Improvement of Small Object Detection Effectiveness Based on Swin Transformer

This paper proposes a novel approach to improve performance and effectiveness of the detection transformer (DETR) model for small object detection. The proposed approach utilizes the Swin transformer as the backbone for the DETR model and applies a local attention mechanism. The Swin transformer possesses better multi-level representation capability than the conventional CNN backbones, which helps improve training efficiency and object detection capability. The Focal loss function supports the model converge quickly when dealing with imbalanced training datasets. Experimental results demonstrate that the DETR model using Swin transformer as the backbone and Focal Loss function achieves higher training performance and better small object detection capability than the DETR model using conventional CNN backbones. The model almost converged at Epoch 50th, much lower than the convergence rate of the original DETR model.