Design And Implementation Of Convolution Neural Network Based On FPGA

Convolutional neural network is widely used in the field of image recognition because of its high accuracy due to its ability to simulate the behavior of biological optic nerve.With the rapid increase in the demand for terminal applications,the initial neural network computing platform based on CPU and GPU,its disadvantages such as large size and high power consumption have become more and more significant.FPGA is a programmable logic device with abundant programmable logic resources.It has the advantages of low power consumption,small size,and reconfiguration.It matches the hardware deployment requirements of convolutional neural networks and can be used in small embedded systems.This paper mainly studies the method of using FPGA to realize convolutional neural network.The main content includes three parts: the construction of convolutional neural network,the training of convolutional neural network and the deployment of convolutional neural network.A lightweight convolutional neural network structure is proposed based on the analysis of the computation capacity of existing Spartan-6 FPGA logic resources.Since FPGA is not good at handling floating point number operation,the network model parameters are further quantized from32 bit floating point numbers to 8bit fixed point numbers.The Pytorch deep learning framework is used to train the handwritten digital sample images from the MNIST database through the back propagation mechanism.On the basis of ensuring a high recognition rate,the network extracts the weights and biases with higher accuracy after parameter quantization.By studying the structure and computing characteristics of the network,FPGA is used to realize the forward propagation of the network,and a complete circuit model is constructed.Finally,Verilog language is used to realize the designed five-layer convolutional neural network,and the design scheme of FPGA is given to complete the handwritten digit recognition.The FPGA development board selected in this paper is AV6150,and the image recognition system is compiled and integrated on the software ISE,and the final verification is carried out on the lower board.The results show that the images in MNIST database can be correctly recognized by the system.Under the system clock of 50 MHz,it takes about 10 ms to complete the image recognition,which is better than the running speed of the software.Therefore,it proves that the design is feasible and meets the real-time requirements.