Fast Multi-Grid Methods for Minimizing Curvature Energies.

Area of Science:

• Image processing
• Computational mathematics
• Computer vision

Background:

• Geometric regularization methods, like mean and Gaussian curvature, are vital for preserving image features such as edges and corners.
• A significant challenge in high-order regularization is balancing image restoration quality with computational efficiency.

Purpose of the Study:

• To develop fast multi-grid algorithms for minimizing mean and Gaussian curvature energy functionals.
• To address the trade-off between accuracy and computational speed in high-order regularization methods.
• To propose a robust and efficient algorithm for image processing tasks.

Main Methods:

• Fast multi-grid algorithms applied to mean and Gaussian curvature minimization.
• Operator splitting and Augmented Lagrangian method (ALM) are contrasted with the proposed approach.
• Domain decomposition for parallel computing and fine-to-coarse structure for accelerated convergence.

Main Results:

• The proposed method effectively preserves geometric structures and fine details in image denoising, CT, and MRI reconstruction.
• Achieved a processing time of 40s for a $1024\times 1024$ image, significantly faster than the 200s required by ALM-based methods.
• Demonstrated robustness by avoiding artificial parameters, unlike existing approaches.

Conclusions:

• The developed fast multi-grid algorithms offer a superior solution for high-order geometric regularization in image processing.
• The method provides a significant improvement in computational efficiency for large-scale image processing tasks.
• This work overcomes the efficiency roadblock in high-order regularization, enabling practical applications.