| With the development of SAR imaging technology and the increasing application requirements,the amount of SAR image data and resolution increase greatly,and the USE of CPU for SAR image registration can not meet the requirements of real-time processing.High performance computing based on GPU has high parallel efficiency,CPU-GPU heterogeneous computing platform and NVIDIA Unified Computing Device Architecture(CUDA)provide the possibility to use GPU to accelerate image processing.Research on using GPU to accelerate SAR image registration is of great significance to improve the real-time performance of SAR image registration.Focusing on SAR image registration acceleration,this paper studies GPU-based SAR image registration parallel optimization implementation,including CUDA programming model,thread synchronization and communication,parallel task division,access optimization,loop unrolling,instruction optimization and other technologies.Specific engineering practice results are as follows:(1)Based on CUDA platform,the third chapter introduces technologies such as two-dimensional convolution separation,multi-thread parallel computing,memory optimization,merge access,thread synchronization and atomic computing to realize the parallel acceleration of SAR-SIFT algorithm and improve the real-time performance of SAR-SIFT algorithm.In this method,two dimensional convolution is decomposed into two one-dimensional convolution to reduce the computational complexity of convolution.Then,CUDA memory model and merge alignment access optimization technology are combined to realize the parallel acceleration of SAR-Harris multi-scale space.Then,image boundary zero filling and parallel task division are used to reduce the branch judgment in kernel function and improve the efficiency of feature point detection.Then,in order to realize the parallel optimization of feature point main direction assignment and descriptor construction,thread synchronization and communication mechanism and atomic computing technology are adopted in this method.Finally,the method combines matrix splitting and access optimization techniques to realize the parallel calculation of matrix multiplication,and further speeds up the matching speed of feature point descriptors.Through the analysis and verification of experimental data,compared with the CPU version of the SAR-SIFT program,the speed of CUDA accelerated SAR-SIFT algorithm can reach more than 30 times.(2)In view of the slow speed of the improved SAR and visible image registration algorithm based on OS-SIFT framework,the fourth chapter of this paper introduces parallel task division,access optimization,CUFFT library,loop unrolling,instruction optimization and KD tree technologies to accelerate the parallel optimization of the algorithm and greatly improve the operation efficiency of the registration algorithm.Firstly,CUDA memory model and two-dimensional convolution separation were combined to improve the speed of gradient calculation.Then,multi-thread and thread synchronization mechanism were used to calculate the comparison factor in parallel,and additive operator splitting method was used for splitting and parallel computation to improve the speed of nonlinear diffusion multi-scale space construction.Then,in order to improve the speed of main direction allocation and construction of feature point descriptors,CUFFT library functions are used to accelerate the FFT,and techniques such as allocation thread structure,merge memory,instruction optimization and loop unrolling are introduced to realize the parallel optimization of frequency shift operation.Finally,the KD tree and CUDA parallel retrieval scheme are used to reduce the rough matching time of feature points,and the RANSAC algorithm program is optimized to improve the speed of feature points fine matching.Through experimental data analysis and verification,compared with CPU program,the running speed of CUDA accelerated registration algorithm is more than 30 times higher. |
