טכניון מכון טכנולוגי לישראל
הטכניון מכון טכנולוגי לישראל - בית הספר ללימודי מוסמכים  
M.Sc Thesis
M.Sc StudentPorat Dror
SubjectContext-Based Multiple Description Wavelet Image Coding
DepartmentDepartment of Electrical Engineering
Supervisor Professor Emeritus David Malah
Full Thesis textFull thesis text - English Version


Abstract

Multiple description (MD) coding is a coding technique that produces several descriptions of a single source of information (e.g., an image), such that various reconstruction qualities are obtained from different subsets of the descriptions. The purpose of MD coding is to provide error resilience to information transmitted on lossy networks (i.e., networks that cannot avoid possible loss of packets). Since there is no hierarchy of descriptions in MD coding, representations of this type make all of the received descriptions useful (unlike, for example, layered coding, where a lost layer may also render other enhancement layers useless). Thus, MD coding is especially suitable for networks with no priority mechanisms for data delivery, such as the Internet.

Among previous works, MDs for image coding were generated via the utilization of a decomposition into polyphase-like components (a polyphase transform) and selective quantization, performed in the wavelet domain. In this research work, we present an effective way to exploit the special statistical properties of the wavelet decomposition to provide improved coding efficiency, in the same general framework.

We propose a novel coding scheme that efficiently utilizes contextual information, extracted from a different polyphase component, to improve the coding efficiency of each redundant component (aimed to provide an acceptable reconstruction of a lost polyphase component in the case of a channel failure), and thus enables the proposed MD coder to achieve improved overall performance. This is accomplished by means of various coding procedures, such as context-based classification of the wavelet coefficients, parametric model-based adaptive quantization, efficient optimal bit allocation (performed in the general framework of Lagrangian optimization), and adaptive entropy coding.

Our experimental results clearly demonstrate the advantages of the proposed context-based MD image coder. Specifically, we also show that the proposed coder outperforms its predecessor-the original polyphase transform-based coder-across the entire redundancy range, and that the improvement in coding efficiency can indeed be attributed primarily to the effective utilization of contextual information.