Semi-supervised atmospheric component learning in low-light image problem.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Image quality is significantly impacted by ambient lighting and atmospheric conditions.
• Existing methods often lack instance-adaptive performance or require complex optimization.
• Data efficiency and post-prediction tuning are common challenges in low-light image restoration.

Purpose of the Study:

• To develop a semisupervised training method for low-light image restoration.
• To enhance image quality by incorporating physical properties and atmospheric models.
• To achieve instance-adaptive performance without extensive post-prediction tuning.

Main Methods:

• Utilizing a semisupervised training approach with no-reference image quality metrics.
• Incorporating a classical haze distribution model to analyze physical image properties.
• Minimizing a single objective function for efficient image restoration.

Main Results:

• The proposed method achieves competitive performance on six low-light datasets using no-reference metrics.
• Demonstrates improved generalization capabilities compared to state-of-the-art methods.
• Effectively preserves facial identities in extreme low-light scenarios.

Conclusions:

• The semisupervised method offers an efficient and effective solution for low-light image restoration.
• Integrating physical models enhances the adaptability and performance of image enhancement.
• The approach shows promise for real-world applications requiring high-quality image recovery.