Cascading auto-regressive exponential smoothing of image sequences for reducing turbulence induced motion

Abstract

Atmospheric turbulence can significantly degrade images taken over a long horizontal path near the ground. This can hinder the identification of objects in a scene. We consequently introduce the Cascading Auto-Regressive Exponential Smoothing (CARES) algorithm, which is a fast real-time algorithm that suppresses the effects of atmospheric turbulence in image sequences. CARES is a spatial/temporal filtering algorithm that decomposes the image into a Laplacian Image Pyramid (LIP). Each component of the LIP represents the image smoothed to a specific length scale, which is then temporally filtered using an Auto-Regressive Exponential Smoothing (ARES) filter. The ARES filters have a cut-off frequency that are adjusted in such a way for each LIP component to define a critical velocity. Objects in the scene moving below the critical velocity pass through the CARES filter with little distortion or delay. We assess the performance of CARES using turbulent imaging data. We find that CARES improves image quality using a variety of image quality metrics. We use a simple CARES simulation to show that the salient features of a moving object lag behind by one pixel or less.