GPU-parallel implementation of the autoregressive model interpolation for high-resolution remote sensing images

Abstract

The autoregressive modeling image interpolation scheme is noticeably closer to ideal interpolation aiming at obtaining a high-resolution (HR) image from its low-resolution (LR) version than conventional methods. The basic idea is to first estimate the covariance of HR image from the covariance of the LR image and then adjust the covariance coefficients of HR image according to a feedback mechanism that takes into account the mutual influence between the estimated missing pixels in a local window. In spite of its impressive performance, the time-consuming computation is usually the bottleneck of the method when it is applied in time-critical scenario. Graphics Processing Units (GPUs) are attractive candidates to expedite the computation process. In this paper, an efficient GPU-based massively parallel version of the autoregressive modeling image interpolation scheme was proposed. Because all pixels which need to be interpolated have no dependence, each estimated pixel is assigned to independent thread in our parallel interpolation scheme. Experimental results show that we reached a speedup of 21.2x when I/O transfer time was taken into account, with respect to the original single-threaded C CPU code with the -O2 compiling optimization.