Automatic detection of cracks during power plant inspection

Abstract

Robust inspection is important to ensure the safety of nuclear power plant components. Manually inspecting 100+ hours of video for rarely occurring cracks is a tedious process. However, automatic inspection is challenging as the images often contain highly textured area including weld and concrete surface which causes fragmented and noisy segmentations. Moreover, lack of crack samples cause challenges in training classification methods. In this paper, we propose to improve the detection of cracks by (1) reducing the fragmentation of segmentation by iteratively linking of possibly broken short lines that we call “linelets,” (2) minimize the false positive rate by filtering out area with weld, and (3) using anomaly measure to improve the classification. Testing of 42 real images demonstrates 38% improvement over prior method.