Application and Effect Evaluation of Improved DETR Algorithm in Small Object Detection

Object detection, as the core task of computer vision, has important application value in automatic driving, medical image analysis and other fields. Small object detection is a difficult problem in current research because of the lack of pixel information and environmental interference. Although the traditional DETR (Detection Transformer) framework simplifies the detection process through end-to-end architecture, its global attention mechanism lacks local modeling ability, resulting in insufficient feature extraction of small targets, slow training convergence and difficulty in adapting to high-resolution inputs. In order to overcome these limitations, this paper proposes Windows-DETR model, which constrains feature computation in local windows by multi-scale window self-attention mechanism and window cross-attention mechanism, and effectively improves the capture efficiency of small target details. Dynamic query mechanism is introduced to adjust the number of queries adaptively according to target density, which significantly reduces the missed detection rate. At the same time, high resolution feature map and loss weighting strategy are combined to strengthen the sensitivity to small objects. Experiments on COCO2017 dataset show that the model achieves breakthrough improvement in small object detection accuracy while maintaining computational efficiency, providing a new solution for small object recognition in complex scenes.