Dynamic multi-channel metric network for joint pose-aware and identity-invariant facial expression recognition

Facial expression recognition (FER) is challenging because the appearance of an expression varies significantly depending on head pose and inter-subject characteristics. With existing techniques, it is often difficult to learn both pose-aware and identity-invariant representations of facial expressions effectively due to the complex distribution of intra-class variation and similarity caused by these two factors. In this study, we propose a dynamic multi-channel metric learning network for pose-aware and identity-invariant FER, called DML-Net, which can reduce the effects of pose and identity for robust FER performance. Specifically, DML-Net uses three parallel multi-channel convolutional networks to learn fused global and local features from different facial regions. Then it uses joint embedded feature learning to explore identity-invariant and pose-aware expression representations from fused region-based features in an embedding space. DML-Net is end-to-end trainable by minimizing deep multiple metric losses, FER loss, and pose estimation loss with dynamically learned loss weights, thereby suppressing overfitting and significantly improving recognition. We evaluate DML-Net on three widely-used multi-view facial expression datasets, namely, KDEF, BU-3DFE, and Multi-PIE, as well as a wild dataset SFEW2.0. Extensive experiments demonstrate that our approach outperforms several other popular methods with accuracies of 88.2% on KDEF, 83.5% on BU-3DFE, 93.5% on Multi-PIE, and 54.36% on SFEW.