In recent years, convolutional neural network has achieved good results in the field of face recognition, but it ignores the local features of face in feature extraction. Therefore, this paper proposes a method based on improved LBP and deep learning. In this method, the feature image extracted by ULBP operator is used to reduce the dimension of LDA data, and then combined with the original image as the input of SDResNetSt-50 network, which makes it more comprehensive to extract the features of face image. Among them, sdresnetst-50 network is a network that combines the features of the last three convolution layers of traditional ResNetSt-50 network, which enhances the ability of network feature expression; at the same time, it combines softmax with center loss function to shorten the data distance between peers, and further improves the recognition effect of face image. Experiments on CASIA-WebFace face dataset show that the proposed method achieves 98.75% recognition accuracy, which is superior to all other comparison algorithms, and proves the effectiveness and feasibility of the method.