Rotation-Equivariant Feature Learning and Dual-Head Decoupling for Oriented Remote Sensing Object Detection

Oriented remote sensing object detection (ORSOD) remains challenging because objects usually appear at arbitrary rotations, with large aspect-ratio variations and dense spatial layouts. Although recent detectors have achieved notable progress, most still rely on standard convolutional backbones, whose limited inherent rotation equivariance weakens robustness to rotated targets and reduces localization reliability. To address these issues, this paper proposes RedhDet, an ORSOD framework based on rotation-equivariant feature learning and dual-head decoupling. Specifically, a Ladder-Structured Implicitly Equivariant Network (LSIEN) is developed as the backbone to extract rotation-equivariant features in a soft manner, without relying on strictly constrained kernels or equivariant pre-trained weights. In addition, a decoupled dual-head architecture is introduced, in which a fully connected branch is used for classification and a convolutional branch is used for bounding box regression, thereby better matching the distinct requirements of semantic recognition and precise localization. Experimental results on the DOTA-1.0 and HRSC2016 datasets demonstrate the effectiveness of the proposed method. On the DOTA-1.0 test set, RedhDet improves mAP50 by 2.09% and mAP75 by 14.84% over the baseline. On HRSC2016, gains of 14.12% and 51.85% are achieved for mAP50 and mAP75, respectively. With multi-scale training and testing, the proposed method achieves 80.40% mAP on DOTA-1.0, indicating its effectiveness and competitiveness for ORSOD.