A Face Recognition System On Mobile Platforms

As one field of computer vision,face analysis has received extensive research attentions,due to human faces' rich social information.As the rapid development of computer vision,face anal-ysis has also been widely used in real world applications.Recently,deep learning has achieved state-of-the-art performance in many tasks of computer vision,including face analysis,which even surpasses human performance.Compared to traditional machine learning methods,deep learning based face analysis methods have better performance in terms of recall and precision and thus are widely used in many applications.However,deep learning based methods require extremely larger computation resources,e.g.,CPU and memory than traditional methods,which limits them to be used on mobile devices.Towards fast and accurate face detection and recognition on mobile de-vices,this thesis presents a lightweight face detection and recognition framework.First,we propose several strategies to reduce the number of parameters of neural networks to speedup the computa-tion of detection and recognition,mainly including depthwise convolution,the CReLu activation function and using small size of convolution input.Then in order to achieve high accuracies,we train our model using millions of data as previous works used.Finally,it allows us to build a com-plete face analysis system on mobile devices,including face detection,tracking and recognition.The experiments show our face analysis system has acceptable overhead and good performance on TX2...