SFPNet: Self-Learning Small Object Detection for Large-Scale Remote Sensing Images

Compared with generic object detection, remote sensing image small object detection typically encounters problems such as information loss, noise characteristic interference, and disturbance of the bounding box. Different from the con ventional method of bounding box regression, we propose a self-learning feature point network (SFPNet) for small object detection, which can captures the geometric information of the object in arbitrary directions as well as enhance the representation of small objects by leveraging feature points. Specifically, a feature points orientation transformation module is proposed to accurately obtain classification and localization information. To improve the learning ability of weak sample objects, a positive and negative sample evaluation and balancing strategy is proposed. Additionally, we introduce a self-learning method of feature points under spatial constraints, which enhances the self-refinement and object expression ability of the feature points while eliminating false alarms. To enhances the training of hard-to-distinguish samples, we propose online feature point hard example mining. Experimental results on three benchmark datasets (including AI-TOD, SODA-A, and DOTA v2) show that the proposed method achieves more than 1.3% improvement in small object detection accuracy at different scales.