Fast subspace-based tensor data filtering.

Abstract

Subspace-based methods rely on dominant element selection from second order statistics. They have been extended to tensor processing, in particular to tensor data filtering. For this, the processed tensor is flattened along each mode successively, and singular value decomposition of the flattened matrix is classically performed. Data projection on the dominant singular vectors results in noise reduction. The numerical cost of SVD is elevated. Now, tensor processing methods include an ALS (Alternating Least Squares) loop, which implies that a large number of SVDs are performed. Fixed point algorithm estimates an a priori fixed number of singular vectors from a matrix. In this paper, we generalize fixed point algorithm as a higher-order fixed point algorithm to the estimation of only the required dominant singular vectors in a tensor processing framework. We compare the proposed method in terms of denoising quality and speed through an application to color image and hyperspectral image denoising.