Vector quantisation for wavelet based image compression

Abstract

Summary form only given. The present work arose from a need to transmit architectural line drawings over relatively slow communication links, such as telephone circuits. The images are mostly large line drawings, but with some shading. The application required good compression, incremental transmission, and excellent reproduction of sharp lines and fine detail such as text. The final system uses an initial wavelet transform stage (actually using a wave-packet transform), an adaptive vector quantiser stage, and a final post-compression stage. This paper emphasises the vector quantiser. Incremental transmission makes it desirable to use only actual data vectors in the database. The standard Linde Buzo Gray (LBG) algorithm was slow, taking 30-60 minutes for a training set, tended to use 'near-zero' vectors instead of 'true-zero' vectors introducing undesirable texture into the reconstructed image, and the quality could not be guaranteed with some images producing; artifacts at even low compression rates. The final vector quantiser uses new techniques with LRU maintenance of the database, updating for 'exact matches' to an existing vector and for 'near matches', using a combination of mean-square error and magnitude error. A conventional counting LRU mechanism is used, with different aging parameters for the two types of LRU update. The new vector quantiser requires about 10 seconds per image (compared with 30-60 minutes for LBG) and essentially eliminates the undesirable compression artifacts.