GPU Parallel Implementation And Optimization Of SAR Target Recognition Method

SAR Target Recognition algorithm,whose achievements have been widely used in military and civilian domains,has become a hot topic of research in recent years.With the rapidly increasement of the data dimension and the amount of data of SAR images,due to the development of High-resolution SAR Imaging technology,the traditional CPUbased target recognition algorithm can not meet the requirements of real-time processing of high-resolution SAR target recognition software,and the calculation cost is too high.In recent years,GPU common computing is able to provide powerful computing power and storage bandwidth with low development costs,short cycle and other advantages.Therefore,the research of GPU-based parallel target recognition algorithm is of great significance to the establishment and improvement of target recognition software system processing data in real time.The thesis will discuss the system architecture firstly.Then the CUDA programming model is introduced in detail.The target recognition algorithm is divided into the feature extraction part and the classifier part,and then describing that how the computing tasks will be parallel decomposed.The thesis also introduces how to implement various computational tasks through CUDA parallel programming.Finally,a series of optimization measures are used on CUDA programs to maximize the acceleration of the algorithm.The main works is summarized as follows:(1)Analyzing the CUDA programming model,storage model,execution model and programming language.Then principle and the realization methods are studied,including analyzing the Principal Component Analysis,Linear Discriminant Analysis and Nonnegative Matrix Factorization feature extraction technologies,and the Support Vector Machines classifier.All of these provide the theoretical basis and technical basis for the parallel analysis.(2)Researching feature extraction and classifier computing tasks and realizing the parallel improvement of calculation process.Then parallel decompose and realize Jacobi iterative algorithm,matrix multiplication,reduction,and the construction of interclass and intraclass divergence matrix.Introduce the calculation process and parallelism of SMO algorithm to realize the parallel transplantation of SVM on CUDA.Based on the MSTAR public database,the experiments can get the results that the target recognition algorithm is running on the CPU and GPU.Comparing and analying the experimental results to verify the acceleration effect of GPU parallel computing on target recognition algorithm.(3)Based on the general evaluation method and optimization strategy of CUDA program,the causes of CUDA program running speed in target recognition algorithm are deeply analyzed,and the algorithm is optimized from three aspects: communication,access and instruction flow.And through experiments show that GPU-based parallel implementation of the target recognition algorithm has been optimized to obtain about 20-30 times performance upgrade.