Design And Implementation Of Image Recognition System Based On RISC-V SoC

Image is an essential way for humans to obtain information.Image recognition plays a vital role in many fields.Among them,face recognition is a method of rapid identification of human identity features.Its application is more widely,and the scenarios are more concentrated in embedded and mobile terminals.However,as face recognition accuracy improves,the convolutional neural network model of face recognition is becoming more and more complex.So the deployment of face recognition algorithms in hardware resource-limited platforms has become a hot issue to be studied.RISC-V instruction set architecture can realize simple,flexible,and low power consumption customization design due to its characteristics of minimalism,modularization,and scalability.This topic established RISC-V embedded platform and completed the design and deployment of two lightweight face recognition algorithms,the final design and implementation of RISC-V So C based face image recognition system.Firstly,based on the face recognition process,an embedded face recognition platform based on RISC-V So C is designed and implemented on FPGA with RISC-V kernel as the primary control core,and in order to solve the problems of the embedded platform resources presents a video down-sampling technology,effectively improve the system of resource utilization and enhance the portability of the system.Secondly,this paper proposes two lightweight convolutional neural network models based on mainstream face recognition algorithms.One is the open-set face recognition algorithm – Emfacenet.The network takes an inverted residual structure as the main body.Further,it reduces the number of network layers and parameters by combining deep convolution,point-by-point convolution,and other efficient computing methods.Squeeze and excitation modules are also introduced to enhance the information transmission between channels.Thus,the recognition speed of the neural network can be improved without sacrificing precision.Under the condition of keeping the recognition accuracy at90.65%,the recognition speed of the neural network on the RISC-V embedded platform is56.65 times,2.09 times,and 3.41 times of Res Net50,Mobile Net V3,and Mobile Face Nets,respectively.The other is the closed-set face recognition method – Mbfacenet,which takes inverted residual structure as the main body and introduces the improved nonlinear activation function to achieve the linear fitting of piecewise function to further reduce the complex calculation in the neural network.After the last inverted residual network layer,a lighter ECA attention module was introduced to ensure the accuracy of feature classification results further.Under the condition that the recognition accuracy of the neural network is kept at 98%,the recognition speed of the neural network on the RISC-V embedded platform is 4.76 times,4.00 times,26.71 times,and 14.33 times of Shuffle Net V2,Mobile Net V3,Efficient Net,and Mix Net respectively.