Deformable facial fitting using active appearance model for emotion recognition

Abstract

In facial emotion recognition, facial features of the person have to be detected first. For this, active appearance models (AAMs) are used in this work. The main drawback of AAM is that it cannot generalize to unseen faces. To overcome this drawback, the training images are pre-processed before using them for model construction in this work. For automatic initialization of the model, the output of Viola–Jones face detector is used. In literature, principal component analysis (PCA) is used to capture the main variations of the training data in AAM. As the variance in PCA is fixed, they produce large models which are difficult to optimize. In this work, fisher face method is used in which PCA is first used to reduce the dimensions of the feature space so that the resulting within class scatter matrix is non-singular and then apply the linear discriminant analysis (LDA) to reduce the dimensions still further. The experimental results on unseen and seen faces show that fitting accuracy is better and takes less time to fit when compared with existing AAMs.