Transferable discriminative non-negative matrix factorization for cross-database facial expression recognition

Abstract

Traditional facial expression recognition (FER) systems often have an ideal assumption that the training and testing facial images are collected from the same or similar databases. However, it cannot always be satisfied in real applications, and the recognition accuracy would decrease greatly. To solve this shortcoming, in this paper, we present a novel transferable discriminative non-negative matrix factorization (TDNMF) approach for cross-database FER. In TDNMF, we use the similarity and dissimilarity information obtained from labels to guide matrix factorization and feature transfer learning. Meanwhile, we devise a graph regularization, which considers the inter- and intra-database similarity, to further mitigate the distribution shift across databases. At last, we conduct extensive experiments on several benchmarks for evaluation. The results demonstrate the effectiveness of the proposed method in comparison with some state-of-the-art transfer learning algorithms.