The Image Crowd Counting Of Light-Weight Network Based On Knowledge Distillation

Since the development of crowd counting research,a variety of excellent statistical models have been proposed for this research.From the beginning of a light-weight network with a shallow network layer and few training parameters to the current network layer with a deeper layer and relatively high training parameters with many heavy-weight networks,different statistical models have different advantages,and their disadvantages cannot be ignored.The light-weight network with shallow network layers and few training parameters is easy to train,and the crowd counting effect is better in the sparse crowd scene,but when the scene crowd is dense,the counting effect is worse than the network layer is deeper and the training parameters are heavy-weight.The heavy-weight network needs more resources during training because of its more parameters and bulkiness.When the crowd scene is sparse,the light-weight network performance is good enough.At this time,the heavy-weight network model is used to count the sparse crowd scenes.It seems too much.Therefore,it is meaningful to think about how to transfer the counting effect of the heavy-weight network model to the light-weight network,so that the light-weight network can maintain its counting effect in the sparse crowd scene,and at the same time improve its counting effect in the dense crowd scene extraordinary,so that while rationally using resources,it can maximize the role of light-weight networks.In response to this problem,this paper proposes a method of knowledge distillation to transfer heavy-weight network knowledge to light-weight networks to improve the statistical performance of light-weight networks.In this study,the heavy-weight learning network is called the teacher network,and the light-weight learning network is called the student network.This article fully investigates crowd counting research,selects several representative models as students and teachers,and investigates the application of knowledge distillation in crowd statistics research,and selects reasonable knowledge distillation methods for the research.This paper first distills the output density map of the teacher network,then distills the feature map of it,and transfers its knowledge to the student network.The experimental results show that the research of knowledge distillation method in this paper on image crowd counting can improve the statistical effect of light-weight networks in dense crowd scenarios.The experiment in this paper is trained and tested on several representative public datasets and densely populated datasets,ShanghaiTech Part?A dataset and UCF?CC?50 dataset.Among them,on ShanghaiTech Part?A,the average value of student network population increased by 9 %,on the UCF?CC?50 data set,the student network demographics have increased by an average of 21%.In this paper,through knowledge transfer,the counting effect of light-weight networks in dense crowd scenarios is improved,and resources can be used reasonably,which has good research and application value.