Residual Quotient Learning for Zero-Reference Low-Light Image Enhancement.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Neural networks dominate low-light image enhancement (LLIE).
• Many LLIE networks use Retinex-related architectures.
• Existing Retinex-based LLIE structures are suboptimal for non-uniform illumination.

Purpose of the Study:

• To address the limitations of current LLIE architectures.
• To introduce a novel variant learning framework for improved low-light image enhancement.
• To develop a lightweight and effective network for real-world LLIE tasks.

Main Methods:

• Proposed a novel framework: residual quotient learning.
• Reformulated light enhancement as predicting a latent quotient.
• Developed ResQ-Net, a lightweight network leveraging residual quotient learning.
• Employed a reference-free loss function for zero-reference optimization.

Main Results:

• ResQ-Net demonstrates enhanced non-uniform illumination modeling.
• The proposed method outperforms state-of-the-art LLIE techniques.
• Experiments show significant qualitative and quantitative improvements on benchmark datasets.
• Preliminary results in dark face detection confirm practical feasibility.

Conclusions:

• Residual quotient learning offers a superior approach to LLIE compared to existing methods.
• ResQ-Net is a flexible, generalizable, and effective network for real-world LLIE.
• The proposed method shows promise for practical applications like dark face detection.