Anchor-SAM: Active Mining of Latent Anchors from SAM Encoder for Road Extraction

Road extraction in high-resolution remote sensing imagery remains a persistent challenge due to occlusions and complex backgrounds, which lead to fragmented road topologies. However, existing road extraction methods constrained by limited pre-training remote sensing data often lack the generalization capability to distinguish roads from complex backgrounds. To address this issue, we propose Anchor-SAM, a novel framework that actively mines latent semantic anchors embedded in the SAM encoder to guide topological reconstruction. Our approach stems from a pivotal insight: the SAM encoder is able to abstract complex scenes into sparse semantic anchors at deep layers, thereby implicitly encoding the global structural skeleton. To harness these implicit cues, we introduce the Multi-scale Deformable Context Perceiver (MDCP) and the Deformable Bayesian Conditional Interaction Module (DBCIM). The MDCP explicitly utilizes spatial cues to aggregate global semantics across distributed anchors, establishing a robust initial context for the decoder. The DBCIM facilitates the diffusion of semantic cues to surrounding regions and effectively suppresses noise. Specifically, by leveraging the semantic certainty of anchors to guide deformable sampling trajectories, this mechanism proactively filters out background regions while precisely repairing fragmented road topologies. Our method achieves competitive performance on both the DeepGlobe and Massachusetts datasets. The source code will be publicly available at https://github.com/Hmbb0606/Anchor-SAM.