Brightness-Aware Synthetic-to-Real Learning for Nighttime Hazy Image Enhancement.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Nighttime hazy vision significantly degrades visual perception due to atmospheric haze and complex lighting.
• Existing daytime image dehazing methods are insufficient for the unique challenges of nighttime scenarios.
• Game engine simulations offer generalization but present brightness inconsistencies for nighttime dehazing.

Purpose of the Study:

• To develop an effective method for nighttime image dehazing.
• To address the limitations of synthetic data in real-world nighttime image enhancement.
• To improve visual clarity and reduce haze in nighttime imagery.

Main Methods:

• A three-step, brightness-aware synthetic-to-real learning approach was introduced.
• Supervised learning trained a spatial-frequency network (SFN) on synthetic data for pseudo-label generation.
• Semi-supervised learning (SFN+) minimized domain discrepancy using brightness consistency loss, followed by fine-tuning (SFN++) for relative brightness improvement.

Main Results:

• The proposed method demonstrated superior performance compared to state-of-the-art approaches on benchmark datasets.
• The brightness-aware strategy effectively handled unrealistic brightness in synthetic data.
• The three-step approach successfully improved nighttime image quality and dehazing.

Conclusions:

• The developed brightness-aware synthetic-to-real learning approach significantly advances nighttime image dehazing.
• The method offers a robust solution for enhancing visibility in challenging nighttime conditions.
• This work provides a foundation for future research in low-light and adverse weather image enhancement.