Azam, S., Gavrilova, M.
Abstract:
Images have rarely been used for psychological behavior analysis or for person identification in the information technology domain of research. In this paper, we present one of the first methods that allows to accurately predict gender from a collection of person’s favorite images. We select 56 image aesthetic features, and propose a mixture of expert models consisting of support vector machine, K-nearest neighbor and Decision tree. Final decision is taken based on the weighted combination of probability generated by individual classifiers. We introduce a genetic algorithm based method to improve the prediction accuracy of the model, which allows us to find the best combination of feature subset in 56D binary search space. Moreover, feature dimension is reduced significantly that decrease the testing time. Finally, three weights of the prediction model are adjusted using genetic algorithm in 3D real-number search space. Experimental results conducted on a true image database of 24000 images provided by 120 Flickr users. The experimental results demonstrate superiority of the proposed method over other approaches for gender prediction from perceptual image aesthetics preferences.