EVALUATION OF FUZZY IMAGE QUALITY MEASURES USING A MULTIDIMENSIONAL SCALING FRAMEWORK Dietrich Vander Weken, Etienne Kerre Ghent University Fuzziness and Uncertainty Modeling Group Krijgslaan 281-S9, Ghent, Belgium Ewout Vansteenkiste, Wilfried Philips Ghent University Image Processing Group, TELIN Dept. Sint-Pietersnieuwstraat 41, Ghent, Belgium ABSTRACT In this paper we present a comparison of fuzzy instrumental image quality measures versus experimental psycho-visual data. An existing as well as a new psycho-visual experiment we recently performed at our departments were used to col- lect data. The Multi-Dimensional Scaling (MDS) frame- work was applied in order to test which of our fuzzy im- age similarity measures correlates best to the human visual perception. Based on Spearman’s Rank Order Correlation coefficient we will show the M p 6 and M h i3 measures outper- form their peers as well as the commenly used MSE and PSNR measures, in the case where image distortions are less trivial to distinguish with the bare eye. 1. INTRODUCTION In this paper we focus on instrumental image quality mea- sures that are based on similarity measures initially intro- duced to express the degree of equality between two fuzzy sets. Fuzzy similarity measures can be applied in different ways to digital images. First of all, we used the different similarity measures to construct neighborhood-based sim- ilarity measures which also incorporate homogeneity [5]. Secondly, we have illustrated how fuzzy similarity measures can be applied to image histograms [6]. In this case similar- ity measures turned out to be useful for the comparison of two different kinds of histograms. In the first place, the similarity measures were applied directly to normalized histograms of the images considered. Secondly the similarity measures were applied to normal- ized ordered histograms. These combined similarity mea- sures clearly outperform the classical image quality mea- sures, like the Root Mean Square Error, in the sense of im- age quality evaluation. In a next step we now want to confirm these results based on experimental psycho-visual data. In former re- search, a Multi-Dimensional Scaling (MDS) approach to- wards analyzing and modeling image quality variations has extra co-authors Mike Nachtegael, Vladimir Zlokolica, Joost Rombaut (a) Wanda (b) Terrace (c) Mondrian Fig. 1. Test scenes used in the psycho-visual experiment of [2]. been shown successful [1]. This approach is based on the fact that image quality is often determined by several un- derlying attributes (such as noise and blur e.g.) and uses a multidimensional geometric model to describe the mutual relationships between different perceptual attributes, as well as the relationships between these attributes and the overall image quality. The dimensionality of this model is deter- mined by the number of independently varying perceptual attributes. The output or geometrical stimulus configuration can be used to create a perceptual distance matrix for the different stimuli, which can then be compared to the distance matri- ces created by our own fuzzy similarity measures. Herefore, Spearman’s Rank Order coefficient is used. In the next sections we will first explain our psycho- visual experiment, Section 2. Then explain the fuzzy sim- ilarity measures in more detail, Section 3. We present our results in Section 4 and end with a concluding Section 5. 2. EXPERIMENTAL SETUP A psycho-visual experiment for the image quality assess- ment has been described in detail in [2] for images artifi- cially degraded by noise and blur. Three scenes (Wanda,