Tensor Based Analysis Dictionary Learning for Color Video Denoising

In this paper, by utilizing the tensor computation, we propose a novel analysis dictionary learning (ADL) algorithm named tensor analysis dictionary learning (TADL) to effectively represent multidimensional (MD) signals and apply it for color video (four-dimensional signal) denoising. By incorporating correlations among various channels, TADL can handle an arbitrary order tensor of the MD signals more naturally. Different from what is performed in existing ADL algorithms, in TADL the dimensional features of the MD signals are represented by a series of dictionaries collaboratively, and the dictionary atoms are updated simultaneously by adopting the simultaneous codeword optimization (SimCO) on tensors, leading to a computationally efficient optimization process. Experiments show that, 1) the proposed TADL algorithm can deal with color video denoising well, which is impossible for the standard ADL algorithm, and 2) when the noise intensity is relatively high, TADL outperforms state-of-the-art ADL algorithms in terms of computational efficiency.