A Deep Learning Approach to Single Image Dehazing Inspired by Euler Numerical Schemes
Yin Li, Wenze Shao · 2023
Haze concentration variation challenges traditional dehazing methods in accurately estimating atmospheric haze. Conventional techniques can result in over-smoothed images and visible artifacts. Recently, deep neural network-based methods have gained attention, with network module performance being crucial. This paper introduces ProgNet, a numerical scheme-inspired module enhancing interpretability by connecting network structure with the forward Euler progressive multi-step numerical scheme. ProgNet is integrated into PFF-Net and AOD-Net dehazing models, addressing limitations of the classical atmospheric scattering model using PFF-Net's extended dehazing model. Experiments on public datasets show the module's effectiveness in dehazing various scenes and improving PFF-Net and AOD-Net's performance.