MEG: Texture operators for multi-expert gender classification

In this paper we focus on gender classification from face images. Despite advances in equipment as well as methods, automatic face image processing for recognition or even just for the extraction of demographics, is still a challenging task in unrestricted scenarios. Our tests are aimed at carrying out an extensive comparison of a feature based approach with two score based ones. When directly using features, we first apply different operators to extract the corresponding feature vectors, and then stack such vectors. These are classified by a SVM-based approach. When using scores, the different operators are applied in a completely separate way, so that each of them produces the corresponding scores. Answers are then either fed to a SVM, or compared pairwise to exploit Likelihood Ratio. The testbeds used for experiments are EGA database, which presents a good balance with respect to demographic features of stored face images, and GROPUS, an increasingly popular benchmark for massive experiments. The obtained performances confirm that feature level fusion achieves an often better classification accuracy. However, it is computationally expensive. We contribute to the research on this topic in three ways: 1) we show that the proposed score level fusion approaches, though less demanding, can achieve results that are comparable to feature level fusion, or even slightly better given that we fuse a particular set of experts; the main advantage over the feature-based approach relying on chained vectors, is that it is not required to evaluate a complex multi-feature distribution and the training process: thanks to the individual training of experts the overall process is more efficient and flexible, since experts can be easily added or discarded from the final architecture; 2) we evaluate the number of uncertain/ambiguous cases, i.e., those that might cause classification errors depending on the classification thresholds used, and show that with our score level fusion these significantly decreases; despite the final rate of correct classifications, this results in a more robust system; 3) we achieve very good results with operators that are not computationally expensive.