Photo-Realistic Image Dehazing and Verifying Networks via Complementary Adversarial Learning.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Physical model-based dehazing methods struggle with environmental variables and artifacts like non-collected illuminance, halo, and saturation.
• Accurate estimation of illuminance, light transmission, and airlight is challenging, leading to high computational complexity.
• Existing methods often fail to produce natural-looking results and are prone to undesired artifacts.

Purpose of the Study:

• To develop a novel dehazing and verifying network (DVNet) for direct radiance estimation of hazy images.
• To overcome the limitations of physical model-based approaches by avoiding complex environmental variable estimations.
• To generate naturally dehazed results with improved quality and reduced artifacts.

Main Methods:

• Proposed a novel dehazing and verifying network (DVNet) comprising a correction network (CNet) and a haze network (HNet).
• CNet enhances clean images using ground truth to learn HNet, while HNet restores haze images.
• Employed a self-supervised learning method for error verification and a complementary adversarial learning approach for natural results, utilizing an unpaired dataset.

Main Results:

• The DVNet demonstrated superior performance in generating dehazed images compared to state-of-the-art methods.
• Experimental results confirmed that DVNet outperforms existing approaches under various hazy conditions.
• The complementary adversarial learning produced more natural-looking dehazed images.

Conclusions:

• The proposed DVNet effectively addresses the limitations of physical model-based dehazing methods.
• DVNet offers a robust and efficient solution for image dehazing, producing high-quality results.
• The network's ability to learn from unpaired datasets and its self-supervised verification enhance its practical applicability.