A study of large-scale ethnicity estimation with gender and age variations

In this paper we study large-scale ethnicity estimation under variations of age and gender. The biologically-inspired features are applied to ethnicity classification for the first time. Through a large number of experiments on a large database with more than 21,000 face images, we systematically study the effect of gender and age variations on ethnicity estimation. Our finding is that ethnicity classification can have high accuracy in most cases, but an interesting phenomenon is observed that the ethnic classification accuracies could be reduced by 6~8% in average when female faces are used for training while males for testing. The study results provide a guide for face processing on a multi-ethnic database, e.g., image collection from the Internet, and may inspire further psychological studies on ethnic grouping with gender and age variations. We also apply the methods to the whole MORPH-II database with more than 55,000 face images for ethnicity classification of five races. It is the first time that ethnicity estimation is performed on so large a database.