A Multiscale Denoising Framework Using Detection Theory with Application to Images from CMOS/CCD Sensors

Area of Science:

• Image processing
• Signal processing
• Computational imaging

Background:

• CMOS and CCD imaging sensors produce noisy images due to electronic fluctuations and low photon counts.
• Image noise is often modeled as signal-dependent Poisson or a mix of Poisson-Gaussian noise.
• Existing denoising methods struggle with complex noise models.

Purpose of the Study:

• To develop a generalized framework for denoising images affected by Poisson or Poisson-Gaussian noise.
• To improve image quality from CMOS and CCD sensors.
• To provide a robust solution for signal-dependent noise removal.

Main Methods:

• Utilized variance stability transformation (VST) to convert signal-dependent Poisson noise into signal-independent Gaussian noise.
• Applied multiscale transforms to separate image signal and noise components.
• Employed local binary hypothesis testing with empirical distribution functions (EDF) for noise detection and removal.

Main Results:

• The proposed framework effectively denoises images with Poisson and Poisson-Gaussian noise.
• Demonstrated successful application across various multiscale transforms.
• Validated effectiveness on diverse input datasets, showing significant noise reduction.

Conclusions:

• The generalized framework offers a robust and effective solution for denoising images from CMOS and CCD sensors.
• VST combined with multiscale analysis and hypothesis testing provides a powerful approach for complex noise models.
• The method shows promise for enhancing image quality in various scientific and technical applications.