts, As upconversion input, group-wise convolution (GWC) and point-wise convolution (PWC) are used to extract rich information features. As the down-conversion input, pointwise convolution PWC extraction is performed; the Fuse operation is used to extract the output features of the up- and down-conversion channels. and Pooling operations are performed to collect global spatial information, and channel attention is used to generate importance vectors for the upper and lower parts of the global information. and The output channel reconstruction features are represented as .

[0017] Furthermore, the output characteristics of the up-conversion channel and These are respectively manifested as:

[0018]

[0019]

[0020] In the formula, For GWC's learning weights, , These are the PwC learning weights for the up and down conversion channels, respectively.

[0021] Furthermore, the BiFPN neural network structure used in the Neck network adds bidirectional propagation between different layers to the original FPN structure. On the original top-down connection structure, bottom-up connections are added, enabling the network to fully utilize the feature information between low-level and high-level layers. Specifically, the feature maps of layers P2, P3, P4, and P5 output from the Backbone network are input into the BiFPN neural network structure. Multi-level feature fusion is performed through multiple upsampling, downsampling, Concat, and C2f modules, and finally, feature maps of four scales, D1, D2, D3, and D4, are output.

[0022] Furthermore, the introduction of the P2 small target detection layer is achieved by upsampling the P3 layer and fusing it with the P2 layer. The specific fusion process includes: fusing the P5 layer and the P4 layer to obtain F1; fusing F1 with the P3 layer after passing through the C2f module and one upsampling to obtain F2; fusing F2 with the C2f module and one upsampling and then through a convolutional layer to obtain F3; fusing F3, F2, and the P2 layer to obtain T1; then fusing T1 with the C2f module to obtain D1; fusing D1 with the P3 layer and F2 with the C2f module to obtain T2; fusing T2 with the C2f module to obtain D2; fusing D2 with the P4 layer and F1 with the C2f module to obtain T3; then fusing T3 with the C2f module to obtain D3; finally, upsampling the P5 layer to obtain F4; fusing F4 and D3 with the convolutional layer to obtain T4; and then fusing T4 with the C2f module to obtain D4.

[0023] Furthermore, the method also includes model training and ablation experiments on a self-constructed transmission line defect dataset. By comparing the precision P, recall R, mean precision mAP, model size and number of parameters under different module combinations, the model combining the ScConv module and the BiFPN neural network architecture is selected as the final recognition model.

[0024] Compared with the prior art, the present invention has the following beneficial effects:

[0025] 1. The method of this invention integrates the C2f module of the feature extraction network in the YOLOv8 model with the ScConv convolutional module, and effectively removes feature redundancy by utilizing its spatial reconstruction unit and channel reconstruction unit. While maintaining model performance, it significantly reduces the number of parameters and computation, realizes the lightweighting of the model, and is more suitable for embedded deployment.

[0026] 2. This invention improves the Neck layer connection method of the model by using a BiFPN network architecture. Through bidirectional connections and adaptive feature adjustment, it enhances the performance of target detection. Simultaneously, a P2 small target detection layer is introduced to increase the network's focus on small targets, thereby improving the accuracy of bolt detection and reducing false positives and false negatives.

[0027] 3. The ScConv module of this invention enhances the expressive power of useful features by reducing redundant features. Combined with the multi-scale bidirectional fusion mechanism of BiFPN, it improves the robustness of the model in complex backgrounds, lighting changes and electromagnetic interference environments, and reduces the false detection rate. Attached Figure Description

[0028] Figure 1 This is a diagram of the improved YOLOv8 model network structure in Example 1.

[0029] Figure 2 This is a structural diagram of the C2f_ScConv module in Example 1.

[0030] Figure 3 This is a structural diagram of the ScConv module in Example 1.

[0031] Figure 4 This is a diagram of the FPN network structure in Example 1.

[0032] Figure 5 This is a diagram of the BiFPN network structure in Example 1. Detailed Implementation

[0033] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0034] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0035] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0036] Example 1

[0037] Reference Figures 1 to 5This is the first embodiment of the present invention, which provides a method for identifying small target defects in transmission lines based on an improved YOLO v8. The YOLOv8n network is the base model, which consists of three networks: Backbone, Neck, and Head. The Backbone network primarily performs feature extraction, employing the CSP (Concurrent Feature Processing) approach to fix the vector size of feature maps at different scales. The Neck network primarily performs feature fusion, employing the PAN-FPN approach and removing the 1×1 convolution used in the upsampling stage, directly upsampling the features output from different stages of the Backbone network. The Head network employs Anchor-Free architecture, eliminating parameter sharing between the classification and regression tasks of YOLOv8.

[0038] This embodiment addresses the characteristics of transmission line defects, such as small size, dense distribution, and strong background interference. It improves the Backbone and Neck sections of the basic model, resulting in the following improved model structure: Figure 1 As shown, it mainly includes basic modules such as ConvModule, C2f, improved C2f, SPPF, Concat, and Upsample. Improvements to the Backbone and Neck sections are as follows:

[0039] Backbone Network Improvement: The C2f module in the Backbone network is replaced with an improved C2f_ScConv. The improved C2f_ScConv structure is as follows: Figure 2 As shown, the improved module effectively limits feature redundancy, reducing not only the number of model parameters and FLOPs, but also enhancing the ability of feature representation.

[0040] Specifically, the C2f_ScConv improvement module merges the ScConv module with the original Bottleneck module in the C2f module, fusing the ScConv module between the two convolutional modules in the feedforward channel of the original system module. The feature map after convolution in the Bottleneck module is then reconstructed spatially and channel-wise by ScConv, outputting a feature map with reduced redundancy. The fused C2f_ScConv improvement module reduces redundant features and improves model performance, achieving a lightweight improvement. The SCConv module structure is as follows... Figure 3 As shown, it mainly consists of spatial reconfiguration units (SRU) and channel reconfiguration units (CRU);

[0041] SRU is a sequence model, a recurrent neural network that is faster and simpler than LSTM. The structure of SRU includes two operations: Separate and Reconstruct. Separate involves steps such as group normalization, the sigmoid activation function, and setting a threshold to transform the input feature map... Feature maps with less corresponding information in the spatial content are separated. The parameters of feature map X are the batch processing axis, channel axis, spatial height and width axis, respectively. The input feature map X is subjected to group normalization to standardize the input features.

[0042] In the GN (group normalization) layer, the variance of batch and channel spatial pixels is measured using a trainable parameter γ, and the relevant weight W can be calculated using the following formula. γ :

[0043] ;

[0044] After group normalization, the obtained weights are mapped to the (0,1) range using the Sigmoid function, and then gating is performed using a threshold. Weights above the threshold are set to 1, and weights below the threshold are set to 0. The resulting weights are then set to... and Combine feature X with and Element-wise multiplication yields features rich in information. and features with less information Next, the features from the two different pieces of information are added element-wise to reconstruct richer information features while saving space. Finally, the reconstructed features are... and Perform connections to obtain spatially reconstructed feature maps (Specifically, here the input feature X will be compared with...) and Element-wise multiplication yields two weighted features: information-rich and redundant features with less information Subsequently, to reduce spatial redundancy, a reconstruction operation was proposed to reconstruct information-rich features. Features with less information Element-wise addition is used to generate more informative features and save space. However, instead of directly adding the two parts, the addition is done along the channel dimension. and Each is divided into two parts , and , and to , and , The reconstruction operation is obtained by pairwise cross-elemental addition. and This allows for the full combination of two different weighted information features and strengthens the interactive information flow between them. Finally, we connect the cross-reconstructed features. and To obtain spatially refined reconstructed feature maps )

[0045] The CRU structure further reduces feature map redundancy across channels, primarily through three operations: Split, Transform, and Fuse. First, the spatial feature map is reconstructed. Perform a segmentation operation to divide the C channels of the feature map into... and Two parts. Then, the two segmented channels are compressed using a 1×1 convolutional module. The compressed channel features are divided into... and Two parts. In the Transform stage, As upconversion input, rich information features are extracted using group-wise convolution (GWC) and point-wise convolution (PWC). For down-conversion input, perform PwC extraction:

[0046] ;

[0047] ;

[0048] In the formula, For GWC's learning weights, , These are the PwC learning weights for the up and down conversion channels, respectively.

[0049] Then, through the Fuse operation, the output characteristics of the up and down conversion channels are modified. and Pooling operations are performed to collect global spatial information, and channel attention is used to generate importance vectors for the upper and lower parts of the global information. and Finally, the output channel reconstruction features are represented as follows: Compared to traditional convolution modules, all parameters in the ScConv module are located in the Transform stage, thus significantly reducing the number of parameters involved in the computation.

[0050] Neck network improvement: The traditional YOLOv8 Neck connection layer uses an FPN network structure

[18] , such as Figure 4As shown. However, the top-down information propagation path of FPN may lead to the problem that lower-level features cannot be fully integrated with higher-level semantic information, thus affecting the model's ability to detect small targets. At the same time, feature maps of different scales need to be adjusted through upsampling and downsampling, which may lead to repeated calculation of feature maps and increase computational cost. This embodiment improves the BiFPN network structure by adding bidirectional propagation between different layers in the original FPN structure, and adding bottom-up connections to the original top-down connection structure, so that the network can fully utilize the feature information between lower and higher levels.

[0051] BiFPN (Bi-directional Feature Pyramid Network) is a neural network architecture for object detection and semantic segmentation tasks, designed to improve the performance of Feature Pyramid Network (FPN) [16-17]. Figure 5 As shown, the BiFPN network structure introduces a bidirectional information transmission path, which can better integrate feature information at different scales and levels.

[0052] Specifically, in the improved Neck layer of this embodiment, a small target detection layer P2 is introduced. The Backbone outputs P2, P3, P4, and P5 are input into the BiFPN network structure, where the multi-scale feature maps are fused. P5 and P4 are fused to obtain F1. F1 is then passed through a C2f layer and upsampled once before being fused with P3 to obtain F2. F2 is then passed through a C2f layer, upsampled once, and then through a convolutional layer to obtain F3. F3, F2, and P2 are fused to obtain T1. T1 is then passed through a C2f layer to obtain D1. D1 is then passed through a convolutional layer, and P3 and F2 are fused through a C2f layer to obtain T2. ​​T2 is then passed through a C2f layer to obtain D2. D2 is then passed through a convolutional layer, and P4 and F1 are fused through a C2f layer to obtain T3. T3 is then passed through a C2f layer to obtain D3. Next, P5 is upsampled once to obtain F4. F4 and D3 are then fused through a convolutional layer to obtain T4, which is then passed through a C2f layer to obtain D4. The final outputs of D1, D2, D3, and D4 are the improved Neckc layer outputs.

[0053] The introduction of the P2 small target detection layer improves the model's ability to detect small targets and its overall performance, while reducing false positives and false negatives in dense bolt detection. Its network connections reduce redundant computation between feature maps of different scales through shared feature computation, improving computational efficiency. Furthermore, the BiFPN network's fewer parameters and lower computational complexity make it easier to deploy on robot vision platforms.

[0054] In summary, this invention effectively reduces feature redundancy and parameter count by replacing the C2f module in the Backbone network with a C2f_ScConv module that integrates ScConv. Furthermore, the introduction of a BiFPN structure and a P2 small target detection layer in the Neck part enhances multi-scale feature fusion capabilities, significantly improving the accuracy and robustness of small target detection. This method, while maintaining a lightweight model, significantly improves the recognition performance for small defects such as power line bolts, making it suitable for practical deployment in embedded robot vision systems.

[0055] Example 2

[0056] This is the second embodiment of the present invention. Based on the first embodiment, this embodiment performs statistical analysis on the results of each improved model, fuses the base model with different attention mechanism modules, and finally selects the optimal model improvement method through experiments. Ablation comparison experiments were conducted on different models on a self-collected dataset, and the experimental results are shown in Table 1.

[0057] Table 1

[0058]

[0059] Conclusion: As shown in Table 1, the ablation experiment results in Experiment 6, with the introduction of the ScConv module and the BiFPN structure, show that the model achieves optimal precision (P), recall (R), and mean accuracy (mAP), which are improved to 96.3%, 94.7%, and 97.7%, respectively. At the same time, the model size (4.28 MB) and the number of parameters (2033092) are further reduced. This indicates that the improved strategy of combining ScConv and BiFPN proposed in this invention can improve the small target detection performance while achieving model lightweighting, and has good practicality and deployment feasibility.

[0060] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A power transmission line small target defect recognition method based on improved YOLOv8, characterized in that, The application relates to a power transmission line defect detection method based on a YOLOv8n model. The structure of a Backbone network and a Neck network is improved based on the YOLOv8n model; In the Backbone network, a C2f_ScConv module integrating a ScConv module is used to replace an original C2f module to reduce feature redundancy and reduce the parameter quantity; In the Neck network, a BiFPN neural network architecture is used, and a bidirectional information transmission path is introduced to improve the performance of a feature pyramid network; In the improved Neck network, a P2 small target detection layer is introduced to enhance the multi-scale feature fusion capability and improve the small target recognition precision. The C2f_ScConv module is realized by embedding the ScConv module into the bottleneck network of the C2f module, and specifically comprises the following steps: inserting the ScConv module between two convolution layers of the bottleneck network, reconstructing the spatial and channel features of the convolution features, and outputting the features after reducing the redundancy.

2. The power line small target defect recognition method based on improved YOLOv8 according to claim 1, characterized in that, The ScConv module is composed of a spatial reconstruction unit SRU and a channel reconstruction unit CRU, and is used for reducing the spatial and channel feature redundancy.

3. The power line small target defect recognition method based on improved YOLOv8 according to claim 2, characterized in that, The BiFPN neural network structure used in the Neck network is a bidirectional propagation between different levels in the original FPN structure, and the bottom-up connection is added to the original top-down connection structure, so that the network can fully utilize the feature information between the low-level and high-level, and specifically comprises the following steps: inputting P2, P3, P4 and P5 layer feature maps output by the Backbone network into the BiFPN neural network structure, performing multi-level feature fusion through multiple up-sampling, down-sampling, Concat and C2f modules, and finally outputting D1, D2, D3 and D4 four-scale feature maps.

4. The power line small target defect identification method based on improved YOLOv8 according to claim 3, characterized in that, The space reconstruction unit SRU includes two operations of Separate and Reconstruct; the Separate operation separates the input feature map by group normalization, Sigmoid activation function processing and setting a threshold from the feature map with less corresponding information in the spatial content; each parameter of the feature map X is respectively a batch axis, a channel axis, a spatial height, and a width axis, and the input feature map X is standardized by group normalization processing. After group normalization processing, the obtained correlation weight is mapped to the range of (0, 1) through a Sigmoid function and gated using a threshold; the weight higher than the threshold is set to 1 and the weight lower than the threshold is set to 0, and the obtained weight is respectively set to With ; the feature map X is element-wise multiplied with With to obtain a feature with rich information and a feature with less information ; then, the two features with different information are element-wise added to obtain a feature with more rich information With and save space; finally, the reconstructed feature With is connected to obtain a spatial reconstruction feature map .

5. The power line small target defect identification method based on improved YOLOv8 according to claim 4, characterized in that, In the group normalization processing, the variance of the batch and the channel spatial pixels is measured by using a trainable parameter γ, and the relevant weight W is calculated by the following formula γ : ; In the formula, Indicates the number of channels. No. One trainable parameter, Indicates all channels The sum of trainable parameters.

6. The power line small target defect identification method based on improved YOLOv8 according to claim 3, characterized in that, The channel reconstruction unit (CRU) includes three operations: Split, Transform, and Fuse; the Split operation reconstructs the spatial feature map. Perform a segmentation operation to divide the C channels of the feature map into... and Two parts; the Transform operation compresses the features of both parts using 1×1 convolutions, and the compressed channel features are divided into... and Two parts, As the upconversion input, group-wise convolution (GWC) and point-wise convolution (PWC) are used to extract rich information features. As the down-conversion input, pointwise convolution PWC extraction is performed; the Fuse operation is applied to the output features of the up- and down-conversion channels. and Pooling operations are performed to collect global spatial information, and channel attention is used to generate importance vectors for the upper and lower parts of the global information. and The output channel reconstruction features are represented as .

7. The power line small target defect identification method based on improved YOLOv8 according to claim 6, characterized in that, Output characteristics of the up-conversion channel With respectively embodied as: wherein is the learning weight for the GWC, , are the PWC learning weights for the up and down conversion channels, respectively.

8. The power line small target defect identification method based on improved YOLOv8 according to claim 1, characterized in that, The P2 small target detection layer is introduced by fusing the up-sampled P3 layer with the P2 layer; 9. The power line small target defect identification method based on improved YOLOv8 according to claim 8, characterized in that, The specific fusion process comprises the following steps: fusing the P5 layer and the P4 layer to obtain F1, fusing F1 through a C2f module and once up-sampling with the P3 layer to obtain F2, fusing F2 through a C2f module and once up-sampling and through a convolution layer to obtain F3, fusing F3, F2 and the P2 layer to obtain T1, fusing T1 through a C2f module to obtain D1, fusing D1 through a convolution layer and the P3 layer with F2 through a C2f module to obtain T2, fusing T2 through a C2f module to obtain D2, fusing D2 through a convolution layer and the P4 layer with F1 through a C2f module to obtain T3, fusing T3 through a C2f module to obtain D3, then, up-sampling the P5 layer to obtain F4, fusing F4 with D3 through a convolution layer to obtain T4, and fusing T4 through a C2f module to obtain D4. The method further comprises model training and ablation experiment on a self-constructed power transmission line defect dataset, the precision P, the recall rate R, the average precision mAP, the model size and the parameter quantity under different module combinations are compared, and the model combining the ScConv module and the BiFPN neural network architecture is selected as the final recognition model.

10. The power line small target defect recognition method based on improved YOLOv8 according to claim 1, characterized in that, ​

Images