Face Recognition And Attribute Analysis Based On Single-model Multi-task And System Implementation

In recent years,deep neural networks have made breakthroughs in the field of computer vision,such as image classification,image segmentation,target detection and tracking.Compared to the traditional machine learning method which based on the classification of prior features,the substantial breakthrough by the deep neural network is mainly due to its multi-level nonlinear representation architecture,which brings the capable of self-learning to have high discriminating features and integration ability.As an important branch of computer vision,face recognition based on deep neural network,as well as gender,age,expression and other related attributes analysis,has important research significance and wide application value in public security,video surveillance,human-machine interaction and other fields.The face recognition and attribute analysis system based on deep neural network is composed of many functional modules,such as image acquisition,face detection,feature extraction,recognition and attribute analysis,and the extraction of discriminative features is the key to promote the accuracy of face recognition and attribute analysis.In addition to the need for a certain scale and high quality training data set,the decisive factors influence on the ability of obtaining high discriminability features are come from two aspects:the design of network architecture and loss function.Accordingly,on the basis of fully mining the multitask correlation,the interested research in this field is focus on exploring the network architecture to acquire high discriminability features quickly and efficiently,designing a loss function that can effectively constrain the difference between inter-class and intra-class.Since the deep neural network can only achieve a single task,the achievement of multi-task simultaneously will face the challenge of huge amount of computation,therefore,the exploration of using the single model to achieve multi-task at the same time,reducing the computational complexity,improving the model efficiency and utilization rate are the research challenges currently.Besides,the loss function oriented back propagation optimization mechanism enables multi-task learning to be implemented in a simpler and elegant way,so the collaborative performance of multiple related tasks,the promotion of the model's generalization ability and the effective reduction of model size are difficulties existed.Based on the above analysis,the face recognition and attribute analysis system in the practical application with abilities of high speed,high efficiency and high discriminability features is required.This thesis provided the design of deep network architecture and loss function of single-model multi-task to realize face recognition and attribute analysis.The main research and innovation of this thesis are summarized as follows:(1)A face recognition and attribute analysis algorithm with single-model multi-task is proposed,the main idea is first use the LightResNet which combined the residual network(ResNet)and the MFM(MaxFeature-Map)activation function as the single model network structure,using MFM to reduce the parameters of the network model while guaranteeing the powerful feature expression of ResNet,reduce computational complexity,improve efficiency,and provide possible real-time performance for simultaneous multi-task.Secondly,the loss function of single model is the weighted sum with constant balance factor ? of SoftmaxLoss and CenterLoss,by reducing the difference in inter-class,the discriminability of the extracted features can be increased,and the recognition rate can be improved.Finally,the fine tune multi-task training is potentially increasing the amount of data,and the collaborative performance of multiple related tasks as well as improving the generalization ability of the model.Experimental results show that the algorithm can effectively reduce the single model parameters,reduce the computational complexity and improve the utilization of single model on the premise of ensuring the effectiveness and robustness of face recognition and attribute analysis.(2)A completed real-time single-model multi-task face recognition and attribute analysis system is designed and implemented.The main function modules of the system are introduced,including video capture,face detection and alignment,feature extraction,logging,registration,and other settings,this single-model can rapidly realize many functions such as face authentication,expression recognition,age recognition,gender recognition and attitude estimation.The processing speed of the system is 31 frames per second,achieved a real-time practical system with the characteristics of fast,automatic and efficient.