Guided Filter-Inspired Network for Low-Light RAW Image Enhancement.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Low-light RAW image enhancement (LRIE) is crucial for practical applications.
• Deep learning, particularly fusion-based methods, shows promise for LRIE.
• Existing fusion methods inadequately exploit correlations between source images.

Purpose of the Study:

• To propose a novel network, GFNet, for improved LRIE.
• To enhance fusion-based methods by exploring physical correlations between source images.
• To better preserve details and reduce noise in low-light images.

Main Methods:

• Developed a Guided Filter-inspired Network (GFNet).
• Fused source images in a guided filter (GF)-like manner.
• Inferred fusion coefficients within image and feature domains.

Main Results:

• GFNet effectively captures intrinsic correlations between source images.
• Achieved superior performance in detail preservation and noise reduction for LRIE.
• Demonstrated broad applicability through extended use in guided low-light image enhancement.

Conclusions:

• GFNet significantly advances LRIE by improving information fusion.
• The guided filter-inspired approach offers a robust framework for image enhancement tasks.
• GFNet shows potential for various low-light imaging applications.