LS-UNet: A Lightweight Real-time Segmentation Network

In response to the challenges of low efficiency and extended computation time in road crack detection, we propose a lightweight real-time segmentation model, LS-UNet, built upon improvements to the UNet architecture. This approach utilizes deep separable convolutions to reduce feature redundancy. Furthermore, it incorporates deep skip connections with SE attention mechanisms to extract cross-level encoder and decoder features. Additionally, a hybrid loss function combining the generalized Dice and L1 loss is employed to enhance network training performance. Experimental results on the publicly available DeepCrack and CFD datasets demonstrate that the enhanced method, compared to UNet, achieves improvements in Mean Intersection over Union (MIOU) by 2.99% and 1.84%, with a simultaneous 7% reduction in parameter count and a computational load accounting for only 13%. The frames per second (FPS) reached 33, signifying an enhancement in model detection accuracy and a reduction in both model parameter count and computational load, thereby ensuring real-time requirements are met.