Deep subspace learning for expression recognition driven by a two-phase representation classifier

Recent research has shown that the deep subspace learning (DSL) method can extract high-level features and better represent abstract semantics of data for facial expression recognition. While significant advances have been made in this area, traditional sparse representation classifiers or collaborative representation classifiers are still predominantly used for classification purposes. In this paper, we propose a two-phase representation classifier (TPRC)-driven DSL model for robust facial expression recognition. First, the DSL-based principal component analysis network is used to extract high-level features of training and query samples. Then, the proposed TPRC uses the Euclidean distance as a measure to determine the optimal training sample features (TSFs) for the query sample feature (QSF). Finally, the TPRC represents the QSF as a linear combination of all optimal TSFs and uses the representation result to perform classification. Experiments based on several benchmark datasets confirm that the proposed model exhibits highly competitive performance.