Facial expression recognition using HessianMKL based multiclass-SVM

Multikernel learning (MKL) has recently received great attention in the field of computer vision and pattern recognition. The idea behind MKL is to optimally combine and utilize multiple kernels and features instead of using a single kernel in learning classifiers. This paper presents a novel framework for MKL problem by expanding the HessianMKL algorithm into multiclass-SVM with one-against-one rule. Our framework learns one kernel weight vector for each binary classifier in the multiclass-SVM compared to the SimpleMKL based multiclass-SVM which jointly learns the same kernel weight vector for all binary classifiers. The proposed method is utilized to recognize six basic facial expressions and neutral expression by combining three kernel functions, RBF, Gaussian, and polynomial function and two image representations, HoG and LBPH features. Our experimental results show that our method performed better than SVM classifiers equipped with a single kernel and a single type of feature as well as the SimpleMKL based multiclass-SVM.