New reconstruction techniques for image/video communication

Resumen

Motivation Recent advances in computing and communication technologies boosted the bandwidth expansion and the processing capabilities of personal computers and battery-powered terminals. These advances naturally yield a rapid growth of multimedia applications. In particular, video streaming (e.g. mobile TV, video-calling, etc.) and image transmission comprise nowadays by far the largest fraction of all the consumer traffic. Thus, achieving high quality of service (QoS) for these signals is of utmost importance and it is a challenging task since multimedia streams are usually transmitted over error-prone channels such as the internet. In many common multimedia applications the retransmission of the lost data is not possible due to real-time constraints (video-calling, sport events retransmissions, etc.) or lack of bandwidth. Thus, error concealment (EC) techniques are required and even mandatory when packet losses occur [1]. Objective and Methodology In this thesis we have studied the possibility of improving the reconstruction quality provided by the state-of-the-art techniques [2,3,4]. We have designed and implemented several reconstruction algorithms applying different points of view. We have focused our attention on four main issues: 1) EC techniques based on spatial interpolation. 2) Algorithms that pursue the reconstruction by gathering data statistics from known surrounding samples. 3) EC techniques carried out in transformed domain. 4) Improving the filling order so the error propagation is reduced. An exhaustive comparison with state-of-the-art algorithms is provided in order to asses the quality of our techniques. In the simulations, we have considered different loss scenarios and tested the performance over a large variety of images and video sequences. Proposed techniques In this thesis we tackle the four topics detailed above. First, we propose an EC technique based on the concept of visual clearness of an edge. We explore the directional behaviour in the neighbourhood of the missing area applying a novel scanning procedure. Second, since natural images can be locally non-stationary, we design a sparse linear predictor that dynamically adapts itself to the amount of useful and available data. We then generalize this predictor by adopting a multivariate kernel-based MMSE estimation framework. Third, natural images tend to be low-pass signals. This a priori knowledge could be exploited in order to improve both reconstruction quality and robustness against overfitting. We propose a frequency weighting (filtering) to exploit this a priori knowledge. A special low-pass filter is designed for such a purpose. Finally, we propose a novel filling order approach that, exploiting the reconstruction error, improves the quality reconstruction. Regions that yield better reconstructions will be prioritized in order to reduce error propagation and achieve better overall reconstruction quality. Conclusion The experimental results show that our proposals achieve better PSNR and perceptual reconstruction quality than other state-of-the-art techniques. In many cases, improvements superior to 1dB are obtained with respect to modern EC algorithms. Bibliography [1] ITU-T, ¿ITU-T Recommendation H.264,¿ International Telecommunication Union, 2010. [2] A.M. Peinado, V. Sanchez, and A. Gomez, ¿Error concealment based on MMSE estimation for multimedia wireless and IP applications,¿ in Proceedings of PIMRC (invited paper), September 2008, pp. 1¿5. [3] D. Persson, T. Eriksson, and P. Hedelin, ¿Packet video error concealment with gaussian mixture models,¿ IEEE Transactions on Image Processing, vol. 17, pp. 145¿154, 2008. [4] J. Seiler and A. Kaup, ¿Complex-valued frequency selective extrapolation for fast image and video signal extrapolation,¿ IEEE Signal Processing Letters, vol. 17, pp. 949¿952, November 2010.