Information theoretic methods in distributed compression and visual quality assessment
MetadataShow full item record
Distributed compression and quality assessment (QA) are essential ingredients in the design and analysis of networked signal processing systems with voluminous data. Distributed source coding techniques enable the efficient utilization of available resources and are extremely important in a multitude of data intensive applications including image and video. The quality analysis of such systems is also equally important in providing benchmarks on performance leading to improved design and control. This dissertation approaches the complementary problems of distributed compression and quality assessment using information theoretic methods. While such an approach provides intuition on designing practical coding schemes for distributed compression, it directly yields image and video QA algorithms with excellent performance that can be employed in practice. This dissertation considers the information theoretic study of sophisticated problems in distributed compression including, multiterminal multiple description coding, multiterminal source coding through relays and joint source channel coding of correlated sources over wireless channels. Random and/or structured codes are developed and shown to be optimal or near optimal through novel bounds on performance. While lattices play an important role in designing near optimal codes for multiterminal source coding through relays and joint source channel coding over multiple access channels, time sharing random Gaussian codebooks is optimal for a wide range of system parameters in the multiterminal multiple description coding problem. The dissertation also addresses the challenging problem of reduced reference image and video QA. A family of novel reduced reference image and video QA algorithms are developed based on spatial and temporal entropic differences. While the QA algorithms for still images only compute spatial entropic differences, the video QA algorithms compute both spatial and temporal entropic differences and combine them in a perceptually relevant manner. These algorithms attain excellent performances in terms of correlation with human judgments of quality on large QA databases. The framework developed also enables the study of the degradation in performance of QA algorithms from full reference information to almost no information from the reference image or video.