Abstract: To overcome these limitations,such as the computational bottlenecks and the challenge of balancing real-time perfor-mance with accuracy in the DETR(Detection Transformer) family of object detection methods during inference,this paper proposes an enhanced approach that combines a phased training strategy with multi-scale feature fusion.Specifically,the multi-layer encoder structure of DETR is simplified to reduce computational complexity,while the phased training strategy improves feature representation and accelerates model convergence.In the first phase,one-to-many label matching is adopted to obtain high-quality two-dimensional multi-scale features.In the second phase,the weights from the first phase are frozen,and a parallel attention-convolutional fusion module is introduced to further refine the features.Experimental results demonstrate that the proposed method achieves a 5× increase in inference speed and a 1.5-point AP gain over the baseline model on the COCO dataset,effectively alleviating DETR's inference inefficiency.In addition,it yields a 1.4-point AP improvement on the BitVehicle dataset.

Key words: Object detection, Parallel attention convolution, Phased training strategy, Multi-scale feature fusion