GLANet: Global–Local Adaptive Network for Efficient Rotated Object Detection

Rotated object detection plays a crucial role in various visual perception tasks such as aerial photography, remote sensing imagery, and low-altitude unmanned aerial vehicle (UAV) imagery. However, targets in both high-altitude remote sensing images and low-altitude UAV images often exhibit significant scale variations, diverse orientations, and dense spatial distributions, posing formidable challenges to detection algorithms in terms of accuracy and real-time performance. To address these issues, this paper proposes a Global-Local Adaptive Network for Efficient Rotated Object Detection (GLANet), designed to enhance detection precision and efficiency in complex scenarios. GLANet incorporates a lightweight backbone network, Revisiting Mobile CNN From ViT Perspective (RepViT), which balances inference efficiency with an improved capability to represent directional structural features of objects. During feature fusion, we introduce the Geometry-Enhanced Attention guided Rotated Feature Pyramid Network (GEAR-FPN), which jointly models global semantic context and local detailed features, thereby strengthening detection performance for small-scale and densely packed targets. In the detection head, we present a Dynamic Lightweight Geometric-Aware Head (DLGA-Head) alongside a Dynamic Lightweight Global Attention (Dynamic LWGA) mechanism to strengthen the representation of target orientation and boundary information. The effectiveness of the proposed method is validated on both the DOTA and CODrone datasets. GLANet achieves an mAP of 78.12% on DOTA with competitive, near-state-of-the-art accuracy and significantly higher computational efficiency than other top-performing models. Specifically, it contains only 8.64M parameters and 35.69 GFLOPs, ensuring real-time inference while maintaining high precision. On the CODrone dataset, it further delivers improved detection performance while maintaining superior efficiency compared with existing approaches.