Tiny Defect Detection Algorithm for Power System Based on Drone Aerial Images

The secure, stable, and reliable operation of power transmission systems depends on the real-time perception and precise detection of defects in critical components. Traditional manual inspection methodologies are frequently characterized by low operational efficiency, high economic costs, and the presence of significant safety risks. In stark contrast, automated inspection technology based on unmanned aerial vehicles (UAVs), leveraging its superior advantages of high safety standards and low operational costs, has successfully emerged as the mainstream direction in the field of electric power operation and maintenance. However, in the context of practical engineering applications, the images acquired by drones from high-altitude aerial perspectives frequently encounter substantial challenges, such as the extremely diminutive scale of targets and highly complex background environments. When processing such challenging tasks, existing mainstream general-purpose target detection algorithms are severely limited by the continuous strided convolution downsampling mechanisms inherent in their backbone networks. This structural limitation inevitably leads to a situation where the fine-grained features of tiny defects such as insulator cracks and corrosion are excessively compressed or even completely lost within deep feature maps, thereby directly resulting in missed detections. To effectively address and mitigate this specific problem, this paper proposes an improved YOLOv8 detection model based on a collaborative enhancement strategy. This model constructs a structural coupling mechanism that integrates frontend feature fidelity with backend fine-grained parsing within the multi-scale feature fusion network. To be specific, at the frontend of the network architecture, the non-destructive downsampling module, known as SPD-Conv, is introduced to replace the traditional strided convolution layers. By mapping information from the spatial dimension non-destructively directly into the channel dimension, this module is able to meticulously preserve the fine-grained features of tiny defects while simultaneously reducing the resolution. At the backend of the network, a dedicated high-resolution detection head is designed and incorporated. This additional detection head utilizes the bidirectional fusion capabilities of the feature pyramid network and the path aggregation network to specifically parse and reconstruct the detailed information that has been preserved by the frontend. This design is explicitly intended to achieve the end-to-end, precise capture of tiny targets through the synergistic effect generated by the collaboration of both components. Extensive empirical experiments conducted on the VisDrone2019 benchmark dataset demonstrate that the proposed model exhibits superior comprehensive performance. Compared with the baseline YOLOv8 model, the improved model achieves substantial improvements of 3%, 1.57%, and 1.54% in precision, recall, and mean average precision(mAP), respectively, ultimately reaching an mAP of 38.75%. Furthermore, with only about 4.24 million parameters, its detection accuracy outperforms mainstream models such as TIB-Net, YOLOv5s, and YOLOv7, while the inference speed reaches 233 FPS. Featuring both low parameter count and high frame rate, it meets the practical needs for deployment on edge devices. In conclusion, this collaborative enhancement strategy can effectively boost the detection performance for tiny defects in power systems, possessing strong and robust potential for practical engineering applications.