Face frontalization, as an important and commonly used means of face recognition, generating a frontal face image with original identity information from a profile face image, avoiding the problem of facial features extraction difficulty and face recognition inability at large poses. Current face frontalization models based on generative adversarial network technology usually generate frontal face images with poor retention of identity information, low image quality, and low face recognition accuracy when profile facial images under a large pose. By introducing the U-net architecture and Self-Attention mechanism into the Generative Adversarial Network, a single-pathway U-Net Self-attention Generative Adversarial Network (UNSA-GAN) model is proposed to effectively integrate features of various scales and capture more detailed information, thus to enhance the quality of synthesized frontal facial images. During training phase, idendity(id) loss and pose loss are introduced to generate more accurate frontal faces that maintain identity consistency with the original profile faces. Experimental results demonstrate the effectiveness of the proposed model.