A Face Mask Recognition System Based On SSD Neural Network Algorithm

Currently,in order to ensure the safety of people’s own safety,prevent infectious diseases from raging,and meet the requirement of some specific places,this thesis proposes a SOC system design method based on the SSD neural network algorithm,which is used to achieve real-time masks detection.In response to the SOC system architecture design,this thesis proposes that the Hummingbird E203 with the RISC-V architecture is the main processor of the SOC system,and the associate processor module is added to the SOC system to achieve the SSD neural network algorithm to complete the masks detection.In response to the design problems of various modules in the SOC system,this thesis first combines the OV5640 camera group to design image collection,compression,preprocessing,and image display module to complete the collection and display of the image,combined with the VGA protocol to complete the input and output module design in the SOC system.In addition,this thesis proposes that the voice broadcast function is used to use the SD card and WM8731 chip to remind those who do not wear masks,design the relevant module to complete the relevant initialization and control functions.Then this thesis proposes a hardware design method for the SSD neural network algorithm to complete the hardware design of the modules of the neural network,such as convolutional modules,pooling modules,activation modules,etc.Due to the limited FPGA on board resources,a large number of parameters of the neural network are stored in DDR,and the hardware design method can effectively reduce the number of memory access.Finally,this thesis proposes a hardware implementation method for non-maximum suppression algorithms.This method greatly reduces the difficulty of achieving the hardware implementation of the algorithm after simplifying the non-maximum suppression algorithm.In addition,this thesis combines FPGA with high configuration characteristics,designing related modules to complete the number of neural network layers and related specifications parameter configuration.In response to the interaction between the main processor of the SOC system and the collaborator,this thesis combines the provisions of the custom instruction set of the custom instruction set in combination with the RISC-V architecture to complete the control function of the collaborator by the main processor.And use related interfaces to design related modules to implement instructions and related data interaction functions.Finally,this thesis uses related debugging software to write assembly language and related code,and use software and hardware to cooperate with the idea of completing the entire system construction,control the normal operation of the SOC system and perform related testing.This thesis tests the SOC system after the design of the face mask.The accuracy of the face mask detection accuracy of the people without a mask is 100%.The accuracy of the face mask detection of the face mask is 87.88%.In the system,336 DSP units are used to achieve the PE convolution computing unit in the neural network algorithm.After testing,the PE utilization rate is 66.3%.