Alessandro Artusi and Alexander Wilkie Institute of Computer Graphics and Algorithms Vienna University of Technology Karlsplatz 13/186, A–1040 Vienna, Austria ABSTRACT A key problem in multimedia systems is the faithful reproduction of color. One of the main reasons why this is a complicated issue are the different color reproduction technologies used by the various devices; displays use easily modeled additive color mixing, while printers use a subtractive process, the characterization of which is much more complex than that of self–luminous displays. In order to resolve these problems several processing steps are necessary, one of which is accurate device characterization. Our study examines different learning algorithms for one particular neural network technique which already has been found to be useful in related contexts, namely radial basis function network models, and proposes a modified learning algorithm which improves the colorimetric characterization process of printers. In particular our results show that is possible to obtain good performance by using a learning algorithm that is trained on only small sets of color samples, and use it to generate a larger look–up table (LUT) through use of multiple polynomial regression or an interpolation algorithm. We deem our findings to be a good start point for further studies on learning algorithms used in conjunction with this problem. Keywords: Radial basis function networks, regression, colorimetric characterization of printing devices 1. INTRODUCTION In multimedia systems, different color reproduction devices — while serving the same purpose — exhibit large discrepancies in their raw output. This is due to the fact that they usually employ different color mixing technologies (additive and subtractive), use different color spaces and hence have different gamuts, and that their device characteristics can change with time and usage. These facts usually do not permit a faithful matching of colors between devices if no precautions are taken. Colorimetric characterization is one step in the colorimetric reproduction process that permits faithful image reproduction across different display devices. Its goal is to define a mapping function between the device–dependent color spaces in question (such as RGB or CMYK) and device–independent colour spaces (such as CIELAB or CIEXYZ), and vice versa. There are three main approches to defining this mapping function: physical models, empirical models and exhaustive measurements. Physical modeling of images devices involves building mathematical models that find a relationship between the colorimetric coordinates of the input or output image element and the signals used to drive an output device or the signals originating from an input device. The advantage of these approaches is that they are robust, typically require few colorimetric measurements in order to characterize the device, and allow for easy recharacterization if some component of the imaging system is modified. The disvantage is that the models are often quite complex to derive and can be complicated to implement. Physical models are often used for the colorimetric characterization of displays and scanners. Empirical modeling of imaging devices involves collecting a fairly large set of data and then statistically fitting a rela- tionship between device coordinates and colorimetric coordinates. Empirical models are often higher–order multidimensional polynomials, or neural network models. They require fewer measurements than LUT techniques, but they need more than physical models. Empirical models are often used for scanners and printers. Often the colorimetric characterization of printers requires an exhaustive measurement in order to obtain good performances. Typically samples of the device drive signals is sampled and colorimetrically measured. Many more measurements have to be used for devices with poor repeatability. Correspondence: Email: artusi@cg.tuwien.ac.at