RefineDNet: A Weakly Supervised Refinement Framework for Single Image Dehazing.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Single image dehazing is crucial for many computer vision systems.
• Prior-based methods often create artifacts, while learning-based methods lack paired training data.
• Existing methods struggle with limited training data and generalization to diverse scenarios.

Purpose of the Study:

• To develop a novel framework for single image dehazing that overcomes limitations of existing methods.
• To merge the strengths of prior-based and learning-based approaches for improved haze removal.
• To enhance both visibility restoration and the realism of dehazed images.

Main Methods:

• Proposed a two-stage, weakly supervised dehazing framework named RefineDNet.
• Stage 1: Visibility restoration using the dark channel prior.
• Stage 2: Realness improvement via adversarial learning with unpaired data and a perceptual fusion strategy.

Main Results:

• RefineDNet demonstrated outstanding haze removal capability and produced visually pleasing results.
• The proposed perceptual fusion strategy enhanced the quality of dehazed images.
• Outperformed supervised and state-of-the-art dehazing methods on indoor and outdoor datasets, even with basic backbones.

Conclusions:

• RefineDNet effectively merges prior-based and learning-based techniques for superior single image dehazing.
• The weakly supervised approach addresses the challenge of limited paired training data.
• The framework offers a robust and visually appealing solution for haze removal in computer vision.