Orientation-Aware Feature Flow Network for Oriented Object Detection in Remote Sensing Images

Remote sensing oriented object detection faces the challenge of maintaining consistent rotational geometric information throughout feature representation, cross-scale propagation, and bounding-box localization. To address this issue, this paper proposes a direction–scale collaborative optimization method for oriented object detection. A frozen DINOv3 encoder is adopted as the semantic feature source to enhance target discrimination in complex backgrounds through global semantic modeling and dense feature representation. Rotation-Oriented Space-to-Depth Convolution (ROSPD) is introduced to preserve fine-grained spatial structures and local directional-response cues during downsampling. In addition, a Rotation-Aware Residual Fusion (RARF) module estimates local axial agreement between adjacent feature levels and uses it to condition residual multi-scale fusion. A ProbIoU–NWD joint loss with deterministic target-size-adaptive weighting is further constructed: it emphasizes NWD supervision for small objects and gradually returns to ProbIoU as target size increases, thereby improving localization stability under low-overlap conditions. Experiments on DIOR-R and DOTA-v1.5 demonstrate that the proposed method effectively improves the detection and localization performance of multi-scale oriented objects in complex remote sensing scenes.