Progressive fusion and deformable perception for remote sensing tiny object detection

Although anchor-free detectors have shown promise, they still struggle with insufficient multi-scale feature representation and limited geometric adaptability when detecting tiny objects with drastic scale variations and complex geometric structures in remote sensing images. To address these challenges, we propose a fusion-perception framework consisting of a Progressive Fusion-Perception Module (PFPM) and a Dynamic Deformable Perception Head (DDP-Head). PFPM performs progressive multi-scale feature aggregation across channel-split sub-feature groups, while DDP-Head integrates spatial attention, deformable convolution, and dynamic convolution for adaptive prediction. Experimental results demonstrate that the proposed method achieves a 2.7% AP improvement over the baseline on AI-TODv2 and also improves detection performance on LEVIR-Ship.