Hazy-to-Clear Translation with Contrastive Regularization
Qingxin Deng, Weidong Zhang · 2022 41st Chinese Control Conference (CCC) · 2022
Single Image Dehazing is an under-constrained challenge in computer vision due to the severe information degradation. Previous prior-based methods solve this problem based on various summarized observation on hazy or clear images while disciplined priors can hardly fit diverse complex hazy situations. Existing learning-based methods can produce more natural images while most of them focus on exploiting clear images while ignoring hazy images. Therefore, we proposed an novel Hazy-to-Clear translation network for single image dehazing, which equipped contrastive regularization built upon contrastive learning to make the best of both the hazy and clear images as negative and positive samples respectively. We termed our de-hazing network as HCTCR-Net, which consists of two cooperative branches: Hazy-to-Clear Translation Branch(HCTB) and Contrastive Regularization Branch(CRB). Specifically, HCTB take full advantage of the mutual resistance between the generator and discriminator to explore the latent feature distribution and guide the hazy-to-clear translation mapping by advesarial learning. Simultaneously, CRB devoted to pulling the restored images approach clear images and pushing far away from hazy images by contrastive learning, to implicitly regularize the embeddings of the restored image, hazy and clear images in representation space. Extensive experiments demonstrates that our method achieves outstanding dehazing performance on both synthetic and real-world dataset.