Enhanced Facial Expression Recognition Through Deep Subspace-Based Joint Multi-Class Dictionaries

Facial expression recognition (FER) has experienced substantial progress in theoretical development and real-world application. However, some existing methods are still facing challenges associated with nonlinear distribution of limited samples and weak discriminability. To tackle the challenges, an innovative method for FER that combines deep subspace feature representation with joint multi-dictionary learning is proposed. The method consists primarily of two stages: first, deep subspace feature representation is used to construct multi-class filtered feature dictionaries, with the aim of enhancing the representation capability of expression features. Secondly, the diversity among the atoms of multi-class feature dictionaries is jointly learned, aimed at improving the discriminative performance of the learned dictionary. Our experimental results, which were obtained from three well-known datasets (JAFFE, CK+, and KDEF), show that our proposed method achieves superior performance over several existing advanced techniques in small-scale expression sample recognition.