Edge-preserving image denoising via group coordinate descent on the GPU

Area of Science:

• Computer Vision
• Image Processing
• Scientific Computing

Background:

• Image denoising is crucial for enhancing image quality and is a key step in many image reconstruction tasks.
• Increasing data size and stalled serial processing speeds necessitate parallel computing solutions.
• Graphics Processing Units (GPUs) offer massive parallelism suitable for computationally intensive image processing.

Purpose of the Study:

• To develop and present a family of image denoising algorithms optimized for GPU execution.
• To address memory limitations on GPUs by implementing in-place pixel updates.
• To handle diverse edge-preserving roughness penalties, including anisotropic total variation.

Main Methods:

• Algorithms perform iterative, independent, parallel 1D pixel-update subproblems.
• In-place updates minimize memory usage, storing only essential data.
• The majorize-minimize framework is employed to solve the 1D subproblem.
• Support for various edge-preserving penalties like differentiable convex and anisotropic total variation.

Main Results:

• Demonstrated rapid convergence in both iteration count and execution time.
• Successfully applied to large-scale 2D image denoising.
• Validated on 3D medical imaging denoising tasks.
• Achieved efficient denoising on GPU architectures.

Conclusions:

• The proposed GPU-accelerated algorithms offer a significant speedup for image denoising.
• These methods are effective for both standard and complex imaging applications, including medical imaging.
• The parallel approach effectively addresses the computational demands of modern image processing.