Real-world nighttime image dehazing using contrastive and adversarial learning

Abstract

Nighttime image dehazing is a challenging task due to the scarcity of real hazy images and the domain gap between synthetic and real data. To address these challenges, we propose a novel deep learning framework that integrates contrastive and adversarial learning. In the initial training phase, the dehazing generator is trained on synthetic data to produce dehazed images that closely match the ground truths while maintaining a significant distance from the original hazy images through contrastive learning. Simultaneously, the contrastive learning encoder is updated to enhance its ability to distinguish between the dehazed images and ground truths, thereby increasing the difficulty of the dehazing task and pushing the generator to fully exploit feature information for improved results. To bridge the gap between synthetic and real data, the model is fine-tuned using a small set of real hazy images. To mitigate bias from the limited amount of real data, an additional constraint is applied to regulate model adjustments during fine-tuning. Empirical evaluation on multiple benchmark datasets demonstrates that our model outperforms state-of-the-art methods, providing an effective solution for improving visibility in hazy nighttime images by effectively leveraging both synthetic and real data.

Lab Members

Patrick Chan

Associate Professor, Vice Dean

Patrick Chan works on machine learning, deep learning, image processing, adversarial learning, and secure machine learning.