Face Recognition Algorithm Based on CNN-Transformer

The era of intelligence has quietly arrived, high-performance computers have emerged, face recognition technology has really entered the stage of automatic recognition from semi-automatic recognition, and has been rapidly developed as an important application field of artificial intelligence. In view of the problem of receptive field when CNN is used as encoder alone, the modeling ability of the network for global information is weak. This paper combines CNN and Transformer to make full use of the information contained in the image and improve the overall performance of the network. The research focuses on information fusion and multi-head feature enhancement, which includes multi-head self-attention in window and multi-head self-attention in shifted window. One part of multi-head feature enhancement is the fusion feature map generated in the encoder and the other part is the feature map generated by the decoder. The experiment and baseline model were compared and analyzed. Compared with the original network model, the lightweight CnnTrans model proposed in this paper has the advantages of lightweight structure, simple calculation and strong portability.