Unsupervised Single Image Dehazing via Disentangled Representation
Qian Liu · 2019
Image dehazing aims to recover the latent clear content from the corresponding degraded hazy image. In this paper, we propose an unsupervised method for single image dehazing based on disentangled representation. Our proposed method does not rely on the physical scattering model and does not need paired of training data. We propose a content encoder and a haze encoder to disentangle the content and hazy information from a hazy image respectively. We propose a latent regression loss to encourage the generated image to preserve haze information and to force the haze encoder to extract haze information from the haze image. The cycle-consistency loss are introduced to ensure that the dehazed images have the same content structures with the original images. We also use an adversarial loss on the dehazed images to guarantee haze free and visually realistic. Extensive experiment results on the public dehazing dataset RESIDE demonstrate that the proposed method outperforms state-of-the-art unsupervised methods, and can achieve comparable performance with the state-of-the- art supervised methods.