A Non-Linear Regression Model to Predict Aesthetic Ratings of On-Screen Images

It has been found that the perceived appeal (or aesthetic) of an interface plays important role in determining its usability. Predictive model of interface aesthetics can thus be useful for designer to determine and improve usability. Images being an integral part of most of the interfaces contribute significantly to the overall interface aesthetics. In this paper, we propose a computational model to predict the aesthetic quality of on-screen images. We have identified a total of twenty features, divided into two broad categories, to capture image aesthetics. In order to relate the features to aesthetics, we performed a controlled user study with eighty images and hundred participants. The images were created by us and the participants were asked to rate those on a 5-point scale as per their judgment of appeal (or beauty or aesthetics) of the images. The data were used to train and test a non-linear regression model based on a SVM classifier, as the predictor of image aesthetics, with a mean square error of 0.03. The model basically predicts the likely aesthetic rating (on a 5-point scale) for a given image, given the feature values. The proposed model along with the details of the empirical data collection and analysis are discussed in this paper.