MobileCrackNet: lightweight crack segmentation with mobile encoder and channel attention

Abstract

Crack segmentation is critical for civil infrastructure inspection, yet state-of-the-art methods such as MixerCSeg rely on complex State Space Model (SSM) encoders with custom CUDA kernels that limit portability. We propose MobileCrackNet, a lightweight alternative that replaces the SSM encoder with a MobileNetV2 backbone enhanced by Squeeze-and-Excitation (SE) channel attention, while retaining an efficient multi-scale fusion decoder. With only 1.87M parameters and 1.85 GFLOPs (26.4% fewer than MixerCSeg), MobileCrackNet achieves 0.9100 mIoU on DeepCrack—within 0.56% of the state-of-the-art—and 0.8005 mIoU on CrackMap, while delivering 5×faster training throughput without custom CUDA dependencies. Ablation experiments reveal that SE attention contributes +5.51% mIoU, and that training from scratch consistently outperforms ImageNet pre-training across all evaluated datasets.