APFNet: An Adaptive Perception Fusion Network for Multi-Scale Object Detection in Remote Sensing

Abstract

Accurate multi-scale object detection remains challenging in remote sensing (RS) imagery due to severe scale variation, limited pixel occupancy of small targets, and complex backgrounds. Existing multi-scale fusion methods propagate features along predefined cross-scale paths and apply receptive fields fixed once trained, so fusion weights and the spatial range of contextual evidence remain largely input-independent. The Adaptive Perception Fusion Network (APFNet) targets these two aspects through four operations: (1) Split-ELAN (SP-ELAN), a channel-compressed serial refinement backbone whose design is motivated by small-object feature preservation; (2) Adaptive Sampling (AS), which produces sample-dependent channel weights from a global descriptor during cross-scale alignment; (3) Adaptive Gating (AG), which is coupled with AS and aggregates the recalibrated features at complementary receptive-field scales; and (4) Adaptive Context Aggregation (ACA), which aligns three feature levels and applies parallel large-kernel depthwise branches to supply broader multi-scale context. On DIOR, APFNet-L achieves three-run mean values of 86.39% AP50, 63.09% AP, and 26.10% APS. A complete 23 component-group study under one scratch/no-augmentation protocol yields positive conditional contrasts for SP-ELAN, ACA, and the coupled AS–AG pathway. LEVIR and RSOD provide additional cross-dataset and configuration-selection results. These operations use established primitives; the contribution lies in their routing and aggregation.