A Lightweight U-Net for Road Crack Segmentation with Multi-Scale Feature Fusion and Gated Bottleneck Convolution

Road cracks are among the earliest and most common pavement distresses, and their accurate extraction is essential for timely maintenance and early warning of pavement deterioration. However, pixel-level segmentation remains challenging because cracks are typically thin, discontinuous, and multi-scale, while pavement textures, shadows, and stains can cause false detections. This study proposes a lightweight improved U-Net for fine-grained road-crack segmentation. A lightweight multi-scale fusion decoder (LMFD) integrates hierarchical features and strengthens cross-scale interactions to represent small and fragmented cracks. Gated bottleneck convolution (GBC) is further embedded in the skip connections to suppress redundant background information and enhance crack-related responses. Experiments were conducted on Crack500, a public pavement-crack dataset containing 500 high-resolution images (approximately 2000 x 1500 pixels) with pixel-level annotations. On the test set, the proposed method achieved an intersection over union (IoU) of 56.58% and an F1-score of 71.72%, outperforming U-Net by 2.07 and 1.86 percentage points, respectively. Compared with FCN, SegNet, PSPNet, Swin-Unet, and TransUNet, it improved IoU and F1-score by 2.61 and 2.30 percentage points on average, respectively. The model contains 5.445 M parameters and requires 100.424 GFLOPs, representing reductions of 68.47% and 37.35% relative to U-Net. These results indicate that the proposed network improves crack sensitivity and background robustness while maintaining a lightweight structure, providing an efficient approach for automated road crack extraction.