双通道交替最小化算法的GPU并行加速
GPU parallel acceleration of two-channel alternating minimization algorithm
韩雪 1刘金龙 1李松恒 1杨慧珍 2张之光 1李紫薇1
作者信息
- 1. 江苏海洋大学电子工程学院,江苏连云港 222005
- 2. 金陵科技学院网络与通信工程学院,南京 211169
- 折叠
摘要
大气湍流和系统噪声的存在使得天文或空间目标成像模糊.而双通道交替最小化算法是复原湍流和噪声降质图像的有效方法之一.但该算法比较复杂,需要反复迭代运算,处理耗时较长.为了提高算法运行速率,结合算法结构特征,将图形处理器(GPU)加速技术应用于双通道交替最小化算法,重点优化交替最小化迭代过程.实验结果表明:在不同湍流且信噪比(SNR)20 dB的条件下,与直接采用中央处理器(CPU)的算法相比,GPU并行加速用于双通道交替最小化算法,能够实现图像复原的"U-step"运算速率提升80％以上,点扩散函数求解的"H-step"运算速率提升60％以上,且恢复后的图像效果接近衍射极限.并行加速技术与已有的算法相结合的方式能够有效提高运行速率,为湍流和噪声降质图像的复原提供一定的参考.
Abstract
Due to atmospheric turbulence and system noise,the images of astronomical or space objects are blurred and degraded.The dual channel alternating minimization algorithm is one of the effective methods for restoring images degraded by turbulent and noise.However,this algorithm is relatively complex and requires repeated iterative operations,resulting in a longer processing time.In order to improve the algorithm running speed,the graphics proces-sor(GPU)acceleration technology based on the algorithm structure features is applied to the dual channel alternating minimization algorithm,with a focus on optimizing the iterative process of alternating minimization.The experimental results show that under the condition of different atmospheric turbulence and Signal-to-noise ratio 20 dB,compared with the algorithm directly using the central processing unit(CPU),GPU parallel acceleration for the dual channel al-ternating minimization algorithm can achieve the"U-step"operation rate of image restoration increased by more than 80％,and the"H-step"operation rate of point spread function solution increased by more than 60％,and the recon-structed images are close to the diffraction limit.The combination of parallel acceleration technology and existing algo-rithms can effectively improve the running speed,providing a certain reference for the restoration of degraded images caused by turbulence and noise.
关键词
图像复原/双通道/交替最小化/GPUKey words
image restoration/two-channel/alternating minimization/GPU引用本文复制引用
出版年
2024
NETL
NSTL
万方数据