Color image denoising based on low-rank tensor train

Yang Zhang, Zhi Han, Yandong Tang2019 · DOI

TL;DR

This paper introduces a novel approach for denoising color images by leveraging the power of Tensor Train (TT) decomposition, a method that excels at capturing high-dimensional data structures. They propose a new algorithm, TT-Mac, based on block coordinate descent, which effectively recovers clean images from noisy ones, outperforming existing methods both visually and numerically.

What Problem Does This Paper Address?

Image denoising is a fundamental problem in computer vision, aiming to remove noise from images while preserving important details like edges. While traditional methods exist, they often struggle with the high dimensionality and complex spatial structures inherent in natural images, especially color images which can be represented as 3D tensors (height, width, color channels). Existing tensor decomposition methods like CP and Tucker have limitations, such as difficulty in determining optimal ranks or unbalanced mode representations. This paper addresses the need for a more effective tensor-based denoising method that can fully exploit the spatial information in color images.

Methods

The core of this paper is the application of Tensor Train (TT) decomposition for color image denoising. TT decomposition represents a high-order tensor as a series of interconnected lower-order tensors (factors), which is more expressive and computationally efficient for capturing correlations across multiple dimensions compared to traditional tensor decompositions.

The authors propose a novel algorithm called TT-Mac (Denoising algorithm by parallel matrix factorization based on tensor train). This algorithm is built upon the Block Coordinate Descent (BCD) method. BCD iteratively optimizes groups of variables while keeping others fixed. Specifically, TT-Mac addresses the denoising problem by reformulating it as a low-rank tensor train optimization problem and then solving it by iteratively updating the TT factors using a parallel matrix factorization approach. This involves unfolding the tensor, updating the factors using least squares, and then refolding them to reconstruct the denoised tensor. They also introduce Ket Augmentation (KA) to convert a 3rd-order color image tensor into a higher-order tensor, which they argue can better capture richer textures and inter-mode correlations.

Datasets

The paper evaluates its method on real color images. Specific datasets are not explicitly named, but the figures show examples of the "Lena," "Peppers," "Starfish," and "Pine" images, which are common benchmarks in image processing research.

Key Insights and Ideas

  • Tensor Train's Advantage: TT decomposition is shown to be a powerful tool for image denoising due to its ability to efficiently capture high-dimensional spatial correlations in color images, outperforming traditional tensor decomposition methods.
  • BCD for TT Denoising: The proposed TT-Mac algorithm effectively leverages the Block Coordinate Descent method to solve the complex TT-based denoising optimization problem, making it practical for real-world applications.
  • Ket Augmentation for Texture: The introduction of Ket Augmentation (KA) is a clever way to increase the tensor order, potentially enhancing the capture of intricate image textures and inter-channel relationships.
  • Visual and Numerical Superiority: The method not only produces visually pleasing denoised images but also achieves superior quantitative results compared to established low-rank matrix and tensor denoising techniques.

Key Results

The proposed TT-Mac algorithm demonstrates significant improvements over existing methods like RegL1-ALM, HaLRTC, and PARAFAC.

  • Quantitative Performance: In Table 2, the TT-Mac method consistently achieves the highest Peak Signal-to-Noise Ratio (PSNR) and the lowest Relative Square Error (RSE) across all tested images (Lena, Peppers, Starfish, Pine). For instance, on the "Lena" image, TT-Mac achieves a PSNR of 18.0842 and an RSE of 0.1188, outperforming the next best method (PARAFAC) by a notable margin.
  • Qualitative Performance: Figure 3 visually confirms the effectiveness of TT-Mac. The denoised images produced by the proposed method are cleaner and retain more fine details compared to the outputs of other algorithms, especially in the presence of significant Gaussian noise.

Limitations and Future Work

  • Limitations: The paper acknowledges that the computation of matrix rank is NP-hard, which is a general challenge for low-rank tensor methods. While TT decomposition mitigates some of these issues, the complexity of tensor operations can still be a factor. The paper also mentions that the initial parameters (weighting parameters and TT ranks) are assumed to be balanced, which might not always be optimal.
  • Future Work: The authors plan to explore adaptive mode weights to further optimize the denoising process. They also intend to extend their TT-based approach to other image processing tasks, such as classification, tensor completion, and dimension reduction.

Summary updated

Abstract

Tensor has been widely used in computer vision due to its ability to maintain spatial structure information. Owning to the well-balanced unfolding matrices, the recently proposed tensor train (TT) decomposition can make full use of information from tensors. Thereby, tensor train representation has a better performance in many fields compared to traditional methods of tensor decomposition. Inspired by the success of tensor train, in this paper, we firstly apply lowrank tensor train to recovering noisy color images. Meanwhile, we propose a novel algorithm for noise-contaminated images based on the block coordinate descent (BCD) method. The numerical experiments demonstrate that our algorithm can give a better result in the real color image both visually and numerically.
BibTeX
@misc{zhang2019color,
  author = {Yang Zhang and Zhi Han and Yandong Tang},
  title = {{Color image denoising based on low-rank tensor train}},
  year = {2019},
  doi = {10.1117/12.2524189},
}
Figure