Toward sparse and geometry adapted video approximations

Abstract

Video signals are sequences of natural images, where images are often modeled as piecewise-smooth signals. Hence, video can be seen as a 3D piecewise-smooth signal made of piecewise-smooth regions that move through time. Based on the piecewise-smooth model and on related theoretical work on rate-distortion performance of wavelet and oracle based coding schemes, one can better analyze the appropriate coding strategies that adaptive video codecs need to implement in order to be efficient. Efficient video representations for coding purposes require the use of adaptive signal decompositions able to capture appropriately the structure and redundancy appearing in video signals. Adaptivity needs to be such that it allows for proper modeling of signals in order to represent these with the lowest possible coding cost. Video is a very structured signal with high geometric content. This includes temporal geometry (normally represented by motion information) as well as spatial geometry. Clearly, most of past and present strategies used to represent video signals do not exploit properly its spatial geometry. Similarly to the case of images, a very interesting approach seems to be the decomposition of video using large over-complete libraries of basis functions able to represent salient geometric features of the signal. In the framework of video, these features should model 2D geometric video components as well as their temporal evolution, forming spatio-temporal 3D geometric primitives. Through this PhD dissertation, different aspects on the use of adaptivity in video representation are studied looking toward exploiting both aspects of video: its piecewise nature and the geometry. The first part of this work studies the use of localized temporal adaptivity in subband video coding. This is done considering two transformation schemes used for video coding: 3D wavelet representations and motion compensated temporal filtering. A theoretical R-D analysis as well as empirical results demonstrate how temporal adaptivity improves coding performance of moving edges in 3D transform (without motion compensation) based video coding. Adaptivity allows, at the same time, to equally exploit redundancy in non-moving video areas. The analogy between motion compensated video and 1D piecewise-smooth signals is studied as well. This motivates the introduction of local length adaptivity within frame-adaptive motion compensated lifted wavelet decompositions. This allows an optimal rate-distortion performance when video motion trajectories are shorter than the transformation Group Of Pictures, or when efficient motion compensation can not be ensured. After studying temporal adaptivity, the second part of this thesis is dedicated to understand the fundamentals of how can temporal and spatial geometry be jointly exploited. This work builds on some previous results that considered the representation of spatial geometry in video (but not temporal, i.e, without motion). In order to obtain flexible and efficient (sparse) signal representations, using redundant dictionaries, the use of highly non-linear decomposition algorithms, like Matching Pursuit, is required. General signal representation using these techniques is still quite unexplored. For this reason, previous to the study of video representation, some aspects of non-linear decomposition algorithms and the efficient decomposition of images using Matching Pursuits and a geometric dictionary are investigated. A part of this investigation concerns the study on the influence of using a priori models within approximation non-linear algorithms. Dictionaries with a high internal coherence have some problems to obtain optimally sparse signal representations when used with Matching Pursuits. It is proved, theoretically and empirically, that inserting in this algorithm a priori models allows to improve the capacity to obtain sparse signal approximations, mainly when coherent dictionaries are used. Another point discussed in this preliminary study, on the use of Matching Pursuits, concerns the approach used in this work for the decompositions of video frames and images. The technique proposed in this thesis improves a previous work, where authors had to recur to sub-optimal Matching Pursuit strategies (using Genetic Algorithms), given the size of the functions library. In this work the use of full search strategies is made possible, at the same time that approximation efficiency is significantly improved and computational complexity is reduced. Finally, a priori based Matching Pursuit geometric decompositions are investigated for geometric video representations. Regularity constraints are taken into account to recover the temporal evolution of spatial geometric signal components. The results obtained for coding and multi-modal (audio-visual) signal analysis, clarify many unknowns and show to be promising, encouraging to prosecute research on the subject.