Shannon's celebrated separation theorem states that the task of digital communication design can be split into efficient compression of the source without regard to the channel (source coding), and reliable transmission of the compressed source symbols over noisy channels (channel coding) without considering the source characteristics. The elegant information-theoretical results of Shannon theory, however, are based on the following assumptions that may not be realistic in today's digital transmission systems:

  • Stationary sources: Multimedia sources, such as images and video sequences, are highly non-stationary. As a result, they cannot be characterized by the classical measures of rate-distortion or entropy.
  • Unbounded delay/complexity: The classical results of Shannon's theory are asymptotically valid only when the block lengths of source or channel codes tend to infinity. This implies unbounded coding delay and complexity. Large delays are not acceptable in real-time applications such as video streaming and computational complexity is prohibitive in mobile devices, where the processing power is at the premium.
  • Simple, memoryless channels: While the classical theory considers simple memoryless channels such as the binary symmetric, additive white Gaussian noise, or erasure channels, today's digital transmission system have a far more sophisticated operating conditions. Mobile and wireless system suffer from the correlated symbol loss caused by fading, and packet switched networks are vulnerable to bursts of packet loss induced by network congestion. Moreover, a wireline packet network can be augmented by a wireless network, creating a tandem channel where both packet erasures and symbol errors are present.

Our research involves addressing the above challenges in the context of multimedia content delivery. This multi-disciplinary problem has the following components:

  • Source and channel modeling: There is always a trade-off between the tractability/complexity of a model and its accuracy. As an example, a wireless image transmission system may utilize a simple Gauss-Markov model for the source together with a Gilbert-Elliot channel model. While this model provides a tractable analytical tool, it may not accurately capture the characteristics of the real system.
  • Source coding: Multimedia sources are loss-tolerant. Errors in the source can be tolerated so long as the quality of the received signal is not compromised. This is in contrast with the traditional systems in which minimizing the bit error rate is the main objective. As a result, our focus will be on lossy source coding techniques. In this context, a desirable feature of the source coder is successive refinability, or scalability, which provides a mechanism for rate-distortion optimization.
  • Channel coding: The following features of a channel code are desirable for multimedia applications. Variable rate (punctured) codes provide a trade-off mechanism for rate-distortion optimization. Rate-compatibility is required in incremental redundancy based ARQ systems. Codes with low complexity/delay are needed for real-time multimedia applications such as video streaming. Coding schemes capable of protecting against both symbol errors and packet erasures, such as product codes, are required for tandem channels. Finally, codes with known failure/error rates facilitate the rate-distortion optimization problem formulation.
  • Joint source-channel coding: At the core of a multimedia transmission system lays a trade-off between the quality of the received data and the available resources, such as the bandwidth or transmission power. This is where the characteristics of the source and channels coders, as well as the channel model are used to formulate an end-to-end optimization problem. One important aspect of this problem is the graceful degradation of the quality with the worsening channel conditions.
  • Networking: Multimedia transmission over a network introduces a host of new problems that do not apply to a point-to-point communication system. These include, but not limited to, scalable broadcasting and multicasting for a large number of users, exploiting the path diversity offered by the network, and distributed paradigms such as distributed source coding.

While most of the latter technologies have individually matured over the years, it is the integration of these elements that poses the real challenge. Our main objective is to optimize the end-to-end system, and any component-wise innovation that results in an overall gain is considered part of our research agenda.