Single Image Haze Removal from an Image Enhancement Perspective for Real-Time Vision Systems

Abstract

Vision systems operating outdoors are degraded by atmospheric turbidity, so haze removal has come into use as a 
pre-processing step for surveillance, remote sensing, navigation and target identification. This paper implements a 
single image haze removal algorithm intended for real-time use. Rather than inverting the physical model that 
describes how a hazy image forms, which requires estimating the transmission map and atmospheric light and is 
computationally expensive, the approach exploits computationally cheaper image processing operations: detail 
enhancement, multiple-exposure image fusion and adaptive tone remapping. Koschmieder's law is first used to deduce 
that artificially under-exposed versions of a hazy image contain regions of improved visibility, so that fusing several 
such exposures recovers a clearer result than any one of them. Detail enhancement is applied before artificial under
exposure by gamma correction, which mimics physical exposure adjustment, and a dark channel prior weighting 
scheme guides the fusion so that areas with clear visibility are the ones blended into the output. An adaptive tone 
remapping stage then restores brightness to the darkened fused result. The paper also covers the dark channel prior 
pipeline in detail, including atmospheric light estimation, transmission map estimation and refinement by bilateral, 
soft matting and guided filters, and evaluates results using entropy and the image quality index.