Adaptive graph regularized transferable regression for facial expression recognition

Facial expression recognition (FER) has tremendous potential in affective computing and humancomputer interaction fields. Traditional FER algorithms usually perform the training and testing of models on a single domain. However, face images in real environments usually come from different domains, which greatly limits the performance of traditional algorithms. To handle this cross-domain recognition problem, we put forward a novel transfer learning model, called adaptive graph regularized transferable regression (AGTR), which can learn a discriminative projection matrix by embedding relaxed label regression, class sparsity structure, and adaptive graph structure into a unified framework. To be specific, in our method, we develop a relaxed label regression to learn a projection matrix. Then, we exploit a class sparsity structure in each class of samples separately to obtain a consistent subspace. Further, we design a novel adaptive graph structure, which can adaptively discover the geometric relationship between samples. Finally, we verify the advancement of our approach on four public facial expression databases.