Novel Super Resolution Restoration of Remote Sensing Images Based on Compressive Sensing and Example Patches-Aided Dictionary Learning

Abstract

A novel machine learning and compressive sensing (CS) based super-resolution (SR) algorithm for the restoration of remote sensing images is proposed in this paper. This new algorithm relies on the idea that high-resolution (HR) image patches can be correctly recovered from the downsampled low- resolution (LR) image patches under two mild conditions, i.e., the sparsity of image patches, and the incoherence between the sensing and projection matrix. Consequently if most of HR image patches can be represented as a sparse linear combination of elements from a dictionary that is incoherent with sensing matrix, the HR image patches can be recovered accurately from its LR version. To find a dictionary which can sparsely represent HR image patches to guarantee the reconstruction error over a set of patches be minimal, an example patches-aided dictionary learning algorithm named KSVD algorithm is adopted. Moreover, the incoherence between the learned dictionary and sensing matrix is experimentally investigated. The new proposed method is tested on the restoration of remote sensing images came from USC-SIPI Image Database, and the results show that the proposed algorithm can provide substantial improvement in resolution of remote sensing images, and the restored images are superior in quality to that of other related methods.