PERFORMANCE ANALYSIS OF VISUALLY LOSSLESS IMAGE COMPRESSION Nikolay Ponomarenko (*), Alexander Zemlyachenko (*),Vladimir Lukin (*), Karen Egiazarian (**) and Jaakko Astola (**) (*) National Aerospace University, Kharkov, Ukraine (**) Tampere University of Technology, Tampere, Finland ABSTRACT In many applications, it is desirable to provide visually lossless (but, in fact, lossy) compression of still images. This can be done using modern visual quality metrics and iterative image compression/decompression procedure for setting a proper parameter of a coder for a given image. Performance of such a procedure is analyzed for a wide set of grayscale images and components of color images for several lossy compression techniques, both standard and advanced ones. Results are obtained for two human vision system (HVS) metrics, PSNR-HVS-M and MSSIM and they are in good agreement between each other. It is shown that a provided compression ratio (CR) considerably depends upon complexity of an image subject to lossy compression. The provided CR varies in very wide limits (from 3 to 30 for the used values of the metrics). It is also demonstrated that modern advanced HVS-adapted coders are able to produce by 1.21.6 larger CR than standards JPEG and JPEG2000 for the same visual quality. 1. INTRODUCTION Visually lossless compression is important in quite many applications such as medical imaging, remote sensing, digital photography, etc. [1, 2]. The main reasons for using visually lossless (or near-lossless) compression are the following. On one hand, lossless methods of image compression are often unable to provide a desired (large enough) CR especially if an image subject to compression is corrupted by quite intensive noise [3, 4]. On the other hand, there are practical situations where customers of compressed images insist on invisibility of distortions introduced by lossy compression because otherwise images lose their diagnostic or interpretation value. Then, there is a set of requirements to visually lossless methods of image compression. A main of them is that visually lossless compression should be guaranteed and this has to be done in an automatic manner for any compressed image irrespectively to its properties (complexity defined by texture/edge/detail structure and possible noise presence). A second requirement is to produce a larger CR if possible. To ensure invisibility of distortions, one needs to solve a threefold task: (a) to have a proper visual quality metric adequately characterizing image visual quality taking into account HVS, (b) to know a threshold value for this metric for which introduced distortions are still invisible (with high probability) and (c) to provide lossy compression with the value of the selected metric over this threshold [5]. Note that there are several visual quality metrics (indices) such as MSSIM [6], PSNR-HVS- M [7], WSNR [8] and some others that are able to characterize well enough quality of lossy compressed images [9]. Recently, it has been shown for grayscale images [5] that it is enough to provide MSSIM values larger than 0.985 or PSNR-HVS-M values larger than 40dB in order to guarantee practically invisibility of distortions in lossy compressed images. The procedures for reaching a desired (preset) value of a used quality metric exist. They are iterative and presume image compression/decompression, HVS-metric calculation, its comparison to threshold and simple logical operation at any iteration. The goals of this paper are the following. First, we analyze compression results for a wide set of grayscale images (there were only three test images considered in [5]) including images where noise in original images is visible. Second, we present data for seven coders, all based on orthogonal transforms, including standard coders and recently proposed ones specially designed to provide higher visual quality. 2. METHODOLOGY OF ANALYSIS Analysis has been carried out for 19 images including several conventional ones, their color components and several additional test images. The conventional test images (both grayscale and color components) Baboon, Peppers, Barbara as well as grayscale images Lenna, Boat, Airfield, Goldhill, and Cameraman have been widely used