Recent Trends in Deep Learning Based Omnidirectional Image SuperResolution

Omnidirectional imaging is extensively used in fields such as virtual reality, surveillance, and remote sensing. Super-resolution aims at enhancing the detail and resolution of low-resolution images, including omnidirectional ones. This article provides an overview of deep learning methods for achieving super-resolution in omnidirectional images. Using techniques like CNNs-convolutional neural networks and GANs-generative adversarial networks, deep learning has proven highly effective in improving the quality and resolution of omnidirectional images. The study explores various techniques, including models for single-image super-resolution, methods for multi-frame fusion, and the challenges faced in applying deep learning to omnidirectional image super-resolution. The article also highlights the method used in each technique, discusses potential applications, and identifies future research directions. It lays the groundwork for further advancements in this exciting field of deep learning-based methods to apply super-resolution in omnidirectional images.

I. Introduction

The recovery of an image with high resolution (HR) using several low-resolution (LR) versions is the goal of the extensively researched problem of super-resolution (SR) in image processing (Yang et al., 2014). A great deal of research has been put into this area, using either conventional method for image processing or quickly developing deep learning-based techniques [1], [2]. Super-resolution imaging is generally challenging due to limited information at hand and the loss of details, rendering it an ill-posed problem. The two primary categories into which SR algorithms can be divided are SISR- single image super resolution (Tang and Chen; Tsurusaki et al.; Cheng et al.), which aims to retrieve the original information from single image, and multiframe SR (Hung and Siu, 2009; Btz et al., 2016), a conventional approach that uses information derived from several frames. SISR methods can be separated into methodologies based on learning and interpolation. In the present article, we concentrate on the use of deep learning for super-resolution of omnidirectional image.