The USC Andrew and Erna Viterbi School of Engineering USC Signal and Image Processing Institute USC Ming Hsieh Department of Electrical Engineering University of Southern California

Technical Report USC-SIPI-318

“Adaptive Methods and Rate-Distortion Optimization Techniques for Efficient Source Coding”

by Youngjun Yoo

May 1998

In this thesis we study various techniques for efficient quantization and entropy coding in several source coding applications. In particular we discuss adaptive methods and rate-distortion (R-D) optimization techniques as our vehicles to carry out the goal of efficient source coding.

We use adaptive methods to help compression of inputs such as images and video that have little known or non-stationary source statistics. We investigate new R-D optimization techniques to seek fast and optimal solutions in lossy source coding. We also consider the error robustness issue associated with entropy coding and develop techniques for robust image coding.

The specific topics of interest in this thesis are:

× Adaptive vector quantization (VQ) without side information. We employ non-parametric pdf estimation for adaptation of the scalar-vector quantizer and the trellis coded quantizer. The resulting adaptive VQ algorithms are moderate in complexity while achieving significant adaptation gain.

× Adaptive quantization of image subband data. Context-based subband data classification and parametric pdf estimation are used for adaptive image subband quantization. We also employ a forward adaptive mode to obtain robust adaptation performance. The resulting subband image coder is competitive or superior to other state-of-the-art wavelet image coders.

× Image domain coding of simple images. We develop an image domain compression algorithm for images consisting of a few pixel intensity values. We observe that bit-plane coding with an adaptive binary arithmetic coder is effective for these simple images. We introduce a new context modeling technique to enhance the adaptive arithmetic coder efficiency.

× Fast R-D optimization algorithms. We propose a novel hybrid R-D optimization technique by combining the two most popular techniques, Lagrangian optimization and dynamic programming. The proposed technique is endowed with both the speed of the Lagrangian approach and the optimality of dynamic programming.

× Error robust compression of still images. We develop a JPEG-based image coder that transmits the number of coded bits for each group of DCT blocks as resynchronization side information. We show this approach to be useful compared to the resynchronization scheme found in baseline JPEG.

To download the report in PDF format click here: USC-SIPI-318.pdf (24.1Mb)