CFANet: coarse-to-fine adaptive network for oriented object detection in remote sensing images

Abstract Remote sensing object detection seeks to identify and localize targets in images captured by satellites and unmanned aerial vehicles (UAVs). Despite remarkable progress brought by deep learning, the complexity of remote sensing scenes still poses challenges such as dense object distributions, the loss of small targets, and performance degradation from scale variation and sample imbalance. To overcome these issues, this paper proposes a coarse-to-fine adaptive network for oriented object detection. In the feature extraction stage, a ConvNeXt backbone is utilized to capture long-range dependencies and enhance the representation of complex textures and diverse objects. For feature fusion, a Cross-layer Feature Pyramid Transformer is introduced to address the limitations of traditional feature pyramids in exploiting shallow information, integrating shallow details with deep semantics through cross-layer interactions to strengthen multi-scale representation. To improve the discriminability of candidate regions and the accuracy of localization, particularly for densely distributed and variably oriented objects, a Progressive Feature Refinement Head is designed with a stepwise refinement strategy. Benefiting from this coarse-to-fine design, CFANet demonstrates strong and robust performance across multiple benchmarks, achieving 77.16% mAP on DOTA-v1.0, 68.8% AP on RSAR, 90.21% mAP on HRSC2016, 38.25% mAP on FAIR1M-v1.0, and 30.8% mAP on STAR, while maintaining a favorable balance between detection accuracy and computational efficiency.