Comparative analysis of multiple kernel learning on learning emotion recognition

Local appearance descriptors are widely used on facial emotion recognition tasks. With these descriptors, image filters, such as Gabor wavelet or local binary patterns (LBP) are applied on the whole or specific regions of the face to extract facial appearance changes. But it is also clear that beside feature descriptor; choice of suitable learning method that integrates feature novelty is vital. The multiple kernels learning (MKL) framework reportedly shows promising performances on problems of this nature. However, most MKL studies in object recognition domain provide conflicting reports about recognition performances of MKL. We resolve such conflicts by motivating a comparative analysis of MKL using appearance descriptors for facial emotion recognition-in challenging learning setting. Moreover, we introduce a simulated learning emotion (SLE) dataset for the first time in model performance evaluation. We conclude that given sufficient training elements (examples) with efficient feature descriptor, the rapper methods of Semi-infinite programming MKL (SIP-MKL) and SimpleMKL frameworks are relatively efficient on facial emotion recognition task, compare to other kernel combination schemes. Nevertheless we opine that average MKL performance accuracy, especially on learning facial emotion dataset, remains unsatisfactory (around 56%).