Incorporating prior knowledge from the new person into recognition of facial expression

Inter-subject variability plays an important role in the performance of facial expression recognition. Therefore, several methods have been developed to bring the performance of a person-independent system closer to that of a person-dependent one. These techniques need different samples from a new person to increase the generalization ability. We have proposed a new approach to address this problem. It employs the person’s neutral samples as prior knowledge and a synthesis method based on the subspace learning to generate virtual expression samples. These samples have been incorporated in learning process to learn the style of the new person. We have also enriched the training data set by virtual samples created for each person in this set. Compared with previous studies, the results showed that our approach can perform the task of facial expression recognition effectively with better robustness for corrupted data.