Research And Implementation Of Face Recognition System On Multi-Core Embedded Platform

With the development of intelligent society,face recognition has been widely concerned with its advantages of non-coerciveness,universality and non-contact.In a broad sense,face recognition usually refers to face detection and face identity recognition.Among them,face detection refers to accurately finding the position of a face from a large image,while face identity recognition refers to identifying the detected face portrait.In this thesis,a set of face recognition system suitable for multi-core embedded platform is designed,and face detection and recognition are respectively deployed on two development boards for verification.Then,two development boards are connected through the network cable and data transmission is carried out by Socket,forming a complete face recognition system.In the face detection part,this thesis adopts NPD(Normalized Pixel Difference)algorithm to obtain the face portrait from a large image frame,and then adopts LBP(Local Binary Pattern)algorithm to deprocess the face image of the same person in the connected image frame,and then adopts VJ(Viola-Jones)algorithm to filter the wrong detection images of NPD.In this thesis,the detection process was simulated and finally deployed on DM8168 platform.In order to make full use of the computing resources of the platform,In this thesis,DVR_RDK framework is used to calculate the average amount.Among them,NPD algorithm runs in DSP(Digital Signal Processor)core,LBP and VJ algorithm runs in ARM core,and finally makes each algorithm can complete the task interaction in real time.In the part of face recognition,in order to improve the accuracy of extraction,this thesis first use the MTCNN(Multi-task Cascaded Convolutional Networks)algorithm to do affine transformation and cutting preprocessing of face image,and then use the convolutional neural network to determine the face identity.To solve the problem of weak computing power of embedded platform,this thesis designed a shallow network for identity recognition.The network has only 84.6M times of multiplication and accumulation,and the accuracy rate of the trained model is 95.77%.In this thesis,the face recognition part is firstly deployed on the ARM core of AM5728 platform,and two multi-core optimization methods are proposed for the forward propagation of neural network.The first method is to allocate the computation amount to DSP core operation by OpenCL method,which can shorten the operation time by 54.7% under the condition of keeping the accuracy of the model unchanged.The second method is to reduce the complexity of the network model itself through sparse and quantization,and the computing speed can be increased by 6.3 times when the precision is slightly reduced...