A novel neural network for super-resolution remote sensing image reconstruction

Abstract

An accurate super-resolution image (SR image) reconstruction of remote sensing images (RSI) for preserving quality during the process of super-resolution conversion is crucial for many scientific and operational applications. Recent studies on supervised and unsupervised machine learning methodologies of SR image reconstruction have demonstrated their great potential for higher reconstruction performance in obtaining accuracy and quality. In this paper, a novel neural network with barycentric weight function (BWFNN) is proposed as a non-linear mapping function selected from the features of reference images. The whole process includes an online reconstruction phase and an offline training phase. In these phases, an edge orientation-based pre-learned kernel is introduced to describe and reference prior information, and a simple interpolation-like structure is followed to avoid any conventional iterative computation and lead to fast reconstruction. The innovation of this work is the BWFNN, which uses a non-linear barycentric weight function (BWF) to reconstruct the image details. Compared with most of the conventional reconstruction approaches, the proposed algorithm performs better in terms of peak signal-to-noise ratio (PSNR) and structural similarity (SSIM), and the model exhibits significant efficiency in reconstructing the image details.