SID-Net: single image dehazing network using adversarial and contrastive learning

Image dehazing is a fundamental low-level vision task and has gained increasing attention in the computer community. Most existing learning-based methods achieve haze removal by designing different convolutional neural networks. However, these algorithms only consider clean images as optimization targets and fail to utilize negative information from hazy images, which leads to a sub-optimal dehazing performance. Towards this issue, we propose a novel single image dehazing network (SID-Net), and it consists of three core branches: Image Dehazing Branch (IDB), Adversarial Guidance Branch (AGB) and Contrastive Enhancement Branch (CEB). Specifically, IDB achieves an initial hazy-clean translation based on the encoder-decoder framework and enhances its feature representation ability by introducing an Attentive Recurrent Module (ARM) and Attention Fusion Operation (AFO), respectively. Next, AGB takes full advantage of positive information from clean ground truth by an adversarial learning strategy and guides the restored image to be closer to the haze-free domain. Finally, CEB is proposed to exploit the negative information of hazy images and further improve its dehazing performance via a contrastive learning strategy. Extensive experiments on both synthetic and real-world datasets demonstrate that our SID-Net can obtain comparable results with other state-of-the-art algorithms. Code is available at https://github.com/leandepk/SID-Net-for-image-dehazing.