An Image Noise Level Estimation Based on Tensor T-Product

TL;DR

This paper introduces a novel tensor T-product based method for estimating noise levels in color images that preserves tensor data structure and improves estimation accuracy through a new sliding block approach and eigenvalue analysis.

Contribution

The paper proposes a new tensor-based noise estimation method that maintains data structure integrity and enhances accuracy compared to existing algorithms.

Findings

High estimation accuracy demonstrated in numerical experiments

Tensor eigenvalues are related to noise levels in color images

Method outperforms traditional noise estimation algorithms

Abstract

Currently, the noise level of color images is estimated by many algorithms through separate selection of each page of the third-order tensor using sliding blocks of size ${M_1} \times {M_1}$. The data structure of the tensor is disrupted by this method, leading to errors in the estimation results. In order not to disrupt the data structure of the tensor, we directly select the tensor using a sliding block of size ${M_1} \times {M_1} \times 3$ and then re-arrange it. The newly obtained tensor is decomposed into a block diagonal matrix form through T-product. It is demonstrated that the eigenvalues of this matrix are related to the noise level of the color image. Then train the relationship coefficients through learning methods, thereby obtaining the estimated noise level. The effectiveness of the algorithm was verified through numerical experiments, and it also achieved high estimation accuracy.