Substation Defect Detection Network Based on the Attention Model

Abstract

Applications of intelligent robot iterative inspection system in substation are attracting extensive interests. The defect detection in inspection system must consider both real-time and accuracy on small database samples. This paper proposes a defect detection network based on the attention model. We combine a triplet attention module with Yolo-v5 network for producing richer representations of defect features, which can improve the accuracy in calculating fast situation. We also give a robust data enhancement mechanism for extending the database. The experiment dataset contains actual images gathering from substation and they are labelled as PASCAL VOC dataset format by safety inspectors with many years of experience. The experiment results can demonstrate that the proposed method can effectively improve the accuracy without reducing the real time performance.