Tensor completion throughmultiple Kronecker product decomposition

We propose a novel decomposition approach to impute missing values in tensor data. The method uses smaller scale multiway patches to model the whole data or a small volume encompassing the observed missing entries. Simulations on color images show that our method can recover color images using only 5-10% of pixels, and outperforms other available tensor completion methods.