A Novel Small Object Detection Method Based on Improved Transformer Model

Detection Transformer (DETR) transforms queries into unique objects and supports end-to-end object detection by using bipartite matching for training. This end-to-end design has greatly improved the generality of DETR and has been widely used in various downstream tasks. However, small objects are frequently missed or falsely detected due to poor features and high noise. To reduce these problems, we improve the encoder and decoder and introduce feature-enhanced auxiliary head. We advance the spatial correlation between adjacent tokens using the Locally-enhancement FeedForward (LeFF) module to enhance the extraction of local features and establish remote dependencies. Additionally, in order to increase positive queries, we change one-to-one matching for IoU assignment. We have addressed the issue of small object false and missed detection, and achieved better detection results. On the MS COCO datasets, our proposed method has 1.8% improved to the conventional methods and achieves 67.4% mAP@0.5 better than the powerful DINO.