A Study Of Face Recognition System Based On Deep Learning

Face recognition technology,as an important research direction in the field of artificial intelligence,has attracted more and more attention from researchers.Face recognition technology based on deep learning has further developed face recognition from the laboratory research stage into a daily use phase.However,in daily use,face recognition systems still have many challenges,such as light intensity,facial expressions,face occlusion and other factors,and still need better solutions.At the same time,in actual use,face detection and face recognition models all need to consider both speed and accuracy.Finally,in today's society where data is exploding,how to truly implement a face recognition system that supports a considerable amount of data on a terminal with relatively weak computing power is the ultimate challenge facing every researcher.Based on the above issues,the following research contents have been carried out in this dissertation:1)Self-made data set: Through the related cooperation projects,self-built Asian ethnic data set.This data set has a larger age range,more complete daily makeup,and richer facial expressions.Then use the method based on generative adversarial network to amplify the data set,in order to reach a usage level that meets the standards of deep learning.2)Face detection: In the classic face detection method MTCNN,the core idea of the mobilenet network is incorporated to improve the corresponding network structure in MTCNN.While ensuring the original MTCNN algorithm detection effect,by reducing network parameters,the detection speed is improved.3)Face recognition: Analyze the current mainstream face recognition loss function,and design a semi-supervised loss function based on insightface.In the network training process,unlabeled data is added to easily expand the training sample,giving the model a better generalization ability.4)System design: Design and construct a face recognition system to verify the effectiveness of face detection and face recognition algorithms in actual scenes.At the same time,the system optimizes the bottom layer of the entire algorithm,making the system more streamlined and easier to deploy.