Lightweight Design Of SAR Vehicle Image Intelligent Recognition Networ

In recent years,SAR image intelligent recognition technology of convolutional neural network has been highly valued by scholars and scientists at home and abroad owing to its advantages of high accuracy,fast operation speed and so on.In terms of the application of SAR image recognition technology based on neural network to small radar platform,this thesis carries out in-depth research on the lightweight and hardware implementation of neural network.For convenience,this thesis selects the labeled vehicle SAR image as the research object,and adopts the popular lightweight network MobileNetV2 as the basic network for intelligent recognition of vehicle SAR image.The main work of the paper is as follows:First,the meta-learning method was applied to prune and compress the MobileNetV2 network.On the basis of MobileNetV2 network,utilizing the theory of meta-learning to add hidden vector,L1 regularization as well as an additional three-layer weight calculation network and gradually sparse the main network parameters through the process of training and retraining,so as to prune the network structure and compress the network volume.The experimental results on MSTAR data set indicate that the float point arithmetic of MobileNetV2 network after pruning is only 1% of the original one,and the correct rate of vehicle identification is only reduced from 97.35% to 96.57%,which means well results are obtained for the pruning and compression of the model.The Hessian matrix method is adopted to compress the bit width of the weight coefficient of each sub module of the pruned MobileNetV2 network.According to the minimum description length theory,the quantization bit width of each layer parameter is determined based on the main eigenvalue of Hessian matrix,which can reduce the network scale to 23.8% and keep the recognition rate above 96% on the premise of maintaining the recognition performance as much as possible.Then,a model lightweight scheme of pruning quantization integration is proposed,and a pruning algorithm network model combined with quantization characteristics is designed.According to the characteristics of the meta learning network,quantitative links are added in the pruning process and retraining process respectively.In this way,the network structure of the final output,through the meta-learning process,not only achieves the elimination of unimportant channels and the realization of pruning,but also decreases the bit width of parameter,reduces the network scale to 21% and maintains the recognition rate of 97.61%.MobileNetV2 network,after pruning quantization,is successfully transplanted to FPGA hardware platform.The problem of insufficient FPGA resources is effectively overcome after the use of DDR3 + FPGA architecture,successive designation of convolution layer,batch normalization layer,activation function layer and full connection layer modules,as well as the adoption of parallel prefix addition operation and booth multiplier.Then,the mobile netv2 network after lightweight processing is transplanted to Xilinx V7 chip to realize the intelligent recognition of vehicle SAR image.At last,the hardware module is tested and verified on the radar signal processor,and the network works normally.Compared with the GPU,the recognition speed is nearly doubled,and the recognition accuracy is almost consistent with the simulation results.