Single Image Dehazing Based on Depth Map Estimation via Generative Adversarial Networks

Abstract

Images taken in haze weather are characteristic of low contrast and poor visibility. The conventional algorithms are confined to manual designed features, while recently proposed machine learning based algorithms utilize Euclidean loss to generate the transmission map that may result in haze removal incompletely. In this thesis, we propose a single image dehazing method using depth map estimated by the Generative Adversarial Network (GAN). The proposed GAN model is trained to minimize the loss between the given input hazy image and their corresponding depth map, which aim to learn a pixel-wise nonlinear mapping between them. Then the transmission map is calculated by the depth map directly, following with the guided filter to refine the transmission map. The proposed GAN model is trained on synthetic indoor images, although it can be applied to real hazy images. We use the peak signal-to-noise ratio and the structural similarity to evaluate the performance. The experimental results demonstrate that the proposed method achieves better dehazing results against the state-of-the-art algorithms on both the real hazy images and the synthetic hazy images, in terms of qualitative performance and visual performance.