Single Image Dehazing for Homogeneous and Non-Homogeneous Haze using Deep Neural Network

Abstract

Image dehazing is an "ill-posed problem" that has been extensively studied in recent years. Available methods use various constraints/priors, deep learning, or a combination of both to get plausible dehazing solutions. This thesis work reviews some recent advancements and benchmarking, mainly focusing on proposed solutions and their results on both homogeneous and non-homogeneous haze datasets. Intending to achieve haze removal for both types of haze, we propose a new deep learning architecture. This convolution neural network is based on the reformulated atmospheric scattering model(ASM) and haze density estimation model to extract features for both types of haze. Model is trained on perceptual loss. Results on both indoor, outdoor homogeneous haze image datasets demonstrate our superior performance compared with other top deep learning architectures in terms of SSIM and PSNR. while on non-homogeneous haze dataset, the proposed model performs inferiorly compared to the state of the art non-homogeneous haze targeted dehazing model but much better than other homogeneous haze targeted dehazing models.