Subject-Independent Emotion Recognition from Facial Expressions using a Gabor Feature RBF Neural Classifier Trained with Virtual Samples Generated by Concurrent Self-Organizing Maps

摘要

The most expressive way humans display emotions is through facial expressions. This paper is dedicated to the challenging computer vision task of subject-independent emotion recognition from facia! expressions. The original key idea of the proposed model is the increasing of the neural classifier training set size by adding ,,virtual" samples generated with a system of Concurrent Self-Organizing Maps (CSOM). The model consists of the following main processing cascade: (a) Gabor Wavelet Filtering (GVF); (b) dimensionality reduction using Principal Component Analysis (PCA); (c) Radial Basis Function (RBF) neural classifier trained with virtual samples generated by CSOM system (VSG-CSOM). We have evaluated the above proposed model for person-independent facial expression recognition using JAFFE database. One obtains an average recognition score for the test set (leave-one subject out test method) of 69.70%. The advantage of using CSOM-VSG-RBF over a traditional RBF neural classifier means an improvement of recognition score with about 16% (from 53.44% for RBF to 69.70% for VSG-CSOM-RBF).