Spatially Adaptive Super-Resolution Via Deep Flow Estimation And Wiener Filtering For Satellite Imagery

Applications based on satellite imagery depend heavily on the resolution of the images captured by Earth observation satellites. The higher the resolution of the captured images, the better the precision in differentiating fine details in the acquired image, which is a key requirement for critical downstream applications. Though hardware-based methods for improving resolution exist, physical and sensor-based limitations such as optical and motion blur, and noise are difficult to avoid and result in low-resolution images. Super Resolution (SR) is one of the most promising fields in Remote Sensing (RS), which aims to computationally improve the resolution of an image, using one or many input Low Resolution (LR) images. This paper presents a spatially adaptive multiimage super-resolution approach using deep flow estimation and wiener filtering. Our method uses the best of both deep learning and classical approaches. The deep-flow estimation improves the registration performance whereas the spatially adaptive wiener filtering efficiently interpolates the information from low-resolution images to form the super-resolved output. Our method can process an arbitrary number of input LR frames and outputs a super-resolved image. We demonstrate the ability of the method to reconstruct high-frequency details from multiple low-resolution frames using Skysat imagery.