Small Object Detection Algorithm Based on Context Information and Attention Mechanism

As a special scene of object detection task, small object detection has been applied to some remote shooting scenes, especially in the field of unmanned aerial vehicle (UAV) object location and tracking. Because the small object (smaller than 32*32 pixels) only covers a small area, lacks diversity in location, and is usually accompanied by medium size (32*32 pixels to 96*96 pixels) objects, the universality of the small object detection task becomes very difficult. Aiming to deal with these difficulties, this paper proposes CGA-YOLO, a small object detection algorithm based on context information extraction and global attention mechanism. Based on YOLOv5, we use the window-based self-attention mechanism based on swin transformer to extract context information. Then, we also integrate the global attention mechanism to find the global information in the scene with dense objects. In the experimental results, CGA-YOLO improves by about 8% compared to YOLOv5 on the VisDrone2019 dataset, and the performance is improved significantly in the case of dense arrangement, which shows that the proposed strategy is effective for small object detection and recognition.