An Attention-Based Detection Method of Displacement Field on Steel Surfaces

Abstract

Steel structures are prone to fatigue cracks when subjected to cyclic loading, which may lead to catastrophic failure. Generally, the width of fatigue cracks in steel structures is below 0.1 mm at the early stage of crack propagation. Although high-resolution images can be obtained by consumer-grade cameras at low cost, these tiny cracks are difficult to detect by images alone. This paper proposed a crack detection method based on the displacement field on the surface of the structure obtained from images. Video or continuous images of the target structure under loading was first taken and input into an improved Detector-Free Local Feature Matching with Transformers (LoFTR) model, which was capable of densely matching feature points on two pairs of images without distinct visual features. The surface displacement field of the structure was then performed inversely by the coordinate difference of a large number of matched feature points. Eventually, location of the crack was extracted according to discontinuities in the displacement field. A case study was conducted on a cracked steel plate. Results demonstrated a tiny crack with the maximum width of 0.1 mm was detected, which was more effective and accurate in comparison with image-based semantic segmentation methods.