Statistics of natural image sequences: temporal motion smoothness by local phase correlations

Statistical modeling of natural image sequences is of fundamental importance to both the understanding of biological visual systems and the development of Bayesian approaches for solving a wide variety of machine vision and image processing problems. Previous methods are based on measuring spatiotemporal power spectra and by optimizing the best linear filters to achieve independent or sparse representations of the time-varying image signals. Here we propose a different approach, in which we investigate the temporal variations of local phase structures in the complex wavelet transform domain. We observe that natural image sequences exhibit strong prior of temporal motion smoothness, by which local phases of wavelet coefficients can be well predicted from their temporal neighbors. We study how such a statistical regularity is interfered with “unnatural” image distortions and demonstrate the potentials of using temporal motion smoothness measures for reduced-reference video quality assessment.