Full Quaternion Matrix-Based Multiscale Principal Component Analysis Network for Facial Expression Recognition

To acquire a more discriminative feature of facial expression, we propose a multi-scale principal component analysis network based on full quaternion matrix representation. Firstly, the structure feature and color components of facial image constitute a full quaternion matrix. Subsequently, two-staged quaternion principal component analysis is employed to learn convolutional filters. Among them, the feature maps of both stages are activated via nonlinear function. With binarization and coding, the local histograms are stacked together and fed to the classifier for expression matching. Experiments conducted on the RafD, MMI, NVIE, and KDEF datasets have demonstrated that the proposed method achieves higher recognition accuracy than several existing algorithms.