Extension de H.264 au stéréoscopique et le multi-vues : du codage à la qualité de l’expérience

Abstract

Over the last few years, 3D video has been gaining interest with a fast pace. It can be regarded as a type of visual media that expand the capabilities of a 2D video by providing depth sensation of the observed scene. 3D video applications require Multi-View content (MVV) with a very large number of views, captured by several synchronized cameras from different viewpoints. As this huge amount of data is making its way to consumers, its transmission over existing bandwidth-limited infrastructure is a great challenge. Consequently, efficient compression is essential to store or transmit MVV streams. Based on these considerations, Multi-view Video Coding (MVC), an extension of the H.264/MPEG-4 AVC standard, has been developed to address the encoding of MVV content. MVC takes advantage of the redundancies exiting between views to achieve high coding performance. However, one major problem with MVV systems is linked to illumination and color variations between views. These color mismatches may impair both coding performance and rendering quality. To fix this issue, we propose a preprocessing method based on an improvement of the histogram matching for correcting these color discrepancies between views. Furthermore, this thesis tackles the problem of stereoscopic video (i.e., 2 views only) coding by relying on an asymmetric coding method that exploits the binocular suppression theory. For this point, based on recent findings on binocular perception, a fully automated model that dynamically determines the best level of asymmetry is proposed, thus offering optimal 3D visual experience. Finally, in 3D imaging systems, it is desirable to assess degradations and to predict the perceived quality of 3D content in line with the human visual system (HVS). Therefore, reliable quality evaluation metrics are really needed, in order to guarantee the final 3D viewing experience. To this end, this thesis presents three novel quality assessment metrics for stereoscopic images. Thanks to the integration of the perceptual binocular characteristics, these metrics handle effectively the asymmetric distortions of stereoscopic images.

In light of these developments, relying on the perceptual characteristics of the HVS, several methods have been presented in this thesis, thus covering the end-to-end 3D video system, from preprocessing to quality assessment, through coding, with the aim to provide an optimal quality of experience.