非线性图像扩散LB模型的CUDA算法设计与实现  

Design and Implementation of CUDA Algorithms Based on Nonlinear Image Diffusion LB model

作　　者：周明[1] 严壮志[1] 黄彬[1] 

机构地区：[1]上海大学通信与信息工程学院,上海200072

出　　处：《应用科学学报》2014年第1期85-92,共8页Journal of Applied Sciences

基　　金：国家自然科学基金(No.61171146);上海市科委科技创新行动计划基金(No.11DZ1921702)资助

摘　　要：为提高基于格子波尔兹曼(Lattice Boltzmann,LB)模型图像去噪方法的性能,研究了非线性图像扩散LB模型的CUDA算法,即分别利用纹理内存、共享内存以及直接使用全局内存来实现非线性图像扩散LB模型中的迁移过程.利用合成图像和真实图像的去噪实验表明,针对非线性图像扩散LB模型,GPU相对CPU的加速比可达90倍以上;而且加速比的提高与GPU流处理器的数目成正比.To improve the performance of Lattice Boltzmann （LB） method in image denoising, this paper proposes three compute unified device architecture （CUDA） algorithms to realize streaming processes of LB in nonlinear image diffusion, which are based on texture memory, shared memory, and global memory, respectively. To test effectiveness and efficiency of the GUDA algorithms, experiments were carried out with natural and composite images. The results show that GPU acceleration is 90 times faster than CPU acceleration, and the factor is proportional to the number of GPU stream processors.

关 键 词：图像去噪 非线性图像扩散 LB模型 CUDA算法 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]