Novel Geometric Key Frame & Feature Extraction for Four-dimensional Dynamic Facial Expression Recognition

الملخص

This paper presents facial expression synthesis traditionally used in Face recognition systems which leads to good performance under the variations that occurred in poses or expressions face. There are several challenges faced by the 4D FER, and they are not limited to imbalanced, high computational complexity, occlusion, large data requirements, especially for deep learning frameworks. Facial expression synthesis is one of the popular researches used in varying applications. In this paper, we initiate a model to synthesize a natural facial image from given various facial expressions while maintaining the initial facial features. The ability to produce both synthetic images of subjects in the training set which could be employed to models requiring larger training data is one of the novelties of our approach. The feature engineering techniques proposed in this paper may be adapted for real-time embedded systems due to the strategies implemented to reduce computational complexity, memory, and maintain a relatively high degree of accuracy. Also, geometric PSNR, IMMSE and Entropy information is used to detect apex or key frames, and a novel alpha axes-angular feature extracted from geometric facial landmarks data. This leads to the generation of input feature vector from ZY axes for alpha angles with respect to origin in three-dimensional Euclidean space. The NCA feature selections are applied to obtain optimal feature subset and trained using multiclass ECOC-SVM on MATLAB 2020a. The problem of 3D facial expression recognition is modeled as an imbalanced problem and average recognition accuracy are used as a performance metric. The results showed a highly informative alpha angular feature on the BU4DFE dataset and demonstrate the effectiveness of the proposed landmark-based approach to classifying emotions.