Abstract: The research of object detection in natural scenes and remote sensing scenes is extremely challenging. Although many advanced algorithms have achieved excellent results in natural scenes, the complexity of remote sensing images, the diversity of object scales, and the dense distribution of object make the research on remote sensing image object detection slow. This paper proposes a novel multi-category object detection model which can automatically learn the weights of feature fusion and highlight object features at the same time. As a result, the model achieves effective detection of small objects and densely distributed objects in complex remote sensing images. The experimental results of the model on public datasets DOTA and NWPU VHR-10 show that the detection effect exceeds that of most classical algorithms.

Key words: multi-stage, feature fusion, semantic enhancement, object detection, remote sensing image