FINet: An Insulator Dataset and Detection Benchmark Based on Synthetic Fog and Improved YOLOv5
Zhengde Zhang, Bo Zhang, Zhi-Cai Lan, Hai-Chun Liu, Dongying Li, Ling Pei, Wenxian Yu
Abstract
The inspection of insulators and their defects is of great significance for ensuring the safety and stability of power system. Small sample is one of the main issues of insulator defect detection based on neural network. In this research, we release a dataset for insulators and self-explosive defects detection, and provide a benchmark based on improved YOLOv5, named Foggy Insulator Network (FINet). In this work, a synthetic fog algorithm is implemented and optimized. An insulator dataset (SFID) with 13000 images is constructed and released. The YOLOv5 network is improved into SE-YOLOv5 by introducing the channel attention mechanism, and a robust detection model with 96.2% F1 score for insulators and their defects is trained from scratch, and served as benchmark. The synthetic fog algorithm proposed in this paper can be widely used for data augmentation of various datasets. The trained model can be applied in the field of transmission line inspection. The source codes, datasets and tutorials are available on GitHub.