GAN-based dehazing network with knowledge transferring

Capturing images under the condition of haze often shows low contrast and fades the color. Restoring the haze-free image from a single image is a challenging task due to the ill-pose of the problem and high degradation. To solve this problem, we propose a GAN (Generative Adversarial Network) Prior Guided Dehazing Network (GPGDN). While the prior dehazing methods often trained the model with adversarial loss to obtain a photorealistic dehazed result, The proposed method explores to transfer of the rich and diverse priors learned from large clean images to dehazing problem. The GPGDN consists of an Encoder and a GAN-based decoder. The Encoder is designed to generate the latent code and noise input, which are fed to GAN-based decoder and generate the final dehazed result. Due to the high degradation of dense haze areas, it is hard to restore high-quality results for these areas. The proposed method can transfer knowledge from the haze-free images into dehazed results and restore high-quality results. The experiment on simulated outdoor hazy images demonstrates that the proposed method outperforms other methods with a significant gap of 3.40dB. Hazy images dehazing by GPGDN show a clear improvement compared to prior methods.