LAUSR

The fault diagnosis of electrical equipment plays a vital role in the safe operation of the power system. The task of electrical thermal image semantic segmentation is to segment all electrical equipment in thermal images, which is a key step in the automatic fault diagnosis of electrical equipment. However, there lacks of a large-scale dataset in this research field. Therefore, we contribute to a large-scale dataset for electrical thermal image semantic segmentation. It contains 4839 thermal images and 17 types of electrical equipment. We provide pixel-level annotations to facilitate the performance evaluation of different semantic segmentation algorithms. To provide a strong baseline, we propose a cross-guidance network (CGNet), which jointly infers semantic segmentation maps and edge extraction results in an end-to-end learning framework, for electrical thermal image semantic segmentation. Extensive experiments on our launched dataset demonstrate the effectiveness of the proposed CGNet, and it also achieves the best performance on the general thermal image segmentation dataset. We will release our code and dataset at https://github.com/guo49/CGNet-pytorch.