Evaluation of neural network algorithms for atmospheric turbulence mitigation

TL;DR

This paper presents an overview of existing neural networks architectures, including a network trained in an end-to-end fashion, thereby removing the need for a stabilization step and performs experiments to remove the blur caused by atmospheric turbulence.

Abstract

A variety of neural networks architectures are being studied to tackle blur in images and videos caused by a non-steady camera and objects being captured. In this paper, we present an overview of these existing networks and perform experiments to remove the blur caused by atmospheric turbulence. Our experiments aim to examine the reusability of existing networks and identify desirable aspects of the architecture in a system that is geared specifically towards atmospheric turbulence mitigation. We compare five different architectures, including a network trained in an end-to-end fashion, thereby removing the need for a stabilization step.

Figures (4)

• Figure 1: Examples of available sequences in the SOTIS dataset. The used ground-truth images are given in the left column. The corresponding weak and strong turbulence scenarios are illustrated in the center and right columns, respectively.
• Figure 2: Performances results for Non-Deep Learning algorithms.
• Figure 3: Performances results for Re-trained Deep Learning algorithms.
• Figure 4: Performances results for Pre-trained Deep Learning algorithms.