Contents I. Introduction
Image dehazing is an essential pre-processing step in different computer vision applications. Various techniques have been developed based on physical characteristics and imaging models to address this challenging problem, few are summarized here. Ren et al. [1] use fusion based encoder-decoder architecture and a multi-scale gated fusion network (GFN) to achieve dehazed images. Bui and Kim [2] presented dehazing technique with a prior basing upon color ellipsoid and clustering to reduce haze and noise levels while minimizing the computational complexity. Cheng et al. [3] proposed a color priors approach basing upon extracted semantic features for single image dehazing. The design accurately recovers clean scene under strong estimation ambiguity, e.g. strong haze and semi-saturated ambient illumination, with learned semantic priors. Zhang and Patel [4] proposed a deep learning- based dehazing method that including a encoder-decoder which is densely connected along with a pooling module of multi-level and (GAN) framework generative adversarial network based on a joint discriminator. Li et al. [5] proposed a conditional GAN for end to end trained network with an encoder and decoder architecture used to capture useful information. Li et al. [6] presented a flexible cascaded convolutional neutral network (CNN) to dehaze images, which considers global airlight and medium transmission jointly using two task-driven subnetworks.
