Nonparametric multiscale blind estimation of intensity-frequency-dependent noise

Area of Science:

• Computer Vision
• Image Processing
• Signal Processing

Background:

• Image processing chains and camera calibration parameters are often unavailable for received images, such as scanned photographs and JPEGs.
• These images frequently undergo nonlinear contrast adjustments and linear/nonlinear filtering, complicating noise management.

Purpose of the Study:

• To develop a general nonparametric noise model that accounts for intensity and frequency-dependent noise characteristics.
• To establish an efficient noise estimation method applicable to images with unknown processing histories.
• To provide a preliminary step for patch-based denoising algorithms.

Main Methods:

• Introduction of a general nonparametric intensity and frequency-dependent noise model.
• Development of a patch model for noise estimation, requiring over 1000 parameters.
• Implementation of a novel sparse patch distance function to identify patches with similar underlying geometry.

Main Results:

• Demonstrated efficient noise estimation using both simulated and real image experiments.
• Validated the noise model and estimation method against ground-truth noise curves for raw and JPEG images.
• Achieved effective denoising results on real images, validated by visual inspection.

Conclusions:

• The proposed noise model and estimation method offer an efficient solution for noise assessment in images with unknown processing.
• The noise estimation serves as a valuable preprocessing step for patch-based denoising techniques.
• The method shows competitive performance compared to existing state-of-the-art approaches.