Research On Real-time Face Recogniton And Attribute Analysis Based On Lightweight Convolution

Driven by big data accumulation,theoretical innovation and hardware upgrading,AI ushered in a new wave of development.With the application of deep learning,computer vision tasks such as face detection,recognition and attribute learning have made breakthroughs,which have broad application prospects in public safety,mobile payment,smart retailing and other scenarios.Despite its better accuracy and robustness,the face analysis method based on convolution neural network has the disadvantages of large parameters and computational overhead.It can only run in real time on expensive GPU server.Therefore,the research on real-time face recognition and attribute analysis algorithm for personal computer and mobile device has important research and practical value.In this paper,on the basis of accelerated implementation of lightweight convolution module,efficient real-time processing algorithms and lightweight CNNs with high precision but low computational complexity has been designed and trained,according to the characteristics of face detection,recognition and attribute analysis.Specifically,in view of the inefficiency of the depth separable convolution in dnn module of OpenCV,a channel parallel method is designed.In face detection and tracking,aiming at low efficiency caused dense sampling in MTCNN,a multi-scale anchor-based sparse pyramid scheme and a multiface tracking algorithm based on ONet fast detection are proposed.In which,the dual trigger mechanism of countdown and offset check for flexible switching between fast detection and complete detection.The marked face mask is used to reduce the repeated detection.In face landmark localization,in order to solve the problem of unbalanced distribution of pose in datasets,a pose weighted regression loss is proposed and a multi-scale fusion network of feature pyramid is designed.In face recognition,aiming at separable feature in open set problem,a loss constraint based on feature metric is proposed and a lightweight network with fast downsampling and global depth separable convolution is designed.In age estimation and gender classification task,a two-in-one network based on multi-task learning is proposed to realize collaborative training and reduce computing overhead.The practical experiments show that the proposed channel parallel depth-wise convolution acceleration implementation can achieve an effective speedup of about 8 times on different number of channels.The improved face detector achieves better results on FDDB and other datasets,and improves the detection speed and localization accuracy.The fast-tracking scheme doubles the detection speed on three typical videos.Our face recognition network surpasses the best existing methods,MobileFaceNet,on LFW and MegaFace with fewer parameters.And our landmark localization model and mult-task attribute analysis model also ranks first in open benchmarks,reflecting the dual advantages of efficiency and accuracy.