Tuning-free and self-supervised image enhancement against ill exposure

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Complex lighting and limited dynamic range cause ill-exposed images with information loss.
• Existing enhancement methods require manual tuning or exhibit poor generalization.
• Paired datasets for training are often inaccessible or imperfect.

Purpose of the Study:

• To develop a tuning-free image enhancement method for ill-exposed images using self-supervised learning.
• To improve image detail recovery and visual perception.
• To overcome limitations of existing methods regarding manual tuning and generalization.

Main Methods:

• A dual illumination estimation network estimates illumination for under- and over-exposed regions.
• Mertens' multi-exposure fusion strategy combines intermediate corrected images.
• Self-supervised learning enables global histogram adjustment for enhanced generalization.

Main Results:

• The proposed method reveals more details and offers better visual perception than state-of-the-art techniques.
• Significant improvements observed in image naturalness metrics (NIQE, BRISQUE) and contrast metrics (CEIQ, NSS).
• Achieved weighted average score boosts of 7%, 15%, 4%, and 2% on real-world datasets.

Conclusions:

• The correction-fusion approach adaptively handles various ill-exposed image types.
• Self-supervised learning enhances generalization and eliminates the need for paired datasets.
• The method provides superior performance in exposure correction and detail preservation.