Quantitative Research On Convolutional Neural Networks And FPGA Implementation

Convolutional neural network is an important technology in the field of deep learning.Transplanting convolutional neural network to embedded devices is of great significance to the research of algorithm landing.However,conventional convolutional neural network has the characteristics of large model and many parameters,which brings new challenges to deploy convolutional neural network on embedded platform with limited resources.As for the deployment platform,CPU and GPU are usually the main choices.With the in-depth study of edge computing applications,more and more researchers pay attention to FPGA.Therefore,FPGA is an ideal deployment platform since it features low power consumption and flexible configuration.Therefore,to deploy a convolutional neural network on a resource-limited FPGA platform,the scale of the model needs to be compressed.This paper adopts a method of binarizing the weights so that the parameters of the convolutional layer and the fully connected layer are only +1 and-1,the network intermediate calculation result retains full precision.In this paper,the FPGA was chosen as the deployment platform.The XILINX toolchain and HLS development tools are used to convert high-level languages into underlying HDL languages to deploy network models to the ZYNQ7020 chip platform.The quantization of convolutional neural network model and the deployment of FPGA are experimentally validated,the results showed that the quantified model accuracy has no loss compared to the full accuracy of the model without quantification,meanwhile the model size has been effectively compressed.The model was transplanted to chip platform ZYNQ7020,coming to the solution of overall hardware featuring high accuracy,low power consumption.