Orientation-Aware Feature Fusion for Accurate Rotated Object Detection in Remote Sensing Images

Abstract

Conventional feature fusion mechanisms largely overlook orientation information, making it difficult to effectively represent objects with diverse rotational patterns. To address this issue, we propose YOLO-RSL, a lightweight rotated object detector that introduces orientation awareness into feature representation, feature fusion, and localization optimization. The proposed Rotation-Oriented Iterative Attentional Feature Fusion (RO-iAFF) module incorporates implicit orientation information derived from orthogonal asymmetric depthwise convolutions into a two-stage channel attention process, enabling orientation-sensitive multi-scale feature aggregation. In addition, a Swin Transformer block and a large-kernel feature branch are employed to enhance global context modeling and improve small-object representation. To achieve more accurate localization, an enhanced regression loss (LFFE Loss) is designed to jointly optimize overlap quality, rotation angle, center offset, and aspect ratio. Experimental results on the DIOR-R and UCAS-AOD datasets show that YOLO-RSL achieves mAP0.5 scores of 84.17% and 98.44%, surpassing the baseline by 3.15% and 2.91%, respectively. Moreover, the performance gains become more pronounced as target aspect ratios increase, demonstrating the effectiveness of the proposed orientation-aware design.