Research And Application Of Face Recognition Algorithms Based On Multitask Deep Learning

Face recognition is an important subject in the field of artificial intelligence and biometric recognition.It is a very successful application of image analysis and understanding.It has become a dynamic research field because of its wider and wider application in many fields such as business,security,identity authentication and so on.With the continuous improvement of application requirements,face recognition technology is no longer only aimed at face identity recognition,but also face attribute recognition is becoming more and more important.The existing technologies usually use multiple models in parallel to accomplish the recognition task of different face attributes,which not only has a slow training time,but also has the problems of large consumption of system resources and slow running time in practical use.In order to solve the above problems,this paper studies the face recognition algorithm based on Multitask deep learning.Firstly,the related theories of multitask deep learning and face recognition algorithm are introduced.Then a face recognition algorithm based on multitask deep learning is proposed:a simplified multitask face recognition model is designed to speed up the operation;a suitable multitask training sequence is selected,and a new loss function is designed for the training process of face age recognition with slow convergence of parameters,which speeds up the training speed;and improving the accuracy of model recognition by utilizing the correlation among multiple learning tasks.Based on the algorithm proposed in this paper,this paper designs a deep network structure using deep learning framework Caffe,trains multi-task face recognition model with face training data.After model selection,an end-to-end multitask face recognition modelM₄₄ is finally obtained.A single model can complete multiple recognition tasks,including:face identity recognition,face age recognition,face gender recognition and fatigue state recognition.In the experimental environment of this paper,the total time of model training is only about 10hours,and it only takes 42ms to complete four recognition tasks.Compared with other face recognition models,the training speed and running speed have increased by more than 10times,and the recognition accuracy of each task has reached or exceeded the advanced level in the field.In order to further improve the recognition accuracy of each task,this paper combines the principle of add-feature fusion with the principle of concat-feature fusion,and proposes a multitask face recognition algorithm based on multi-layer feature fusion:extracting features from the bottom,middle and upper levels of the modelM₄₄,designing a multi-layer feature fusion network to fuse the three features based on concat-feature fusion principle,and then selecting and combining features based on add-feature fusion principle.This new method can get more generalized features,which can further improve the performance of multitask learning.Based on this algorithm,an end-to-end multi-layer feature fusion network model MFN is trained,and this single model can also accomplish multiple recognition tasks.In the experimental environment of this paper,the overall running time of the model MFN is only21ms longer than the modelM₄₄,which further improves the recognition rate of each task.The multi-task face recognition algorithm designed and implemented in this paper can quickly recognize face identity and face attribute information in a short time.It has the characteristics of short training time,high speed block and high recognition accuracy.It can be widely used in many fields such as intelligent transportation,intelligent city and so on.