Research On Accelerated Multi Pose Face Recognition Algorithm Based On Deep Learning

In twenty-first Century,the development of computer science has made great progress.AI technology is a milestone development stage after the Internet of things,cloud computing and big data.Now,artificial intelligence technology has been applied to the language recognition,image recognition,Natural Language Processing and expert systems and other fields.Face recognition has the characteristics of openness,initiative and stability,and has wide application prospects.Therefore,more and more experts and scholars pay attention to it.However,the current face recognition techniques by video image acquisition of the weather,illumination and other environmental factors,is the collection object expression,posture,shelter and other factors,the accuracy of face recognition is greatly reduced,the distance is not controlled,open,non matching application there is a great distance from the scene.However,the current face recognition technology is affected by many factors such as weather,illumination and other environmental factors,such as facial expression,posture,shadow and so on,and the accuracy of face recognition is greatly reduced.There is still a lot of distance from the open,uncontrollable,non cooperative practical application scenarios.Pose change has nonlinear characteristics.Traditional pattern recognition model or shallow neural network model can not solve complex and nonlinear classification problems,which is a difficult problem of face recognition.Deep learning can extract features from layer by layer,and the process of abstraction and conceptualization is very similar to the working mechanism of human brain,and this similarity also makes deep learning have special advantages in solving nonlinear problems caused by pose changes.In order to identify the model of face recognition accuracy,more and more factors to be taken into consideration,more and more network layers,neural network weight bias increasingly large number of computing,more and more,which makes the intelligent algorithm based on deep learning is difficult to be applied to computing,storage,power sensitive mobile phone and other mobile terminal.This paper takes the face recognition model as an example to study the low power,high performance,low storage depth learning algorithm for mobile terminals.Therefore,the research work of this paper mainly includes the following points:1)This paper proposes an improved face recognition model which contains two independent convolutional neural networks.The convolution kernels of two neural networks are different in size to extract facial features of different granularity.In addition,in order to bionic human eye in the process of object recognition,consciously extract different levels of features,the characteristics of different network levels are used as fully connected input.In order to reduce the influence of illumination and facial rotation on the accuracy of face recognition,this paper proposes a method of face rotation correction and feature extraction for face images using PCA and Gabor transform in face input images.2)Aiming at the problem of large power consumption,large storage and slow operation speed for the deep neural network model,this paper first compresses the network model from the point of view of the algorithm,and uses model dropout and layers fusion to compress the network model.The model is compressed 4.5 times under the condition of smaller accuracy loss.Greatly reduces the storage space and computational complexity of the model.3)The current market mostly through network API,get intelligent service such as face recognition,However,the intelligent application based on network API interface is greatly influenced by the network state of mobile terminal environment.However,the intelligent application based on network API interface is greatly influenced by the network state of mobile terminal environment.And the hardware acceleration platform based on deep learning is mostly in the research and development stage.Therefore,this paper proposes to accelerate the face recognition algorithm based on deep learning through mobile phone GPU.In this paper,a parallel algorithm of face recognition is designed by using platform compatible GPU parallel software framework OpenCL.This paper makes a thorough optimization,For example,asynchronous data reading is proposed to hide the data copy time,and the convolution layer and the pooling layer are fused to reduce the parameters and the amount of computation,and the vectorization calculation is used to improve the computing speed.