Study Of The Parallelization Of Multi-frame Blind Deconvolution Based On GPU Technology

Adaptive Optics(AO)can reduce the influence of the atmospheric turbulence,but it cannot reach the diffraction limit.Multi-frame blind deconvolution technology can enhance the resolution of AO images,which makes it one of the principle method of image post-treatment.The multi-frame blind deconvolution image restoration based on Bayes principle has used alternating variable method and optimization methods to get the object image and the point spread function(PSF),the algorithm is complicated and costs lots of time,it also needs more than 10 minutes to solve the problem.In this condition the monitor and recognize of the space target has been strongly baffled.Via studies and analyses the principle of the blind deconvolution and the structure of the algorithm,this paper takes the highly developing CPU-GPU heterogeneous system,study on the parallelization of the multi-frame blind deconvolution algorithm.This paper has three primarily work showing as follows:First,on the basis of analyzing and studying on the algorithm,this paper discusses the advantage and disadvantage between “joint variable”,“sequential alternating variable” and “parallel alternating variable”.Based on that,the improved parallel alternating variable is came up;thinking of the multi-frame restoration method based on recursion,this paper designs and discusses the data partitioning scheme,and designs the scheme in multi-GPUs system.This paper comes up with asymmetry data partitioning scheme between nodes and in the node.The result of simulating experiment claims that this two method can improve the ability of the algorithm in multi-GPUs system,and lay the foundation of the expansion to multi-GPU system.Second,CUFFT library is used to speed up the FFT which cost most time when running the program.Based on that,the program is optimized to deeper step according to the basic methods and principle,so that the parallel program of multi-frame blind deconvolution is actualized.According to the experiment,the parallel program has a big speedup comparing to the serial program,and accomplished the speed-up ratio of 17 times.The outcome of this paper ameliorate the problem of image restoration takes too long when calculating,improve the speed,reduce the running time,lay the foundation of the expansion to more complicated system.The achievement has an important significance and particial value in the area of monitoring and recognizing the space target in a short time.