Research on Unconstrained Face Recognition Based on Deep Learning

The emergence of deep learning has greatly promoted the development of the field of face recognition. The accuracy of face recognition in real scenes is affected by many factors. Among them, the problem of multiple poses is still an external factor that is difficult to overcome in face recognition. For the identification process in the Central African people face extreme attitude with the state led to the problem of low recognition accuracy, this paper proposes a gesture of deep space based correction feature to improve the recognition accuracy of multi-pose, first proposed in 2019 to use Google's lightweight network MobileNet for attitude correction in deep space, additionally employed ResNet18 verify and compare recognition results. This paper uses the VggFace2 dataset to train the two models in an end-to-end manner, and then test them on the CFP dataset, IJB-A dataset, and LFW dataset. The results show that the two backbone models proposed in the article are not much different from ResNet50 on the CFP. The face recognition on the IJB-A dataset is around 96%. The average recognition on the public data set LFW is about 96%. From the results of the test set, the model in the article is better than other methods. In addition, MobileNetV3 has a better recognition accuracy than ResNet18, and the amount of calculation is smaller.