Scale-Aware Gated Routing Distillation for Lightweight Remote Sensing Object Counting

Abstract

Remote sensing object counting remains challenging for lightweight models due to large scale variation, dense object distribution, and complex background interference. Existing lightweight counting methods usually rely on static multi-scale fusion, which limits adaptive receptive-field selection under heterogeneous scenes. To address this issue, this paper proposes Scale-Aware Gated Routing Distillation (SAGRD), a lightweight framework that reformulates multi-scale feature fusion as an adaptive routing problem. At its core, a scale-aware gated routing pyramid generates dynamic branch weights from three complementary cues—global scale descriptors, branch-response feedback, and spatial routing signals—to adaptively reweight dilated receptive-field responses before weighted aggregation, enabling more flexible scale-aware representation under limited model capacity. In addition, a routing-oriented distillation strategy is introduced to transfer not only density estimation knowledge but also the teacher’s scale-selection distributions to the lightweight student, allowing the compact model to inherit structured receptive-field adaptation strategies. Extensive experiments on multiple remote sensing object counting datasets demonstrate that the proposed framework achieves a favorable accuracy-efficiency trade-off. On the RSOC Building benchmark, the student model achieves an MAE/RMSE of 4.712/6.367 using only 1.395 M parameters and 1.29 G FLOPs, closely approaching the teacher performance.