3D data denoising via Nonlocal Means filter by using parallel GPU strategies.

Area of Science:

• Computer Vision
• Image Processing
• High-Performance Computing

Background:

• The Nonlocal Means (NLM) algorithm is a leading image denoising filter.
• Its computational demands necessitate parallel processing, particularly on Graphics Processing Units (GPUs).
• Existing 2D NLM on GPUs process 3D data slice-by-slice, limiting efficiency.

Purpose of the Study:

• To design and implement a fully 3D Nonlocal Means parallel algorithm.
• To explore various algorithm mapping strategies on GPU architectures and multi-GPU frameworks.
• To demonstrate the scalability and applicability of the 3D NLM approach.

Main Methods:

• Development of a fully 3D Nonlocal Means algorithm.
• Implementation utilizing parallel programming on GPU architectures.
• Testing diverse algorithm mapping strategies across single and multi-GPU setups.

Main Results:

• The fully 3D NLM approach achieves efficient denoising for 3D datasets.
• The method demonstrates significant scalability on GPU and multi-GPU systems.
• Experimental results validate the approach's performance.

Conclusions:

• The proposed fully 3D NLM parallel approach offers a scalable and efficient solution for image denoising.
• This method is highly applicable to various fields, including Magnetic Resonance Imaging (MRI) and video denoising.
• The study highlights the potential of advanced GPU parallelization for complex image processing tasks.