Non-Homogeneous Haze Image Formation Model Based Single Image Dehazing

Abstract

Image dehazing is an image processing technique that involves removing the haze effect from a given image. Restoration of hazy images is currently an established computer vision issue due to its applications in the real world, like automated vehicles, surveillance, etc. This paper proposes a single image dehazing model by estimating atmospheric light and transmission map simultaneously using two networks. Atmospheric light and transmission map are estimated by a non-homogeneous light estimation network and a novel refined transmission map estimation network. The non-homogeneous light estimation network has two subnetworks: a white-balanced network and a deep convolutional neural network called Global Atmospheric Light Estimation Network. The white-balanced error network for correction of white-balanced error in a hazy image and a Global Atmospheric Light Estimation Network for estimation of global atmospheric light in a corrected white-balanced error hazy image. A refined transmission map estimation network is used to create a refined transmission map of the hazy image. The design of a refined transmission map estimation network is based on U-Net with skip connections, which estimates and refines the transmission map of the input haze image. The non-homogeneous haze image formation model takes both estimated global atmospheric light and a refined transmission map as input to restore the hazy-free image. The global atmospheric light estimation network and refined transmission map estimation network are trained using mean square error (MSE) and hybrid loss functions, respectively. In terms of structural similarity index Matric (SSIM) and peak signal-to-noise ratio (PSNR), the proposed model is equal to the cutting-edge approaches on the NR-Haze, I-Haze, and O-Haze datasets.