Cross-domain face expression recognition based on joint constrained reconstruction subspace learning
Mei Qi, Changshan Zou · 2024
Facial expression recognition is difficult to achieve good results in unlabeled cross domain situations, a cross-domain face expression recognition algorithm based on joint constraint reconstruction subspace learning is proposed. This method first measures the relationship between shallow features extracted by convolutional neural networks and uses the Gram matrix as the original feature expression of the expression image; Secondly, by projecting the original features onto the latent subspace, a dual alignment constraint of edge distribution and conditional distribution is constructed to reduce domain distribution differences; To further preserve discriminative geometric features in feature representation, a sparse data matrix is constructed in the reconstructed subspace, and the structure is constructed to maintain constraints that keep similar classes within the class away from each other; Finally, the reconstructed subspace is trained by combining the dual distribution alignment constraint and structure preservation constraint, and the transformed cross domain facial expression invariant feature expression is learned from the reconstructed subspace. It is input into the SVM classifier to complete cross domain facial expression recognition. Experiments on three publicly available facial expression datasets show that the proposed cross domain facial expression recognition method achieves an average recognition rate of 59.97%, which is 2.02% better than the suboptimal method. This proves the effectiveness of our method in cross domain facial expression recognition tasks.