GPU-Accelerated Adaptive Dictionary Learning and Sparse Representations for Multispectral Image Super-resolution

Abstract

Recently, single image super-resolution (SISR) based on sparse representations has been gaining much attention from the research community in the field of remote sensing. In this paper, a fast SISR reconstruction framework is developed for multispectral remote sensing (MSRS) images based on adaptive dictionary learning and sparse representations. It consists of two major parts: first, a novel super-resolution approach is developed for MSRS using sparse coding and adaptive dictionary learning. High-frequency features present in the input low-resolution MS image are extracted by using Butterworth low-pass, difference of Gaussian (DoG), and Sobel filters in horizontal and vertical directions. The proposed feature extraction method reveals the edges and other detailed information present in the MS image effectively. Secondly, massively parallel algorithms are designed for adaptive dictionary learning and sparse reconstruction using the Compute Unified Device Architecture (CUDA)-enabled General Purpose-Graphics Processing Unit (GP-GPU) programming model. The proposed method GP-GPU implementation not only gives better results in terms of visual quality and objective fidelity criteria, but also significantly reduces the computation time compared to its CPU counterparts to achieve near-real time operating speed.