Transformer‐based crack segmentation for concrete structures in complex scenarios

Abstract Efficient segmentation of concrete cracks in complex scenarios is crucial for ensuring infrastructure safety and maintenance. This study proposes an efficient Transformer‐based crack segmentation network (ETCS‐Net) that effectively balances accuracy and computational efficiency. Unlike conventional transformers, ETCS‐Net introduces a unified feed forward network to integrate local information into self‐attention mechanisms, enhancing robustness without explicit positional embedding. Furthermore, an improved multi‐head self‐attention mechanism with local feature injection and cross‐head communication refines crack representation. To optimize efficiency, an attention‐based down‐sampling strategy significantly reduces computational overhead while preserving critical crack details. For enhanced segmentation performance in challenging environments, an online hard example mining strategy enables the model to prioritize hard‐to‐detect crack regions. Experimental results demonstrate ETCS‐Net's superior efficiency, maintaining competitive accuracy while significantly improving inference speed, with an Intersection over Union of 60.28% and an inference speed of 92.14 frames per second (FPS). These findings highlight ETCS‐Net's potential for real‐time and scalable crack detection in complex engineering scenarios.