Segmentation-Detection Ensembled Network for Corrosion Defect Detection

Abstract

Metal parts in substations corrode over time. The corrosion defects in the substation should be inspected to reduce potential safety risks. Traditional detection methods for corrosion defects rely on manual features and have poor robustness and generalization ability. Because the semantic information of the target corrosion area is not significant, the defect detection methods based on deep learning cannot accurately distinguish the corrosion defect. This paper proposes segmentation-detection ensembled network (SDEN) to improve corrosion detection performance. SDEN integrates the semantic segmentation model and target detection model. The semantic segmentation model predicts pixel-level corroded areas. The target detection model predicts the target location. The two networks share backbone and feature pyramid networks and perform prediction independently. The target box containing the rusted area is identified as the rust defect target. By decoupling rust zones from target prediction, SDEN improves the detection performance of targets containing fewer corroded pixels. Experiments on real datasets validate the effectiveness of the propose SDEN.