Image De-Hazing Via Gradient Optimized Adaptive Forward-Reverse Flow-Based Partial Differential Equation

Abstract

In this study, we propose a modified partial differential equation (PDE)-based algorithm for image de-hazing. The algorithm possesses relatively low computational complexity and the core function of the PDE is easily amenable to hardware implementation. New contributions include the optimization and automated processing for dark and hazy images, avoiding manual parameter tuning. Additionally, the regularization parameter is computed adaptively from the binary mask of the input image. This is combined with a gradient-based metric for optimization to automatically determine stopping time of the algorithm for both types of images. The proposed scheme is fast and utilizes spatial or frequency domain filters to achieve illumination and reflectance component estimation without resorting to logarithms. Moreover, there is absence of halos in de-hazed images compared to previous work. Extensive experiments indicate that the proposed approach yields results comparable to or better than several works from the literature.