基于GPU的多帧盲解卷积图像复原技术并行化实现

Parallel implementation of multiframe blind deconvolution based on GPU

多帧盲解卷积图像复原技术能够进一步提高自适应光学图像的分辨力,但其算法比较复杂,处理耗时过长,对序列图像复原经常需要几分钟甚至几十分钟的计算时间,对实际应用造成了极大不便。为了提升算法的运行速度,改善其耗时过长的问题,通过研究和分析盲解卷积算法原理和算法结构,采用目前高速发展的中央处理器(CPU)和图形处理器(GPU)异构加速技术,主要对耗时最长的矩阵卷积运算进行优化,通过使用库函数与算法结构微调相结合的方法并行加速,实现多帧盲解卷积的图像复原算法的并行化。使用并行算法对图像进行复原处理,针对16帧以上分辨率为256×256像素的空间目标图像,可以实现17×的加速比,为图像复原的实时/准实时提供一种可行的方案。

Multiframe blind deconvolution algorithm, which is one of the primary methods for restoring image, can enhance the resolving power of the adaptive optical images. However, the multiframe blind deconvotution algorithm takes the alternate minimization, optimize deconvo- lution methods to solve the images and the point spread functions （PSFs） of the objects, the arithmetic is complexed and time-consuming,it always needs several minutes even up to several quarters to solve the problem. Via developing the central processing unit-graphics processing unit （CPU-GPU） heterogeneous system architecture, through the combination of using library function and modulating the structure of the algorithm, we optimized the most time-consuming part, matrix convolution, and achieved the parallel methods of restoring image. Results show that for more than 16 frames of sPatical target images with 256 N 256 pixels,the speed-up ratio of 17× could be realized. This paper can provide a feasible scheme for real-time/quasi real-time- image restoration.