Research On Optimizing Of SiamRPN++ Video Single Object Tracking Based On Knowledge Distillation

SiamRPN++ is a video single object tracker model of deep learning constructed by siamese neural network and region proposal network,which shows strong tracking performance in the field of single object tracking.However,SiamRPN++ uses the Res Net as the backbone for feature extraction and uses the layer-wise strategy makes the three regions proposal network to calculate the correlation degree of features,resulting in huge demand for hardware resources and strict hardware requirements for running equipment.To solve this problem,the model optimization strategy of SiamRPN++ model is studied and an optimization algorithm of SiamRPN++ vedio single object tracking based on knowledge distillation is proposed after the optimization strategy.The optimization algorithm can greatly reduce the hardware resource requirements of SiamRPN++ without a large single target tracking performance loss.The main research work is as follows:(1)Analysis of model structure and optimization strategy of SiamRPN++.Aiming at the problem that it is difficult to guarantee the compression efficiency and model performance of SiamRPN++ model after optimization,the corresponding optimization strategy to provide theoretical guarantee for the compression efficiency and performance loss of model optimization algorithm is studied.Firstly,the model structure of SiamRPN++ is analyzed,and the dependence of model performance on the structure and the distribution of hardware resource demand of the structure are analyzed;On this basis,the optimization strategy is further analyzed,including the model optimization objects and model optimization method of SiamRPN++.(2)Research on backbone optimization algorithm of SiamRPN++.Aiming at the problem of constructing backbone with low hardware resource requirements by using the cheap network architecture design,an algorithm of constructing backbone by new convolution architecture design Ghost module is proposed.Firstly,Ghost module is used for transfer learning to build a backbone with low hardware resource requirements;Then,aiming at the problems of mismatch of crossing domains and too large receptive field in the backbone after transfer learning,the solutions of domain adaptive convolution layers and receptive field adjustment strategy are proposed;At the same time,efficient model training strategies according to the characteristics of the optimized model are also proposed.(3)Rearch on performance optimization of knowledge distillation for single target tracking.Aiming at the problem of small performance loss of the optimized SiamRPN++ model,a knowledge distillation algorithm based on gap calculation of outputs of models and fused model is proposed to optimize performance.Firstly,aiming at the poorer performance of the optimized SiamRPN++ model,a knowledge distillation algorithm based on gap calculation of outputs of models is proposed.The algorithm will transfer the knowledge contained in the output of the original SiamRPN++ model to the optimized SiamRPN++ model and help the optimized model train more effective.At the same time,aiming at the problem of many differences between the modules of the original model and the optimized model,a knowledge distillation algorithm based on fused model is proposed to realize the knowledge distillation at module level,transfer the knowledge of the backbone module and the regional proposal network module of the original model to the corresponding modules of the optimized model,and force the optimized model to fit to the original model at the module level,so as to further improve the efficiency of knowledge distillation and performance optimization.