A Looseness Recognition Method for Rail Fastener Based on Semantic Segmentation and Fringe Projection Profilometry

Recognizing loose condition of rail fastener is a challenging task. To this end, we propose a novel solution integrating deep learning and 3-D vision techniques. The segmentation network is first proposed to locate the spike and nut, and then, fringe projection profilometry (FPP) is used to recover the corresponding depth map. Finally, the looseness condition of rail fastener can be identified by the distance between spike and nut. Semantic segmentation and phase calculation of spike and nut are critical to the reconstruction accuracy of depth map. Aiming at segmentation problem, we propose a U-shaped network based on triplet attention (TAU-Net $4+\text {)}$ to improve segmentation accuracy. On the one hand, we redesign a skip connection layer to better enhance the interaction ability between deep and shallow features. On the other hand, a triple attention module is added to make the network more attentive to spike and nut during the learning process. For the phase calculation problem, we propose a phase correction and verification method based on dual-slope guidance and contextual information to improve the reconstruction accuracy of FPP. The experimental results first show that TAU-Net 4+ can excellently segment the boundaries of spike and nut, and the Dice coefficient reaches 99.27%, which lays the foundation for depth map acquisition. In addition, the experiments also verify that the proposed phase correction method can recover a more accurate depth map. Finally, the experiments validate that the proposed strategy can recognize the loosening status of rail fastener and the minimum looseness recognition distance is about 0.1 mm.