Dark2Light: multi-stage progressive learning model for low-light image enhancement

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Low-light image enhancement (LLIE) is difficult due to severe noise and low illuminance.
• Simple illumination restoration amplifies noise, hindering effective signal recovery.

Purpose of the Study:

• To develop a multi-stage model, Dark2Light, for progressive restoration of low-light images.
• To address the dual challenges of illumination enhancement and noise removal simultaneously.

Main Methods:

• Converted images from sRGB to linear RGB for linear illumination enhancement.
• Employed a contextual transformer block for coarse-to-fine illumination enhancement.
• Utilized a U-Net shaped denoising block for noise removal and a dual-supervised attention block for progressive restoration.

Main Results:

• Dark2Light progressively restores normal-light images from low-light inputs.
• The model effectively enhances illumination while mitigating noise amplification.
• Quantitative and qualitative experiments show superior performance compared to state-of-the-art LLIE methods.

Conclusions:

• Dark2Light offers an effective solution for challenging low-light image enhancement tasks.
• The multi-stage approach balances illumination restoration and noise suppression.
• The proposed method advances the field of low-light image restoration.