A Novel Method for Thermal Image Based Electrical-Equipment Detection

An accurate and reliable thermal image based electrical-equipment detection is critical in smart power grids such as automatic defect diagnosis. However, few works have provided solutions to the task. To solve the problem, in this paper, we propose a new task named thermal image based electrical-equipment detection which includes two contributions. First, we have created a large-scale thermal electrical-equipment benchmark from 5558 thermal images which were taken during electrical-equipment inspection in reality. Second, we used the self-attention mechanism to get better detection performance. We have made some improvements based on Dual Attention Network (DANet) and applied it to further improve feature representation, we named our method Channel-Position Dual Attention Network (CPDANet). The experiment results show that our novel method can improve the mean Average Precision (mAP) from Faster R-CNN’s 89.9% to 91.4%.