Tensor Decomposition Learning for Compression of Multidimensional Signals

Multidimensional signals like multispectral images and color videos are becoming ubiquitous in modern times, constantly introducing challenges data storage transfer, therefore demanding efficient compression strategies. Such high dimensional observations can be naturally encoded as tensors, exhibiting significant redundancies across dimensions. This property is exploited by tensor decomposition techniques that being increasingly used for compactly encoding large multidimensional arrays. While efficient, these methods incapable of utilizing prior information present training data. In this paper, a novel learning method proposed the signals. Specifically, instead extracting independent bases each example, our learns an appropriate basis dimension from set samples solving constrained optimization problem. As such, sample quantized into reduced-size core coefficients corresponds to multilinear combination learned matrices. Furthermore, employs symbol dictionary binarizing outputs. Experimental results on synthetic real satellite image sequences demonstrate efficacy method, surpassing competing while offering flexibility handle arbitrary structures.