Neutral expression modeling in feature domain for facial expression recognition

Facial expression recognition (FER) is an active pattern recognition problem in the field of computer vision. The objective of FER algorithms is to extract discriminative features from a face. From the Ekman´s theory, any expression is a result from deviation of their neutral state. So, the analysis of expressive images with respect to neutral expression could be important for facial expression recognition. However, neutral images of different subjects comprise large variability in shapes as well as in texture. Hence, alignment is a primary step to minimize shape and texture variations of neutral images of different subjects. We propose to align neutral images of different subjects in the feature domain using Procrustes analysis. Subsequently, modeling of shape-free neutral images is done using Principal Component Analysis (PCA). Projection of expressive image onto the neutral subspace helps to divide an image into two components namely neutral component and expressive component. Proposed method extracts features from both the components. Extracted features are divided into a number of blocks and subsequently, dimensionality of each block is reduced with multiple discriminant analysis (MDA). The reduced feature is used to train supervised support vector machine (SVM) classifier. Experimental results show the efficacy of the proposed approach.