Parallel multi-frame blind deconvolution with the macopt minimizer based on NUMA nodes.

Multi-frame blind deconvolution (MFBD) can obtain high-resolution images degraded by the atmospheric turbulence with limited prior information, but it is time-consuming. To improve the efficiency of MFBD, we integrate the macopt minimizer into the alternate minimization framework and investigate the detailed parallelization methods using a message passing interface (MPI) and open multiprocessing (OpenMP) based on a multicore architecture. Three sets of simulated seasat images with sizes of 128 ∗ 128, 256 ∗ 256, 512 ∗ 512, and one set of real solar observations with size of 1600 ∗ 1600 are used for restoration experiments. Experimental results indicate that the MFBD using the macopt minimizer (MFBD-M) can achieve faster execution speed while maintaining the image restoration quality compared to the MFBD using a Brent line search method (MFBD-B), and the speedup ratios are 2.76, 2.57, 2.48, and 2.81, respectively. The parallel MFBD with macopt using MPI (-MFBD-M) can achieve speedup ratios of 7.87, 9.26, 10.22, and 5.98 compared to the MFBD-M. Meanwhile, the restoration quality of -MFBD-M can hold almost the same as that of the MFBD-M. Furthermore, by changing the data precision from double precision (DP) to single precision (SP) and the FFT type from complex-to-complex (C2C) to real-to-complex (R2C), the -MFBD-M achieves additional speedup ratios of 3.0, 3.21, 3.88, and 3.69 respectively.