Discriminative deep multi-task learning for facial expression recognition

Deep multi-task learning (DMTL) is an efficient machine learning technique that has been widely utilized for facial expression recognition. However, current deep multi-task learning methods typically only consider the information of class labels, while ignoring the local information of sample spatial distribution. In this paper, we propose a discriminative DMTL (DDMTL) facial expression recognition method, which overcomes the above shortcomings by considering both the class label information and the samples’ local spatial distribution information simultaneously. We further design a siamese network to evaluate the local spatial distribution through an adaptive reweighting module, utilizing the class label information with different confidences. In addition, by taking the advantage of the provided local distribution information of samples, DDMTL is able to achieve acceptable results even if the number of training samples is small. We implement experiments on three facial expression datasets. The experimental results demonstrate that DDMTL is superior to the state-of-the-art methods.