Rotated object detection is widely adopted in remote sensing to handle arbitrary object orientations and improve localization accuracy. However, existing methods still suffer from limited global context modeling, degraded feature representation under complex backgrounds, and suboptimal optimization caused by task coupling, which jointly restrict detection performance in challenging scenarios. To address these issues, this paper proposes a novel rotated object detection framework, termed LDA-YOLO, which systematically enhances feature modeling and prediction quality. Specifically, a Large Separable Kernel Attention (LSKA) module is introduced to approximate global spatial interactions through a low-rank separable formulation, enabling effective long-range dependency modeling with linear computational complexity. A Dual-Path Feature Refinement (DPFR) module is designed to improve feature representation by decomposing features into complementary subspaces and performing adaptive fusion to suppress redundancy and noise. In addition, an Angle-Aware Decoupled Head (AADH) is developed to explicitly separate classification, localization, and orientation estimation, thereby reducing inter-task interference and improving optimization stability. The proposed method achieves superior performance compared to existing approaches. Specifically, it improves mAP50 by 1.6% over the baseline YOLOv8n-OBB, while maintaining a lightweight design with significantly reduced computational cost. These results indicate that the proposed framework provides an effective solution for rotated object detection in complex remote sensing scenarios.
Shan et al. (Fri,) studied this question.