Weather-Robust Spatial-Frequency Decoupling Transformer for Crack Segmentation

Using edge devices such as drones for road inspections is critical for monitoring the structural integrity of transportation infrastructure. However, the limited resources of these devices constrain the deployment of high-performance models, while dynamic weather conditions require greater model stability. To address the challenge of efficient crack detection across various environments, we propose the spatial-frequency decoupling transformer (SFDFormer). Our model integrates an all-inone restoration module and a frequency decoupling module. The restoration module, combining AdverseGAN and stacked attention-integrated iterative recovery (AIIR) modules, restores images degraded by adverse weather. The frequency decoupling module employs a lightweight parallel network architecture to establish a correspondence between frequency-domain components and spatial-domain pixels, enabling direct extraction of crack edges and region information from the raw image. This approach effectively leverages semantic details from the frequency domain for pixel-level crack segmentation. Extensive experiments across six benchmark datasets demonstrate that SFDFormer consistently outperforms existing models, setting a new standard for crack detection on edge devices under harsh weather conditions.