UCL-Dehaze: Toward Real-World Image Dehazing via Unsupervised Contrastive Learning.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Training image dehazing models on synthetic data causes domain shift issues.
• Collecting real-world hazy/clean image pairs is challenging.

Purpose of the Study:

• To develop an unsupervised contrastive learning paradigm for image dehazing.
• To address the domain shift problem using unpaired real-world images.
• To enhance the generalization ability of dehazing networks in real-world scenarios.

Main Methods:

• Proposing UCL-Dehaze, an unsupervised contrastive learning framework.
• Utilizing unpaired real-world clean and hazy images as positive and negative samples.
• Formulating a self-contrastive perceptual loss function for training.
• Employing adversarial training to align image distributions.

Main Results:

• UCL-Dehaze effectively leverages unpaired data for superior dehazing performance.
• The method demonstrates superiority over state-of-the-art techniques.
• Achieved strong results even with a limited dataset of 1,800 unpaired images.

Conclusions:

• Unsupervised contrastive learning with adversarial training is a viable approach for image dehazing.
• UCL-Dehaze alleviates the domain shift problem without requiring paired data.
• The proposed method offers enhanced generalization for real-world image dehazing.