Context-awareness and frequency-refinement network for small object detection in aerial images

Due to the high flight altitude and large reconnaissance area of Unmanned Aerial Vehicle, objects in aerial images usually have limited feature information and low resolution, which results in them having a few pixels. In this work, we propose a context-awareness and frequency-refinement network for small object detection in aerial images, and make the following contributions to overcoming the above challenges. First, a dual-branch feature extraction network is constructed, which extracts the local and global feature of small objects to enhance the feature representation of small objects. Then, the boundary context-awareness module is designed, which fuses boundary context information with original images to enhance the local feature of small objects, aiming to improve the feature representation of small objects. Finally, the self-attention frequency-refinement module is studied, which adopts the adaptive technology to filter out redundant interference information of different frequencies, aiming to refine the global feature of small objects and reduce the interference from complex backgrounds. Extensive experiments on aerial image datasets demonstrate the superior performance of our network in both quantitative and qualitative evaluation. It is worth noting that our network reaches 95.37% mAP on NWPU VHR-10 dataset and 68.27% mAP on DOTA-v1.5 dataset, which has significant advantages in small object detection compared to currently popular methods.