Research On Face Recognition Algorithm Based On Deep Learning

Face recognition technology is one of the research hotspots in the field of machine vision and pattern recognition.It has wide application prospect in many fields such as national defense,public security,customs,transportation,finance and medical treatment.The core of face recognition technology is the facial feature extraction technology.However,for a long time,the traditional facial feature definition method can not solve the influence of face pose,age,facial expression and occlusion,making the face recognition technology develop slowly.In recent years,with the rapid development of artificial intelligence,deep learning technology can extract the high-level,abstract and essential features of facial features more efficiently and effectively solve many difficulties that traditional technologies face,and becomes the mainstream of machine vision and pattern recognition research.Face recognition algorithm based on deep learning is still in the development stage,and its main problems are as follows:1)Ways to design a more efficient deep network model from the depth of learning model and the internal structure of the module to extract higher-level,more abstract more essential features of the human face;2)The design of the loss function of the deep learning model is relatively simple,while the design of the loss function to meet the general miscellaneous complexity will involve complex neural network training process.The unreasonable training data generation process will lead to the neural network model loss value hard to be reduced and the convergence speed is very slow and the training is not enough;3)The essence of face recognition algorithm based on deep learning is the process of iterative optimization and approximation,which requires a lot of training and verification of data,lack of face data sets,making the algorithm research and performance analysis subject to greater restrictions.In this paper,aiming at the current research status and existing problems of face recognition,this paper attempts to propose a set of efficient and versatile recognition algorithms.The main contributions include:1)Deeply analyze of the Inception-ResNet network in the face recognition with the feasibility of its full connection layer parameters to design and adjust so that it can extract a higher level of facial features;2)For the facial features within and between the class differences,a hybrid loss function based on central loss,face verification loss and Softmax loss was designed and its performance in neural network training was compared and analyzed;3)When the training based on triplet loss model appeared aiming at the problem of slow convergence,a new algorithm for batch generation of online triples training data was designed;4)Aiming at the different training sets of different data sets,a method of preloading initialization,iterative training based on model parameters is designed to fully learn data distribution over different data sets and solve the problem of overlapping samples in the multi-data set,achieving comprehensive training of multiple data sets.Finally,this paper analyzes the performance of face verification performance on LFW open dataset.The experimental results show that:1)Inception-ResNet achieves 99.45%performance based on the central loss,face verification loss,and SoftMax loss-driven hybrid loss function;2)based on the triplet loss,Inception-ResNet achieves 99.41%performance for the depth model of a batch triplet training data generation algorithm.Finally,this paper summarizes the existing work results,and concludes that based on the deep learning model proposed in this paper,the effectiveness and efficiency of face recognition can be effectively improved,and the performance requirements of practical applications are basically met.In the future,the algorithm is ready to run in a real scenario and the improvement of moving model calculation.