A Second-Order TGV Discretization with $90^{\circ}$ Rotational Invariance Property

Alireza Hosseini,Kristian Bredies

DOI: https://doi.org/10.48550/arXiv.2209.11450

2023-01-28

Abstract:In this work, we propose a new discretization for second-order total generalized variation (TGV) with some distinct properties compared to existing discrete formulations. The introduced model is based on same design principles as Condat's discrete total variation model (\textit{SIAM J. Imaging Sci}., 10(3), 1258--1290, 2017) and shares its benefits, in particular, improved quality for the solution of imaging problems. An algorithm for image denoising with second-order TGV using the new discretization is proposed. Numerical results obtained with this algorithm demonstrate the discretization's advantages. Moreover, in order to compare invariance properties of the new model, an algorithm for calculating the TGV value with respect to the new discretization model is given.

Numerical Analysis

What problem does this paper attempt to address?

The problem that this paper attempts to solve is that the existing discretization methods lack rotational invariance when dealing with image processing tasks (such as image denoising), especially in the case of 90 - degree rotation. Traditional discretization methods (such as the classical discrete total generalized variation (TGV) and isotropic total variation (TV)) cannot well maintain rotational invariance when dealing with image problems due to their inherent discretization methods, which may lead to a decline in the quality of processing results.

To solve this problem, the author proposes a new discretization method, aiming to improve the discretization of the second - order total generalized variation (TGV2) to make it 90 - degree rotation - invariant. Specifically, based on the idea of Condat's discrete total variation model, this method designs a new discretization scheme by introducing staggered grids and some basic operators to ensure invariance under 90 - degree rotation and also provide a better approximation at other rotation angles.

The following are the key points mentioned in the paper:

1. **Deficiencies of existing discretization methods**:
- The classical discrete total variation (TV) and total generalized variation (TGV) lead to a decline in the quality of processing results when dealing with image problems due to the lack of rotational invariance.
- Especially in the case of 90 - degree rotation, the performance of classical discretization methods is not good.

2. **Objectives of the new discretization method**:
- Propose a new discretization method to make the second - order total generalized variation (TGV2) invariant under 90 - degree rotation.
- Improve the quality of image processing tasks (such as denoising), especially when dealing with images with complex structures.

3. **Features of the new discretization method**:
- Based on the idea of Condat's discrete total variation model, design a new discretization scheme by introducing staggered grids and basic operators.
- Verify the effectiveness of the new method through numerical experiments, demonstrating its invariance under 90 - degree rotation and better approximation at other rotation angles.

In summary, the main contribution of this paper lies in proposing a new discretization method, which solves the problem of the lack of rotational invariance of existing methods when dealing with image problems, thereby improving the quality and reliability of image processing tasks.