Lightweight single image dehazing utilizing relative depth information

Abstract

Considering the need for lightweight and fast implementations, this paper presents an architecture based on a MobileVit encoder for efficiency and speed, introducing a fully convolutional lightweight decoder with skip connections for feature extraction. The main purpose of this network is to address the problem of single image dehazing. Recognizing the critical role of depth information in assisting the above task, the merging of these two tasks into a single network was performed in a supervised manner. Taking into account that there is a shortage of datasets that provide both dehazing and relative depth estimation ground truths, Depth Anything was utilized to extract the relative depth values of the images, which is the SOTA network in this task.