Low-Light Image Enhancement Using Hybrid Deep-Learning and Mixed-Norm Loss Functions

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Low-light conditions degrade image quality, impacting various applications.
• Existing image enhancement methods often struggle with noise, color distortion, and artifacts.

Purpose of the Study:

• To develop an advanced low-light image enhancement technique.
• To improve illumination, reduce noise, and minimize color distortion in images captured under poor lighting.

Main Methods:

• A hybrid deep-learning network comprising decomposition-net, illuminance enhance-net, and chroma-net.
• Utilizing YCbCr color space for training and restoration to account for RGB channel correlations.
• Employing a mixed-norm loss function to enhance training stability and image clarity.

Main Results:

• The decomposition-net effectively separates reflectance and illuminance, enabling accurate feature extraction and noise reduction.
• The illuminance enhance-net boosts brightness while mitigating halo artifacts.
• The chroma-net independently addresses and reduces color distortion.
• The mixed-norm loss function contributes to stable training and sharper reconstructed images.

Conclusions:

• The proposed hybrid deep-learning network offers significant improvements in low-light image enhancement.
• The method demonstrates superior subjective and objective performance compared to existing state-of-the-art techniques.