DD‐YOLO: An object detection method combining knowledge distillation and Differentiable Architecture Search

Abstract Although YOLOv4 has a high detection accuracy, its backbone network CSPDarknet53 requires a large number of parameters, which reduces the inference speed of the detection network to some extent. To address this drawback, a high‐efficiency detection network search strategy is proposed. First, the concept of Differentiable Architecture Search (DARTS) is used to search networks with a small number of parameters on the COCO2017 datasets; then, the backbone of YOLOv4 is redesigned by stacking cells. This strategy can reduce the number of parameters of the network. Second, the authors use ResNet101 as the teacher network and DD‐YOLO searched by DARTS as the student network to refine the knowledge model and improve the detection accuracy of the model. The proposed model is evaluated by conducting experiments on the COCO2017 test‐dev datasets. The results show that DD‐YOLO achieves 43.5% mean average precision and 2.3% model accuracy improvement. Meanwhile, the model complexity can be reduced to 61.4% of the original. Moreover, compared with YOLOv4, DD‐YOLO is more suitable for mobile deployment.