Traditional Softmax loss algorithm has only separability for features algorithm. This study proposed an improved Softmax loss algorithm to recognize facial features. The algorithm first applies an intra class cosine similarity loss between the features and weight vectors based on the Softmax loss feature distribution, making the intra class more compact and separating the classes as much as possible; then, on the basis of Softmax loss, we use normalized features to better simulate low-quality facial images, and reduce category imbalance by normalizing weights to ensure consistency with the cosine similarity measurement during testing; finally, the joint normalization of Softmax loss and intra class cosine similarity loss were fine-tuned on the pre trained model. The algorithm achieved recognition rates of >98% and >93% on the face recognition benchmark test sets LFW (labeled faces in the wild) and YTF (YouTube face database), respectively. The experimental results showed that in large-scale face recognition, the algorithm improved the discriminability of features, Enhanced generalization ability of the model can effectively improve facial recognition rate.