Dynamic distillation based multi-scale lightweight target detection

For target detection tasks of autonomous driving, the scales of targets vary greatly, which brings huge challenges to current target detection models. In this paper, a lightweight multi-scale target detection model based on YOLOV7tiny is proposed. Firstly, WASPS, MFUP, and ELAN_DCN modules are designed to improve the feature extraction capability of the original YOLOV7tiny model. Secondly, to reduce the redundant information in the model, and to decrease the number of model parameters and the amount of computation, a Lamp pruning algorithm is used to perform pruning operation upon the model. More importantly, a dynamic knowledge distillation algorithm, which dynamically scales the cross-loss between a teacher model and a student model, is designed to improve the detection effect. The experimental results show that the proposed model has huge advantages in multi-scale target detection over prevalent models. Compared with the original YOLOV7tiny, the proposed model not only reduces the number of parameters and the computation amount, but also enhances map50 by 6.6% and map50:95 by 9.8%. The experiments on autonomous driving multi-scale datasets (such as KITTI and Citysacapes), VOC dataset and COCO dataset, further validate the generalization of the proposed model.