Applications of gain-spectral-block classification in image coding

Hamid Jafarkhani, M. Kerry, Nariman Farvardin

Research output: Chapter in Book/Report/Conference proceedingConference contributionpeer-review

Abstract

This work focuses on the development of a two-level image block classification scheme and its application to low bit rate image coding. Using this classifier, we present two adaptive encoding structures, one based on vector quantization (VQ) and the other based on transform coding. The first stage of our system classifies the image blocks into K 1 classes based on the block grain, similar to the well-known classification scheme of Chen and Smith, but allows for the possibility of a variable number of vectors per class. To do this, we develop an iterative mini-max algorithm that adjusts the vectors among the classes so that the resulting mean-normalized standard deviation of the gain values within any class is similar to all other classes. After classifying based on block gain values, we further classify each gain-class into K 2 spectral classes. This is accomplished by performing a 1D LPC-type analysis of each block, and clustering the resulting LPC vectors using a vector quantizer (VQ) with K 2 codevectors. In order to make this spectral matching meaningful, the VQ is designed and implemented using the Itakura-Saito distortion measure. The resulting two-level classification scheme thus classifies an image into K = K 1K 2 classes. A system consisting of a bank of K fixed-rate Multi-Stage VQ's and a DCT based system are then used to examine the usefulness of the proposed approaches for classification.

Original languageEnglish
Title of host publicationProceedings of SPIE - The International Society for Optical Engineering
EditorsMajid Rabbani, Edward J. Delp, Sarah A. Rajala
Pages88-99
Number of pages12
StatePublished - 1995
EventStill-Image Compression - San Jose, CA, USA
Duration: 7 Feb 19958 Feb 1995

Publication series

NameProceedings of SPIE - The International Society for Optical Engineering
Volume2418
ISSN (Print)0277-786X

Conference

ConferenceStill-Image Compression
CitySan Jose, CA, USA
Period7/02/958/02/95

Fingerprint

Dive into the research topics of 'Applications of gain-spectral-block classification in image coding'. Together they form a unique fingerprint.

Cite this