| As the key technology of artificial intelligence,artificial neural network has become one of the strategic technologies that countries around the world are competing to develop,and it is also an important driving force for a new round of technological revolution and industrial transformation.FPGA has the advantages of reconfigurability,low power consumption,high performance,etc.,which is very suitable for the realization of neural network,but the neural network control algorithm is complex and the calculation is large,so the design has high integration and good real-time performance.High performance realized by FPGA The neural network system has important theoretical value and practical application value.This paper takes neural network as the specific object,and aims at designing and developing neural network system realized by FPGA.It conducts in-depth research around the realization structure of neural network,the key modules of neural network realization,and FPGA hardware realization of handwritten digit recognition system.The main work and conclusions are as follows:(1)The FPGA implementation design and implementation research of Sigmoid activation function based on CORDIC algorithm.The CORDIC algorithm implements complex nonlinear functions,only a series of addition and shift operations,simple operation and strong versatility.The experiment proved the correctness of the design.(2)A new method of variable section nonlinear double fitting based on the initial fitting error is proposed.The fitting approximation experiment of Softmax is carried out using Python,and it is implemented on FPGA.The fitting method firstly finds the interval with larger error through uniform segmentation and random segmentation nonlinear fitting,and then selects the interval with smaller/larger error according to the interval of larger/smaller error to perform the second nonlinear fitting.The analysis of experimental results shows that this method not only solves the problem of large interval errors when using the piecewise nonlinear fitting approximation method to fit Softmax,but also has a great improvement in speed and accuracy.(3)An improved MNIST handwritten digit recognition convolutional neural network algorithm suitable for FPGA hardware implementation is proposed.Python is used to train the neural network parameters,and a handwritten digit is realized by high-level synthesis(HLS)language programming and FPGA.recognition system.The improved convolutional neural network algorithm removes the bias parameter of the neuron activation function.In the FPGA logic design of HLS,design ideas such as parallel,pipeline,and folding of hardware design optimization are integrated into various design optimizations of HLS programming of convolutional neural networks.Experimental results show that the improved handwritten digit recognition algorithm has a recognition accuracy rate of 98.21%,a hardware single frame processing time of 0.031 s,and a power consumption of 1.96 W,which has good performance.At the same time,the handwritten digit recognition system is transplanted and extended to implement an FPGA-based clothing recognition system.The system has a clothing recognition accuracy rate of 88.87% after testing,which shows that the system has good transplant scalability. |
