Facial Expression recognition via neurons partially activated discriminated ELM

Facial expression recognition is a very meaningful and challenging research topic and it has been widely used in many problems in the artificial intelligence era. Such as human computer interaction, intelligent family robot, patient intelligence diagnosis, and driver safety. In this paper, we proposed a new approach for facial recognition based on ELM(Extreme Learning Machine). Traditional ELM is famous for it simple structure and high accuracy both in classification and regression problems. However, the number of neurons is hard to determine in a given problem. The core idea is force the l2,1-norm on the weight matrix of hidden layer to select the activated neurons, thus to obtain a simple network structure. Then, we add the migration factor into loss term to enlarge the distance of different types of facial expressions. Compared to state-of-the-art method at present, our method is competitive in accuracy and robustness. The experiments of facial expression recognition are carried out on JAFFE facial expression database and getting excellent performance.