A hybrid framework for curve estimation based low light image enhancement.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Low-light image enhancement (LLIE) is crucial for improving visibility in underexposed images.
• Existing LLIE methods often use global mapping and struggle with diverse degradations in dark regions.
• Convolutional Neural Networks (CNNs) used in LLIE have limitations in capturing long-range dependencies.

Purpose of the Study:

• To develop a hybrid framework (hybLLIE) for effective low-light image enhancement.
• To address limitations of existing methods by incorporating both transformer and convolutional architectures.
• To improve the modeling of valuable information in low-light regions and capture local context.

Main Methods:

• Proposed a hybrid framework (hybLLIE) combining transformer and convolutional designs.
• Introduced a light-aware transformer (LAFormer) block with a feature reassignment modulator for targeted information modeling.
• Utilized a SeqNeXt block for local context capture and a self-supervised mechanism with high-order curves for image brightening.

Main Results:

• The hybLLIE framework demonstrated strong performance in low-light image enhancement.
• Achieved comparable results against 17 state-of-the-art methods across 7 benchmark datasets.
• The proposed LAFormer and SeqNeXt blocks effectively addressed limitations of previous approaches.

Conclusions:

• The hybLLIE framework offers a robust solution for low-light image enhancement.
• The hybrid approach effectively balances global and local feature extraction for improved image quality.
• This work advances LLIE techniques by integrating advanced deep learning architectures.