Semi-Supervised Domain Alignment Learning for Single Image Dehazing.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Convolutional neural networks (CNNs) show promise in single-image dehazing but struggle with generalization due to reliance on synthetic training data.
• Existing methods face challenges in adapting to natural hazy images because of the domain gap between synthetic and real-world data.

Purpose of the Study:

• To develop a semi-supervised learning approach for single-image dehazing that improves generalization on natural images.
• To address the domain shift problem between synthetic and realistic hazy images.
• To enhance the adaptive response to varying haze densities in real-world scenarios.

Main Methods:

• A semi-supervised learning framework leveraging both synthetic and realistic images for training.
• A domain alignment module to minimize distribution differences between synthetic and real hazy image features.
• A haze-aware attention module for adaptive processing based on regional haze density.
• Integration of the dark channel prior to refine unsupervised learning using haze-free image statistics.

Main Results:

• The proposed method effectively bridges the domain gap, significantly improving generalization performance on natural hazy images.
• Experimental results demonstrate state-of-the-art performance on both synthetic and realistic datasets.
• The approach yields superior visual quality in dehazed images compared to existing methods.

Conclusions:

• The semi-supervised approach with domain alignment and haze-aware attention offers a robust solution for single-image dehazing.
• This method enhances the practical applicability of deep learning models for image restoration in real-world conditions.
• The integration of domain adaptation and attention mechanisms is crucial for achieving high-performance dehazing.