Edge-Preserving Single Remote-Sensing Image Super-resolution Using Sparse Representations

Multispectral (MS) sensors are mostly of low resolution (LR) and fail to give promising results in remote-sensing applications. In the recovery of edge information from LR images, the sparse representation-based single image super-resolution (SISR) employing patch-based dictionary alone does not give satisfactory results. To overcome this, we propose a parallel SISR framework based on edge-preserving dictionary learning and sparse representations on compute unified device architecture (CUDA)-enabled graphics processing units (GPU). To recover edges, multiple coupled dictionaries, namely, the scale-invariant feature transform (SIFT) keypoints and non-keypoints patch-based dictionaries, are learned. A joint reconstruction model is also designed based on SIFT keypoints-guided patch sparsity and non-local total variation (NLTV)-based gradient sparsity. Simulation results show that the proposed method not only performs better than state-of-the-art methods in terms of visual quality and objective criteria, but also enhances the speed, implying a great potential for real-time applications.