Construction of a feature enhancement network for small object detection.

• To more effectively expand the possibility of small objects appearing, we improve current copy-paste based data augmentation method (CDCI) by introducing collision detection and spatial context position extension to avoid object collision and incorrect context information caused by random copy-paste. • To solve the problem that the small objects are vulnerable to scale variation, we construct a multi-granular deformable convolution network to learn and capture the changes in the shape and scale of the object, and offset feature representations in different granularity are acquire by granulating and fusing the offset features. • A high-resolution block (HR block) is designed to bring more semantic information while maintaining high-resolution features, and high-resolution block-based Feature Pyramid is built by parallel embedding HR block in FPN to further enhancing the feature representation. • A large number of experiments are reported to demonstrate the effectiveness of the proposed method. At the same time, we set up ablation experiments to analyze the rationality of proposed different strategies. Limited by the size, location, number of samples and other factors of the small object itself, the small object is usually insufficient, which degrades the performance of the small object detection algorithms. To address this issue, we construct a novel Feature Enhancement Network (FENet) to improve the performance of small object detection. Firstly, an improved data augmentation method based on collision detection and spatial context extension (CDCI) is proposed to effectively expand the possibility of small object detection. Then, based on the idea of Granular Computing, a multi-granular deformable convolution network is constructed to acquire the offset feature representation at the different granularity levels. Finally, we design a high-resolution block (HR block) and build High-Resolution Block-based Feature Pyramid by parallel embedding HR block in FPN (HR-FPN) to make full use different granularity and resolution features. By above strategies, FENet can acquire sufficient feature information of small objects. In this paper, we firstly applied the multi-granularity deformable convolution to feature extraction of small objects. Meanwhile, a new feature fusion module is constructed by optimizing feature pyramid to maintain the detailed features and enrich the semantic information of small objects. Experiments show that FENet achieves excellent performance compared with performance of other methods when applied to the publicly available COCO dataset, VisDrone dataset and TinyPerson dataset. The code is available at https://github.com/cowarder/FENet. [ABSTRACT FROM AUTHOR]