Unsupervised Single Image Dehazing Using Dark Channel Prior Loss.

Area of Science:

• Computer Vision
• Artificial Intelligence

Background:

• Single image dehazing is crucial for autonomous systems.
• Prior-based methods are slow; deep learning methods often use synthetic data, risking domain shift for outdoor scenes.

Purpose of the Study:

• To develop an unsupervised deep learning method for effective single image dehazing.
• To address the domain shift problem in dehazing by using real-world outdoor data.

Main Methods:

• Proposed a novel unsupervised training approach for deep neural networks (DNNs).
• Tuned network parameters by directly minimizing the Dark Channel Prior (DCP) energy function.
• Utilized real-world outdoor images exclusively, avoiding synthetic data.

Main Results:

• The 'Deep DCP' method significantly improves upon traditional DCP results.
• Achieved performance comparable to large-scale supervised dehazing methods.
• Demonstrated effective dehazing on real-world outdoor images without domain shift.

Conclusions:

• Unsupervised minimization of DCP via DNNs offers a powerful alternative to supervised learning.
• The proposed method provides a fast yet accurate solution for single image dehazing.
• This approach enhances the robustness of autonomous vision systems in various environmental conditions.