COMPUTER VISION AND IMAGE UNDERSTANDING Vol. 71, No. 2, August, pp. 198–212, 1998 ARTICLE NO. IV980707 Progressive Content-Based Shape Compression for Retrieval of Binary Images Corinne Le Buhan Jordan, Touradj Ebrahimi, and Murat Kunt Signal Processing Laboratory, Swiss Federal Institute of Technology EPFL, CH-1015 Lausanne, Switzerland E-mail: corinne.lebuhan@epfl.ch Received July 1, 1997; accepted April 21, 1998 This paperdeals with content-based compression of binary-shape images. The proposed method is based on a polygonal approxi- mation of the shape contours. A well-known approximation algo- rithm, from computer vision applications such as shape analysis and boundary pattern matching, is adapted to achieve a progres- sive representation. The resulting various levels of shape quality are encoded, from a coarse representation forfast browsing up to a loss- less representation for final rendering. In order to perform efficient compression of the progressive shape information, discrete geomet- rical constraints inherent to the image grid quantization are ex- ploited. While the proposed scheme offers a content-based descrip- tion (shape boundary as opposed to bitmap blocks) together with a quality scalable representation, it remains comparable, in terms of compression efficiency, with state of the art shape coding methods that do not combine such functionalities. c  1998 Academic Press 1. INTRODUCTION The need for content-based image compression is increas- ing with the amount of available visual data. On the one hand, as digital images are very expensive to store and transmit un- compressed, efficient digital compression schemes are required. On the other hand, in order to keep acceptable computational complexity in image processing, browsing, and retrieval appli- cations, a content-based representation should be stored that fa- cilitates visual data handling at the decoder side. These concerns justify the need for content-based compression, which intends to trade-off compression efficiency with application-driven rep- resentations. Shape constitutes a major feature of interest in visual databases (image, graphics, and video), as it carries meaningful semantic information about the associated visual object. Although shape information is not available yet in the existing image and video compression standards, the future standards MPEG-4 [16] and MPEG-7 [39] will support arbitrarily shaped visual-objects. In this context, it is important to design shape-coding schemes that will offer not only efficient compression to decrease the storage cost, but also a content-based representation which shall facili- tate further access, indexing, manipulation, and retrieval of the compressed data. Intelligent visual data management should fulfill the follow- ing requirements: • Efficient data compression. The shape information should be stored separately from the texture/motion information, to fa- cilitate bitstream editing. A lossless, or quasi-lossless, shape compression may be required since shape distortion is visually disturbing and affects the texture/motion coding quality. • Progressive (also called scalable) representation. It is possible to distinguish between quality scalability, when an im- age is approximated with an increasing accuracy at a constant size, and spatial scalability, when the image size is simply in- creased through the progressive representation levels. In the case of natural images, quality scalability is often called signal-to- noise-ratio (SNR) scalability, like in MPEG-2 terminology [38]. A progressive representation shall allow rough browsing as well as accurate final rendering. • Semantic representation. On the one hand, the existing image and video compression standards consist of bitmap-based and block-based coding methods, which bring a compromise between high compression ratios, acceptable complexity, and content-based representation. On the other hand, visual databases require a feature-based image description, which involves the analysis of the image/video content to extract a set of mean- ingful parameters. Feature extraction benefits from the amount of work achieved in pattern recognition and computer vision during the past decades and will not be directly dealt with in this paper; the reader is referred to some major schemes re- viewed in [37]. A growing field of interest is the indexing and retrieval of compressed data, mainly texture and motion data in available standards, taking advantage of the image analy- sis/synthesis step performed at the encoding stage (such as DCT transform, motion estimation) to decrease both the additional storage and computational complexity required by a database index add-on [4]. Existing compressed images and videos can then be processed without having to perform the whole decod- ing stage (although entropy decoding remains necessary), and directly using the bitstream-embedded visual features such as frequency-domain coefficients for texture description or motion and scene change information instead of storing them as side information: the more content-based the compression method, 198 1077-3142/98 $25.00 Copyright c  1998 by Academic Press All rights of reproduction in any form reserved.