Semi-Blindly Enhancing Extremely Noisy Videos With Recurrent Spatio-Temporal Large-Span Network

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Capturing videos in extremely dark environments presents significant challenges due to complex noise.
• Existing physics-based and learning-based noise modeling methods have limitations, including complex calibration or performance degradation.

Purpose of the Study:

• To propose a semi-blind noise modeling and enhancing method for low-light videos.
• To develop an adaptive denoising process that overcomes the limitations of current methods.

Main Methods:

• Incorporation of a physics-based noise model with a learning-based Noise Analysis Module (NAM) for self-calibration.
• Development of a recurrent Spatio-Temporal Large-span Network (STLNet) featuring a Slow-Fast Dual-branch (SFDB) architecture and Interframe Non-local Correlation Guidance (INCG) mechanism.
• Investigation of spatio-temporal correlations in large spans for enhanced video denoising.

Main Results:

• The proposed method demonstrates effective self-calibration of noise model parameters.
• The STLNet architecture successfully leverages spatio-temporal correlations for improved denoising.
• Extensive experiments confirm the method's superiority in both qualitative and quantitative evaluations.

Conclusions:

• The semi-blind approach offers adaptive and robust performance for low-light video enhancement.
• The integration of NAM and STLNet provides a significant advancement in handling complex noise in dark environments.
• The proposed method achieves state-of-the-art results in challenging low-light video denoising scenarios.