Application of single-pass adaptive VQ to bilevel images

Abstract

Summary form only given; substantially as follows. Constantinescu and Storer (1994) introduced a new single pass adaptive vector quantization algorithm that maintains a constantly changing dictionary of variable sized rectangles by "learning" larger rectangles from smaller ones as an image is processed. For lossy compression of gray scale images, this algorithm with no advance information or training typically at least equals and often exceeds the compression obtained by the JPEG standard for a given quality. All of the authors' past work with this approach has been with lossy compression of images where pixels are 8 or more bits. The present authors provide experimental evidence that their generic single pass adaptive VQ algorithm is highly effective for bilevel images. They examine not only lossless compression, but also very high quality lossy compression as well as mixtures of lossless and lossy compression applied to scanned images that contain text, gray scale images, and line drawings. New distortion measures are introduced for high quality lossy compressed bilevel images. The authors have also experimented with an image that is a mixture of text and gray scale imagery.