In the task of detecting foreign objects on power transmission lines, overcoming complex background interference and acquiring small target features are the key points. Although YOLOv8 achieves effective feature extraction in the backbone network through C2f module, the traditional small-scale convolution kernel (3 × 3) in its backbone network has problems such as limited receptive field and insufficient long-range dependency modeling capabilities. It is difficult to capture global contextual information in scenarios with large transmission line spans. This paper integrates the LarK block module of UniRepLKNet into C2f module to construct the C2f_LarK module. The LarK block module captures more contextual information through a larger receptive field, enabling the backbone network to capture richer features. At the same time, the re-parameterized nature of the LarK block can significantly improve the efficiency of the model during inference, making it more conducive to deployment on edge devices. The structure design of C2f_UniRepLKNet is shown as Figure 2.

In Figure 2, the LarK Block module achieves efficient feature extraction through a dual-path design of dilated reparameterization and channel attention enhancement (Hu et al., 2018). Its core is to use dilated convolution to reconstruct the parameter space of small-kernel convolutions, re-parameterizing the multi-branch structure into a single sparse large-kernel convolution through mathematical equivalence. Specifically, during the training phase, the module deploys non-dilated large-kernel convolutions (e.g., 13 × 13) and dilated small-kernel convolutions (e.g., 3 × 3, with a dilation rate of d = 4) in parallel, complementing their parameter spaces through gradient co-optimization. During the inference phase, leveraging the additivity principle of convolution, the multi-branch structure is mathematically equivalently reparameterized. Specifically, in physical implementation, we deploy it as a depth-wise dilated convolution. This implementation directly utilizes the underlying operator optimizations of dilated and depth-wise convolutions in mainstream hardware (such as GPUs and NPUs), achieving a receptive field equivalent to a 31 × 31 large core without explicit sparse matrix operations. This avoids the hardware efficiency degradation issues caused by unstructured sparsity. Furthermore, the module integrates the Squeeze-and-Excitation (SE) attention mechanism. Global Average Pooling (GAP) is used to compress the spatial dimensions to generate channel description vectors. Channel attention weights are then generated through a fully connected layer and sigmoid activation to dynamically calibrate feature channel responses.

Regarding the selection of the core kernel size for this module, this study strictly adheres to the ablation conclusions from the original UniRepLKNet (Ding et al., 2024) and designates 31 × 31 as the optimal kernel dimension. Research has substantiated that this size represents a “saturation point” for balancing receptive field gains and inference latency. Compared to larger kernels (e.g., 51 × 51), a 31 × 31 kernel provides a large-scale receptive field sufficient to cover the wide-span scenarios of transmission lines with negligible increases in Memory Access Cost (MAC). This enables efficient feature capture while effectively circumventing excessive computational overhead.

In traditional YOLOv8 object detection model, 25% of its parameters come from the detection head. Conventional convolutional layers (Conv + BN + SiLU) are often used for category classification and bounding box regression, but the introduced parameter amount and computational complexity limit their deployment capabilities on edge devices. To optimize this problem, this paper proposes to replace the standard convolution operation in original classification branch with Partial Convolution (PConv) of FasterNet. This allows for the construction of a new lightweight object detection head, Det_Tiny. Through adaptive feature redundancy compression and hardware-friendly computational optimization, the algorithm significantly reduces computational complexity while maintaining classification accuracy. The PConv structure design is shown in Figure 3.

Traditional standard convolution performs intensive calculations on all input channels in the spatial dimension. However, in the scenario of foreign object detection on power transmission lines, the feature distribution of background areas (such as the sky and vegetation) and target areas (kites and plastic bags) is significantly spatially sparse. Direct application of standard convolution will lead to a large number of redundant calculations. Especially in classification tasks, high-frequency detail features (such as foreign object edge textures) contribute much more to category judgment than smooth background areas. PConv performs spatial convolution only on some channels through channel grouping and selective calculation. The remaining channels are processed using low-cost point-by-point convolution (PWConv) to reduce redundant operations. Its mathematical expression is as Equation 1. F o u t = Concat Conv k × k F i n 1 : g , PWConv F i n g + 1 : C i n (1) Where g is the number of channels involved in spatial convolution. By separating high-frequency feature channels (required for spatial detail extraction) from low-frequency channels (required for channel information fusion), a balance between computational efficiency and feature expression is achieved.

In this study, adhering to the ablation experimental conclusions from the original FasterNet authors (Chen et al., 2023), we set this parameter to 1/4 of the total input channels (g = 1/4). Research has demonstrated that this ratio serves as the optimal balance point between computational efficiency and feature representation capability. It ensures that a sufficient number of channels are dedicated to extracting critical high-frequency spatial features of foreign objects, while simultaneously maximizing the utilization of the remaining channels for low-cost fusion to reduce Giga Floating-point Operations Per Second (GFLOPs). An excessively low ratio leads to insufficient feature extraction, whereas an overly high ratio diminishes the lightweight advantages. Thus, the configuration of g = 1/4 effectively satisfies the dual requirements for both accuracy and speed.

To address the accuracy loss caused by lightweight model, this paper proposes a multi-scale feature fusion network, Fusion, based on CGAFusion. Structure design of CGAFusion is shown in Figure 4.

The CGAFusion module fuses shallow features with deep features. The Content-Guided Attention (CGA) module, shown in Figure 5, assigns a unique SIM, W ∈ R C × H × W , to each channel. This coarse-to-fine fusion of channel and spatial information enhances feature representation. The spatial attention module captures the spatial information of the feature map through global average pooling and global max pooling operations. The channel attention module utilizes adaptive average pooling and convolutional layers to highlight the importance of different channels. Finally, by combining the input feature map and the outputs of the first two attention modules, the SIM in the feature space is finely adjusted through channel shuffling.

The Fusion network proposed in this study is composed of three CGAFusion modules, which respectively fuse the low-level features from the final three C2f_LarK layers of the backbone with the high-level features from the final three C2f layers of the neck. Given the backbone low-level features F l o w ( F l o w ∈ R C × H × W ) and the neck high-level features F h i g h ( F h i g h ∈ R C × H × W ) the module first constructs ground-state features via initial feature superposition X = F l o w + F h i g h ( X ∈ R C × H × W ) to serve as the input for the CGA module. Finally, the fusion results are fed into the Det-Tiny detection head. The specific steps are detailed below:

The spatial attention modeling of the CGA module captures local salient regions through dual-path feature compression. The calculation is expressed as shown in Equation 2.

Specifically, AvgPool and MaxPool compress the data along the channel dimension to produce 2-channel features. A 7 × 7 reflected padded convolution (Padding = 3) and a Sigmoid activation function are then used to generate a spatial weight map W S ∈ R C × H × W , highlighting the potential spatial distribution of foreign objects.

2. The channel attention modeling of the CGA module is based on the Squeeze-and-Excitation (SE) framework of SENet, which is utilized to model inter-channel dependencies. The calculation is expressed as shown in Equation 3.

Specifically, the spatial and channel attention maps are fused via element-wise addition. The resulting integrated attention map is then concatenated with the initial features. A 7 × 7 grouped convolution (groups = C) is subsequently applied to extract local correlations, generating refined spatial importance weights W ∈ R C × H × W . These weights are utilized to dynamically balance the contributions from the backbone and neck features.

4. Synergetic generation via the CGA-based Multi-scale Feature Fusion Network (CGAFusion). The calculation is expressed as shown in Equation 5.

This fusion approach achieves information complementarity, enabling the model to capture high-level semantic information while retaining underlying details. Furthermore, the CGAFusion module’s adaptive attention mechanism highlights key features and suppresses redundant information, effectively enhancing the model’s ability to detect objects of varying scales. This improves the model’s overall detection performance in complex scenarios.

Table 1 provides the main configurations for the experimental environment, including the operating system, CPU, GPU, Python, Cuda, and PyTorch version.

The hyperparameter settings for model training are shown in Table 2.

The input image size was uniformly set to 3 × 640 × 640. The number of training epochs was 200. The training batch size was 16. The initial learning rate was 0.01. And the cosine annealing decay algorithm was used to gradually change the current learning rate.

In view of the pronounced scarcity of publicly available datasets for power line foreign object detection, this study employs the RailFOD23 (Chen et al., 2024b) benchmark to evaluate the generalization capability of the proposed lightweight algorithm in complex scenarios. This dataset, which has been widely adopted in the literature for assessing the performance of power line detection models (Hao et al., 2024), comprises video frames sourced from multi-national railway surveillance systems. It spans a broad spectrum of illumination conditions, weather patterns, and object categories. Notably, its structural and environmental characteristics share a high degree of scene commonality and challenge with power line inspection tasks, providing a rigorous platform for validating model robustness.

The RailFOD23 dataset contains 14,615 images and 40,541 annotated objects. The annotated objects include four common foreign objects: bird nests, balloons, plastic bags, and floating objects. It is randomly divided into three group sets with a 7:2:1 ratio. The training set consists of 10,230 images. The validation set consists of 2,923 images. And the test set consists of 1,462 images.

This study uses recall (R), precision (P), and Mean Average Precision (mAP) to evaluate the accuracy of the model in detecting small objects. Calculations are shown from Equations 6–9. To comprehensively evaluate the computational efficiency of the model, other metrics introduced include the total number of parameters, model size, and number of floating-point operations (GFLOPs). R e c a l l = T P T P + F N × 100 % (6) P r e c i s i o n = T P T P + F P × 100 % (7) A P = ∫ 0 1 P R d R (8) m A P = ∑ i = 1 K A P i K × 100 % (9) Where TP represents the true positive that is correctly predicted. FP represents the false positive that is incorrectly predicted as positive. FN represents the false negative that is incorrectly predicted as negative. And K represents the number of classes.

C2f_LarK module, Det_Tiny module, and CGAFusion module are considered in ablation experiments.

To visually verify the effectiveness of proposed method and explain the underlying mechanism of performance improvement, we used Gradient-weighted Class Activation Mapping (Grad-CAM) technology to perform a feature map visualization comparison analysis of the models before and after the improvement, as shown in Figure 7. As can be seen from the figure, the original YOLOv8n model, limited by its small effective receptive field, often struggles to capture the complete features of the target, resulting in weak response or even feature loss in the heatmap at the target region.

The improved YOLOv8-FOD model exhibits stronger feature extraction capabilities, with its high-response regions in the heatmap more densely and completely covering the foreign object target. Notably, compared to the original model, the improved model demonstrates stronger focusing ability in the target edge region, enabling more accurate delineation of the object’s outline. This complete perception of the target’s shape is mainly due to the large receptive field of the C2f_LarK module, ensuring that the model can capture the target’s contextual information from a global perspective. Simultaneously, the introduction of the Fusion module effectively enhances the preservation of details during feature fusion, allowing the model to keenly capture object boundary information, thereby achieving accurate localization of the foreign object from its center to its edge.

To further verify overall effectiveness of the improved model proposed in this paper, we conducted a comprehensive comparative experiment with other models.