Deep belief network has good performance on facial expressions recognition, but its initial weight matrix between the last hidden layer and the labeled layer is usually generated randomly with less discriminative ability, which leads to that the features mapped from the initial weight matrix cannot guarantee to be suitable for classification tasks. To address the problem, a new architecture of deep belief network is proposed, named linear discriminant deep belief network. First, the traditional linear discriminant analysis is improved by designing a new between scatter matrix, which addresses the rank problem of the traditional linear discriminant analysis. Then, the improved linear discriminant analysis is used to initialize the weight matrix between the last hidden layer and the labeled layer of deep belief network to make sure that the weight matrix is suitable for classification task. In the experiments, our proposed linear discriminant deep belief network obtains respectively the recognition rates of 78.26% and 94.48% on the JAFFE database and the Extended Cohn-Kanade database.

facial expression recognition; restricted Boltzmann machine; deep belief network; linear discriminant analysis