Soft Global-Local Representation Learning and Labeling for Cross-Domain Facial Expression Recognition

Cross-Domain Facial Expression Recognition (CD-FER) is concentrated on learning transferable knowledge to understand human emotion states automatically in the cross-domain scenario and has achieved much progress in the past decade. Most works extract both global and local features and adopt adversariallearning mechanism to learn domain-invariant features, which takes facial landmarks as prior knowledge and overlooks the unreliability, resulting in suboptimal performance. To solve this limitation, we propose a Soft Global-Local Representation Learning and Labeling (SGLRLL) framework for reliable and effective unsupervised facial expression domain adaptation. Specifically, instead of extracting local features with unreliable prior facial landmarks, we propose a Local Clue Searching (LCS) module to look up the local regions adaptively and extract the local features in a soft manner. We also propose a Joint Soft Pseudo Label Generator (JSPLG) based on learned Joint Gaussian distributions for assigning pseudo labels and weights to the features of unlabeled samples from the target domain, which considers the correlation between global-local learning and achieves a trade-off between the quality and quantity of pseudo labels. The cross-dataset results demonstrate that the proposed framework outperforms existing competing methods while requiring fewer parameters and enabling faster inference.