Fast and Accurate Poisson Denoising With Trainable Nonlinear Diffusion.

Area of Science:

• Image processing
• Computational imaging
• Signal processing

Background:

• Poisson noise degrades image quality in applications like medical imaging and microscopy.
• Existing Poisson denoising methods prioritize performance over computational efficiency.

Purpose of the Study:

• To develop an efficient Poisson denoising model balancing high computational speed and image recovery quality.
• To adapt the Trainable Nonlinear Reaction Diffusion (TNRD) model for effective Poisson noise reduction.

Main Methods:

• Adapted the TNRD model using proximal gradient descent, overcoming limitations of direct gradient descent.
• Retrained model parameters, including linear filters and influence functions, incorporating Poisson noise statistics.
• Leveraged the diffusion process for parallel computation on Graphics Processing Units (GPUs).

Main Results:

• Achieved competitive denoising results compared to state-of-the-art methods.
• Demonstrated a simple model structure with high computational efficiency.
• GPU implementation processed images in under 0.1 seconds, delivering state-of-the-art performance.

Conclusions:

• The proposed TNRD-based model offers an efficient and effective solution for Poisson denoising.
• The model's suitability for GPU parallelization enhances its practical applicability.
• This approach provides a strong balance between performance and speed for various imaging applications.