Timely and accurate detection of foreign objects is crucial for the safe operation of transmission lines in power grid. Currently, object detection models have more and more parameters and their calculations are becoming increasingly complex. Therefore, sufficient computing power is usually required. To deploy models on resource-constrained edge devices, this paper proposed a lightweight detection algorithm based on improved YOLOv8, which is called YOLOv8-FOD (Foreign Object Detection). Firstly, the backbone network is optimized by incorporating large kernel block (LarK Block) into c2f module, forming a new C2fLarK module. This achieves a wider receptive field, captures more contextual information, and effectively reduces network redundancy. Secondly, a lightweight detection head, DetTiny, is proposed. Through adaptive feature redundancy compression and hardware-friendly computational optimization, computational complexity is significantly reduced. Finally, a new feature fusion network structure (Fusion) is designed, utilizing the CGAFusion module to fuse high-dimensional and low-dimensional features. This captures important information at different semantic layers and improves the detection of edge details, effectively enhancing detection accuracy. Experimental results show that, compared with standard YOLOv8n, the proposed model reduces parameter count by 36. 6%, model size by 31. 1%, and computational complexity (GFLOPs) by 40. 7%. While maintaining detection accuracy of mAP@0. 5 and improving mAP@0. 5: 0. 95 by 0. 3%, the model is more lightweight and has high practicality for deployment on edge devices.
Liu et al. (Tue,) studied this question.