International Journal of Computer Science Issues

More than a traditional journal...

Reconstruction of image and video from sparse observations attract a great deal of interest in the filed of image/video com- pression, feature extraction and denoising. Since the color image and video data can be naturally expressed as a ten- sor structure, many methods based on tensor algebra have been studied together with promising predictive performance. However, one challenging problem in those methods is tuning parameters empirically which usually requires computational demanding cross validation or intuitive selection. In this pa- per, we introduce Bayesian Tucker decomposition to recon- struct image and video data from incomplete observation. By specifying the sparsity priors over factor matrices and core tensor, the tensor rank can be automatically inferred via vari- ational bayesian, which greatly reduce the computational cost for model selection. We conduct several experiments on im- age and video data, which shows that our method outperforms the other tensor methods in terms of completion performance.