End‐to‐end power equipment detection and localization with RM transformer
Jian Fang, Youyuan Wang, Weigen Chen · IET Generation Transmission & Distribution · 2022
Abstract Power equipment detection and localization is the key component of automatic inspection tasks in substations. To solve the challenges such as complex environment and lack of training data, utilization of context information is considered here. For substation scenes, object relation modelling (RM) is proven to be useful, but an end‐to‐end efficient framework which is suitable for non‐Convolutional Neural Networks (CNN) is still missing. Therefore, an extended Transformer network with an elaborately designed RM module is proposed. As the foundation, transformer network is better than CNN in terms of context dependency construction. On top of that, an RM module is plugged to adjust the decoded feature embeddings based on their appearance, position and class information. The module is based on a graph attention neural network which uses similarity as weights of nodes. The experiments show that the proposed method has a 16.2% improvement in accuracy compared to pipeline, and even 6.4% higher than the most recent models, largely promoting the construction of intelligent substations.