AnyRoad: A Frequency-Aware Adapter Framework for Road Segmentation With Segment Anything Model

Abstract

Automated road extraction from high-resolution satellite imagery is critical for geospatial applications. However, accurate segmentation requires balancing global topological continuity with local boundary precision. Existing methods often struggle with this tradeoff, while directly adapting large-scale foundation models introduces challenges with geometric discontinuities and computational cost. We propose AnyRoad, an asymmetric dual-encoder framework integrating a trainable SegFormer for domain semantics and a frozen SAM-2 for universal structural priors. To fuse these distinct representations, we introduce a frequency-domain collaborative fusion module (FCFM). Using the discrete wavelet transform (DWT), FCFM decouples features: low-frequency components are aligned via bidirectional cross-attention to preserve macrolevel connectivity, while high-frequency details are processed with a WaveMLP-based anisotropic operator to refine geometric boundaries. A Deformable UNet++ decoder is then employed to accommodate diverse road shapes. Experiments on the Massachusetts Roads and DeepGlobe datasets show that AnyRoad performs well. Cross-regional tests on the LSRV dataset also show stable transfer to unseen geographic areas.