Possible Extensions

We describe here some of the possible extensions of the work presented in this dissertation. $\bullet$

• Multimedia Quality Assessment: As multimedia quality is affected (when encoded and transmitted in real time using packet networks) by a large number of parameters, one of the directions in future research is to build a robust database, by conducting a series of MOS experiments taking into account different combinations of these parameters and to build a generic tool that may be used in different applications. A generic tool can be built to evaluate the quality of multimedia (audio + video) taking into account several codecs and parameters.
• Understanding the Parameters' Impact: Some future research directions for this work include the study of other codecs like MPEG2/4 and the analysis of the effect of audio and video synchronization. The combined effects of other parameters can be studied by following the approach presented in this dissertation. Network loss is one of the most annoying factors on multimedia quality, and there are several techniques proposed to tackle this problem. A general solution is the use of Forward Error Correction (FEC). There are also some error concealment techniques depending on the media type (for example, waveform substitution for speech or linear interpolation for video). It is interesting to study the impact of these different techniques on human perception.
• Network Technologies: Our work focused on IP networks, but it can also be applied to other network technologies including ATM and wireless networks. ATM networks can guarantee some levels of QoS (e.g. low loss and better bandwidth availability). Wireless networks technology evolves and spreads rapidly in these days, with emphasis on real-time multimedia applications, including videophone teleconferencing and video streaming. Thus, it will be interesting to further investigate the possibilities of adapting our methods to these networks in the future. This concerns the following points: the real-time multimedia quality assessment, the study of the impact of the quality-affecting parameters on the quality, the rate control protocol, and the traffic prediction methodology.
• Encoder Design: A method called Perception Based Spectrum Distortion (PBSD) is proposed in [161]. A very important issue in that work is that the authors used their metric instead of MSE to redesign the ADPCM speech encoder in order to set the quantization level. This led to an improvement in the subjective quality over the normal ADPCM codec. Similarly, there is ongoing research in video encoding, focusing on finding other metrics that give better results than PSNR to be used in encoding design. A preliminary result [98] shows good improvements. But as we have shown, most of the available objective measures have several limitations (mainly, their high computational cost and the need to access the original signal). Thus, they cannot be used in real time to improve the encoding based on human perception. We believe that our method is a good candidate to do such work. In addition, it can take into account the impact of network distortion.
• Rate Control: We used, with our rate control, the equation-based TCP-Friendly congestion control protocol that has some good properties for real-time multimedia applications. It is important to investigate the use of other protocols jointly with ours. We provided a list of the possible controlling parameters to be used with our proposal. By using these parameters, we believe that the quality can be improved. The effect of the other parameters and building a complete controller are some possible future research directions.
• Traffic Prediction Model: Based on this model, many applications can be implemented. Dynamic bandwidth allocation, dynamic long-term contract negotiation, pricing, traffic shaping and engineering, are some examples.
• Random Neural Network Training: It is known that Levenberg-Marquardt methods can converge to local minima in some cases depending on the initialization of the weights. One simple solution to this problem is to restart the training process whenever a local minimum convergence has taken place, but by initializing the weights differently. As shown by our results, when the proposed algorithms converge to the global minimum, they converge very fast. For example, it takes only one or two iterations to reach 0 error in the XOR problem. More sophisticated solutions can be studied in the future, for instance combining the Levenberg-Marquardt methods with the ``simulating annealing'' technique which is based on a well-studied mathematical model. Another point that deserves attention in future research work is the study of the best ranges of the learning parameters (namely, $\mu, \beta, \zeta$ and $dP$).