Domain Adaptation for Synthesis of Hazy Images

Most existing image dehazing methods based learning are less able to perform well real hazy images. An important reason is that they trained on synthetic images whose distribution different from To relieve this issue, paper proposes a new scene generation model domain adaptation, which uses variational autoencoder encode the pairs and into latent space adapt. The guide learn mapping of clear images, used adapt images’ through generative adversarial loss, so as make close possible. By comparing results with traditional physical scattering models Adobe Lightroom CC software, generated in more realistic. Our end-to-end adaptation also very convenient synthesize without depth map. Using method dehaze by paper, both SSIM PSNR have been improved, proved effectiveness our method. non-reference haze density evaluation algorithm other quantitative illustrate advantages