Self-adaptive feature learning based on a priori knowledge for facial expression recognition

Conventional feature extraction methods generally focus on extracting global and local features from the original data or converting a high dimensional space to a lower dimensional one. However, they tend to overlook the discriminative information of pixel values hidden in the original data. Pixel values in some local parts of a face, such as the mouth, eyebrows and eyes, provide extremely useful information for expression recognition, as they reveal the correlation between these local parts. While this information can be learned manually, being able to automatically identify important location information in this context is highly desirable. Given this, we propose a self-adaptive feature learning approach based on a priori knowledge for facial expression recognition in this paper. The proposed approach aims to adaptively select active features. It first generates an intra-class, low-rank dictionary that can decouple the original space from the expression subspace and mitigate the dependence on individual facial identities. Next, the active feature dictionary is formed, taking both global and local importance into account simultaneously. After that, the principal component of the active feature dictionary is extracted to address the influence of redundant features and reduce the dimension. We also introduce an active feature learning model as the final classification framework to make the features more discriminative and reduce the computation time. Results of comprehensive experiments on public facial expression datasets confirm the efficacy of the proposed approach, in terms of accuracy and computation time, compared to some state-of-the-art feature extraction and dictionary learning methods.