To overcome the problem that the face model is initialized by manual location in the first frame for face tracking, we propose an automatic initialization method based on deep learning. We establish a stack of sparse self-encoding with neural networks, use a large number of unlabeled samples to calculate the nodes of each hidden layer using the approximate identical method, and a back-propagation approach is used to fine-tune the weights. After pre-training the network that connects a softmax classifier, we do the supervised training using labeled samples. Thus, a classifier which can automatically initialize the first frame in face tracking is built up. The results show that the proposed method can significantly improve the efficiency of the first frame's initialization in face tracking, and the correct recognition rate can reach 92%, which basically achieves the automatic initialization of the first frame.