Facial Expression recognition using higher order moments on facial patches

Abstract

The discriminative features between different facial expressions are mostly constrained to some regions of a face when an expression occurs on a face. Extracting these discriminating features is one of the most important aspect of a facial expression recognition (FER) system. Though significant amount of work has been done in identifying optimum features for expression recognition, it is still a challenging part of FER systems. Identifying the most significant regions and selectively applying higher order moments on these patches of the face image is found to give good feature vectors that have high discriminative ability between the classes (expressions). Point distribution model (PDM) has been used in the proposed method for locating the facial landmarks required to extract facial patches. Higher order Zernike moment (ZM) invariants were evaluated for these facial patches which form the features to be used for classification. ZMs are rotation invariant orthogonal moments having a high degree of image representation capability, hence they can satisfactorily extract significant amounts of both local and global information contents from the facial patch. The final feature vector consists of higher order Zernike moment invariants from all facial patches. Finally, support vector machines (SVM) classifiers are trained which are used to predict the facial expressions. Experiments on the extended CK+ dataset and JAFFE dataset show the effectiveness of the proposed FER system.