Analysis and Improvement of a Zero-Shot Image Dehazing Method

Abstract

[eng] Images captured under adverse weather conditions often result in low visual quality. In particular, haze is caused by suspended particles in the air, which scatter and attenuate both the light reflected from the scene and the atmospheric light from the hazy medium. Hazy images reduce the effectiveness of image processing and computer vision tasks, and its removal is essential for different applications. The process of eliminating haze is known as image dehazing. In this project, we aim to perform a comprehensive analysis of an existing zero-shot image dehazing method (ZID). ZID is a learning-based approach that employs three networks to generate a haze-free image, a transmission map, and an atmospheric light image. The parameters involving the three networks are trained simultaneously in an unsupervised manner. The examination of ZID will include an analysis of its current limitations and, apart from the reconstructed dehazed image, we plan to specifically inspect the resulting transmission and atmospheric light maps. Then, we aim to propose several modifications to cope with the found disadvantages present in the original model. Finally, we will include an extensive series of experiments with different datasets to evaluate both quantitatively and qualitatively the original method and the effect of the proposed modifications