PhD Defense: "Common Information and Hidden State Dependencies in Layered Source Coding"

Mehdi Salehifar

July 12th (Wednesday), 10:00am
Harold Frank Hall (HFH), Rm 4164

This talk is concerned with approaching optimality in source coding strategies to exploit common information and hidden state dependencies.

The first part of the talk is about optimal coding of hidden Markov sources (HMS), which represent a broad class of practical sources obtained through noisy acquisition processes, beside their explicit modeling use in speech processing and recognition, image understanding and sensor networks. A new fundamental source coding approach for HMS is proposed, based on tracking an estimate of the state probability distribution, and is shown to be optimal. Practical encoder and decoder schemes that leverage the main concepts are introduced. An iterative approach is developed for optimizing the system. It also focuses on a significant extension of the optimal HMS quantization paradigm. It proposes a new approach for scalable coding of HMS which accounts for all the available information while coding a given layer. Simulation results confirm that these approaches significantly reduce the reconstructed distortion and substantially outperform existing techniques.

The second part of this talk is about a layered coding framework with a relaxed hierarchical structure. Advances in wired/wireless communication and consumer electronic devices have created a requirement for serving the same content at different quality levels. The key challenge is to optimally encode all the required quality levels with efficient usage of storage and networking resources. The approach to store and transmit independent copies for every required quality level is highly wasteful in resources. Alternatively, conventional scalable coding has inherent loss due to its structure. This work studies a layered coding framework with a relaxed hierarchical structure to transmit information common to different quality levels along with individual bit streams for each quality level. The flexibility of sharing only a properly selected subset of information from a lower quality level with the higher quality level, enables achieving operating points between conventional scalable coding and independent coding, to control the layered coding penalty. Jointly designing common and individual layers’ coders overcomes the limitations of conventional scalable coding and non-scalable coding, by providing the flexibility of transmitting common and individual bit-streams for different quality levels. It extracts the common information between different quality levels with negligible performance penalty. Simulation results for practically important sources, confirm the superiority of the work.

About Mehdi Salehifar:

XXXXX Mehdi Salehifar is a Ph.D. candidate in the ECE department at UCSB. He received his B.Sc. in Electrical and Computer Engineering in 2012 from the Tehran University, and his M.Sc in 2014 from the University of California, Santa Barbara. He joined the research group of Professor Kenneth Rose in 2012. His research interests include signal compression, general quantization theory and information theory.

Hosted by: Professor Kenneth Rose, Signal Compression Lab