Abstract
Many conventional single image dehazing methods use human-designed priors to estimate transmission and recover the latent scene. However, it is hard to design a universal prior for different real-work hazy scenarios. Though convolutional neural networks (CNNs) have also been utilized to perform end-to-end dehazing, which are dependent on the training data tightly. In this work, we provide a novel framework to incorporate deep priors into iteration scheme to adaptively optimize the transmission map. Specifically, we first establish our fundamental iteration by solving the implicitly regularized transmission optimization model based on the widely used alternating direction method of multiplier (ADMM). We then design a residual CNN to extract deep priors from a small training dataset (compared with existing dehazing networks) to guide the ADMM iterations, resulting to our final adaptive transmission optimization formulation. Extensive experimental results show that our proposed approach can produce compelling results superior to those obtained by others.