Abstract: In order to address the challenges posed by complex backgrounds, small object sizes, and arbitrary directions in remote sensing images, this paper presents a novel object detection algorithm. The proposed algorithm consists of several key components.Firstly, a context augmentation feature pyramid network（CAFPN） is introduced. In the feature extraction stage, it integrates adaptively with the dilated convolution to obtain features with rich semantic information and reduce information loss of small objects. Then,midpoint-offset regression（MOR） is employed to detect oriented box in the regression network to reduce the computational complexity caused by redundant anchors. Finally, a double-head network decouples classification and regression features, and incorporates a feature refinement module guided by attention mechanism and polarization functions, enabling the construction of task-specific features that facilitate accurate object detection.Experimental results on public remote sensing datasets, including DOTA, HRSC2016, and UCAS＿AOD, demonstrate the effectiveness of the proposed algorithm. Compared to the Faster RCNN algorithm, the proposed method achieves accuracy improvements of 8.48%,7.60%, and 3.10% on the three datasets, respectively. The proposed method enables high-performance object detection in remote sensing images.

Key words: remote sensing image, context augmentation, attention mechanism, polarization function, feature refinement