Pattern Analysis & Applications (1999)2:285–291  1999 Springer-Verlag London Limited Vectorising and Feature-Based Filtering for Line-Drawing Image Compression P. Fra ¨nti 1 , E. I. Ageenko 1 and A. Kolesnikov 2 1 Department of Computer Science, University of Joensuu, Joensuu, Finland; 2 Institute of Automation and Electrometry, Russian Academy of Sciences, Novosibirsk, Russia Abstract: A three-stage method for compressing bi-level line-drawing images is proposed. In the first stage, the raster image is vectorised using a combination skeletonising and line tracing algorithm. A feature image is then reconstructed from the vector elements extracted. In the second stage, the original image is processed by a feature-based filter for removing noise in the objects out-line. This improves image quality and compression performance. In the final stage, the filtered raster image is compressed using a standard compression technique, JBIG. For a set of test images, the method achieves a compression ratio of 40:1, in comparison to 33:1 of JBIG. Keywords: Filtering; JBIG; Line-drawing images; Near-lossless compression; Preprocessing; Vectorising 1. INTRODUCTION Lossless compression of bi-level images has been well studied in the literature, and several standards already exist [1]. In JBIG the image is coded pixel-by-pixel using a context-based probability model and arithmetic coding [2]. The combination of already coded neighbouring pixels defines the context. In each context, the probability distribution of the black and white pixels is adaptively determined, and the pixel is then coded by binary arithmetic coder, namely the QM-coder [3]. JBIG achieves compression ratios from 10 to 50 for a typical A4-sized image. The pixelwise dependencies are well used, and there is not much room for improvement. Remark- able improvement has been achieved only by specialising to some known image types and exploiting global dependencies. For example, the methods in Witten et al [4] and Howard [5] include pattern matching techniques to extract symbols from text images. The compressed file consists of bitmaps of the library symbols coded by a JBIG-style compressor, the location of the extracted marks as offsets, and a pixelwise coding of the matched symbols using a two-layer context template. One way to improve compression is to preprocess the image by filtering for noise removal. Filtering reduces irregu- Received: 25 November 1998 Received in revised form: 1 April 1999 Accepted: 12 April 1999 larities in the image caused by noise, and in this way makes the image more compressible without affecting the image quality. Noise appears in the image as randomly scattered noise pixels (additive noise), and as content-dependent noise distorting the contours of printed objects (lines, characters) by making them ragged. The noise level may be low enough not to significantly detract from the subjective quality, but it introduces unnecessary details that decrease the com- pression performance. Several methods have been considered for image pre- processing by analysing the local pixel neighbourhood defined by a filtering template [6–8]. These filters (logical smoothing, variations of median filtering, isolated pixel removal and morphological filters [9]) use a set of rules to accept or reject the pixel, such as predefined masks or a quantitative description of the local neighbouring area. Recent research in mathematical morphology has shown that morphological filtering can be used as an efficient tool for pattern restoration in an environment with a lot of additive noise [10–12]. Such approaches, however, are not necessarily suitable for filtering content-dependent quantis- ation noise. Another problem is that the filtering may destroy fine image structures carrying crucial information if the amount of filtering is not controlled. We study content-based noise removal for line-drawing images such as engineering drawings, cartographic maps, architectural and urban plans and circuits (radio electrical and topological) by using global spatial dependencies. This