AACSNet: an angle-aware feature alignment and cross-level synergy framework for oriented object detection in remote sensing imagery

Abstract

Oriented object detection in remote sensing images remains challenging due to arbitrary object orientations, large-scale variations, and dense object arrangements. Feature orientation misalignment and insufficient adaptability of multi-scale feature representations to per-image scale distributions remain two fundamental bottlenecks, rarely addressed jointly in a unified framework. Angle-aware and cross-level synergy network (AACSNet) is proposed as an end-to-end architecture to address both simultaneously. The angle-aware feature alignment module (AAFAM) performs dense feature alignment at the region proposal stage through periodic angle encoding, rotation-aware sampling across multiple angle hypotheses, and joint angle–spatial attention, ensuring geometric consistency at the proposal stage rather than correcting misalignment afterward. Complementing this, the cross-level synergy module enhances multi-scale feature representations through context-conditioned scale importance estimation, neighborhood pyramid aggregation, and conservative residual integration, allowing the network to dynamically adapt to the scale distribution present in each input image. Extensive experiments demonstrate that AACSNet achieves 76.64% mAP50 on DOTA-v1.0, 67.80% mAP50 on DIOR-R, and 90.41% mAP(07)/96.67% mAP(12) on HRSC2016, with ablation studies confirming the contribution of each component.