Zero-Shot Sand-Dust Image Restoration.

Area of Science:

• Computer Vision
• Image Restoration
• Atmospheric Optics

Background:

• Natural sand-dust weather presents complex challenges for image processing.
• Existing sand-dust image enhancement methods struggle with real-world scene variations.
• Robust restoration is crucial for advancing computer vision capabilities in adverse conditions.

Purpose of the Study:

• To develop a robust, unsupervised zero-shot learning method for sand-dust image restoration.
• To address limitations of current enhancement techniques in complex natural environments.
• To improve the performance of computer vision systems dealing with sand-dust imagery.

Main Methods:

• Proposed a novel zero-shot learning approach leveraging an atmospheric scattering physics model.
• Implemented an unsupervised training strategy, eliminating the need for prior knowledge or image pairs.
• Developed a method to learn and infer transmission and atmospheric light from single sand-dust images.

Main Results:

• The proposed method demonstrated superior performance in qualitative and quantitative evaluations.
• Achieved better sand-dust image restoration compared to state-of-the-art algorithms.
• Successfully restored images without reliance on synthetic datasets or paired training data.

Conclusions:

• The developed zero-shot learning method offers a robust solution for sand-dust image restoration.
• Unsupervised learning based on physical models is effective for complex image restoration tasks.
• The technique significantly enhances information processing for computer vision in sand-dust conditions.