LAUSR

Oriented object detection in remote sensing images (RSIs) has received more and more attention due to its broader applicability in natural scenes relative to horizontal bounding boxes (HBBs). The complex scenes and multiscale targets in RSIs make it often difficult for existing studies to extract key features of the targets effectively. At the same time, due to the problem of feature inconsistency in different layers, the direct fusion of these features is likely to cause feature conflicts, resulting in degradation of detection accuracy. To solve these problems, the dual-path multihead feature enhancement detector (DP-MHFE Det), which contains two novel architectures, is proposed in this letter. The dual-path rotation feature aggregation module (DP-RFAM) improves the feature extraction capability of the network for rotating objects through dual-path structure and deformable convolution (DCN). To use these features effectively, the multihead multilevel feature fusion enhancement network (MMFFENet) is proposed to guide the feature layers to learn and retain the key features they need autonomously, and then enhance their features according to the characteristics of different subtasks. Experiments conducted on two remote sensing datasets, DOTA and HRSC2016, show that DP-MHFE Det is faster than almost all detection methods compared to the state-of-the-art (SOTA) methods while showing strong competitiveness in accuracy.