Combining Pre- and Post-Demosaicking Noise Removal for RAW Video

Area of Science:

• Computer Vision
• Image Processing
• Deep Learning

Background:

• Denoising is crucial for converting raw camera data to display-ready images.
• Deep learning has advanced denoising, but challenges remain in adapting to varying noise levels and scenes.
• Traditional denoising often precedes demosaicking, though alternative orders exist.

Purpose of the Study:

• To propose a self-similarity-based denoising scheme for Bayer-patterned CFA video data.
• To balance pre- and post-demosaicking denoising for optimal image quality.
• To enhance texture reconstruction and adapt to diverse noise conditions.

Main Methods:

• A self-similarity-based scheme weighting pre- and post-demosaicking denoisers.
• Integration of temporal trajectory prefiltering before each denoiser.
• Noise model estimation at the sensor for adaptation.

Main Results:

• A balance between pre- and post-demosaicking denoisers yields superior image quality.
• Higher noise levels benefit more from pre-demosaicking denoising influence.
• Temporal prefiltering improves texture reconstruction accuracy.

Conclusions:

• The proposed method adapts accurately to any noise level with minimal input.
• It achieves state-of-the-art performance, making it suitable for real-world videography.
• The approach effectively addresses the limitations of current neural network denoising techniques.