FVS-YOLO: A Lightweight Insulator Material Detection Network with Enhanced Feature Representation for UAV-based Power Line Inspection

Abstract

To address the high computational complexity of existing deep learning models in transmission line insulator material detection and their difficulty in deployment on mobile platforms, this paper proposes FVS-YOLO, a lightweight model. Building upon the YOLOv8s architecture, this model implements three key optimizations: First, an improved FasterNet module is introduced into the backbone network, where embedded convolutional layers with stride=1 enhance feature extraction capability while significantly reducing floating point operations (FLOPs). Second, a novel VoVGSCSPC neck module is designed, integrating group convolution and spatial convolution strategies to minimize computational redundancy during feature fusion. Third, the parameter-free SimAM attention mechanism is incorporated to amplify critical features and suppress irrelevant information. Experiments were conducted on a self-built dataset comprising 3,089 images across three insulator material types (composite, glass, and ceramic). Results demonstrate that FVS-YOLO achieves a mAP@50 of 0.916 while reducing parameters to 5.0M (54.95% decrease), computational cost to 15.1 GFLOPs (46.83% reduction), and model size to 9.8MB (54.21% compression). Moreover, its inference speed reaches 229.9 FPS, representing an 18% improvement. This model achieves a favorable balance between accuracy and lightweight design for insulator material detection, showing potential for deployment on drone-edge computing devices and providing a promising solution for real-time smart grid inspection.