GCaSNet: Global Context-Aware Selective Detection Network for Rail Surface Defect Detection

The accurate and fast detection of rail surface defects is of great significance for the safe operation of rail transport. Convolutional neural network (CNN) is the dominant choice for rail surface defect detection (RSDD) due to its superior feature learning capability. However, current RSDD still suffers from two challenges: high similarity of defects to the background and limited maintenance skylights. To address these challenges, we propose a novel RSDD network, namely global context-aware selective detection network (GCaSNet). Specifically, a rail surface image is first partitioned into non-overlapping patches and then projected into an embedding space for yielding tokens. Afterwards, they are fed into a global context-aware backbone for encoding short- and long-range feature relations. Eventually, a patch-wise detection head is tailored to localize and identify rail surface defects in selected suspicious patches for fast computation. We evaluate the proposed GCaSNet on Type-I and Type-II rail surface discrete defects data sets, and experimental results demonstrate that GCaSNet improves the mAP by at most 10% over existing convolutional networks and Transformer-based baselines while at low computational costs.