A latent tensor factorization framework for non-negative convolutive models

Convolutive models emerge in various domains such as acoustics, image processing or seismic sciences. In this work, we investigate the convolutive models and the related deconvolution problems in a latent tensor factorization framework. We decrease the computational complexity of the inference scheme by utilizing the Fast Fourier Transform. We also demonstrate how this framework can be used in image deblurring and in more complex models like Non-Negative Matrix Factor Deconvolution (NMFD) model.