Research And Implementation Of Artificial Neural Network Based On FPGA

Artificial neural network ANNs(Neural Network Artificial) is a system that imitate the structure and relationship of the human brain neural network, which is composed of a large number of interconnected processing units. Today, artificial neural network has been widely used in many fields, such as signal processing, medical and health, control system, pattern recognition, etc.. Our common method of realizing the neural network is based on the software.Because the software operation is slow, so we need to put forward a new neural network implementation method which can meet the requirements of real-time operation. In this paper, a neural network based on FPGA is realized, which is characterized by parallel computing and fast operation. The neural network which is based on hardware implementation reflects the inherent parallel processing characteristics of neural network, and the processing speed is greatly improved.This paper first introduces the background of the research of artificial neural network and the present situation of the research at home and abroad, as well as the method and significance of the hardware implementation of artificial neural network. Then, the typical BP neural network algorithm is studied, including the specific algorithm implementation steps. The essence of the BP algorithm is to use the gradient descent method to modify the connection weights between neurons, so as to achieve the goal of solving the optimal solution. Since the design’s application background is the real-time identification of 26 English letters, in the consideration of hardware implementation of the neural network, the digital process of the letter images has also been taken into account. Then, the whole BP neural network is divided into modules, and the hardware design of each module is also carried out. Among them, the hardware implementation of the incentive function is the focus of the neural network hardware design that based on the FPGA. In this thesis, we use the widely used Sigmoid function as the activation function of BP neural network, and realize the Sigmoid function on FPGA by using the method of combination of look-up table and the piecewise linear function approximation. Besides, this paper adopts a systolic array architecture in the design of the whole neural network architecture. The structure reflects the parallel characteristics of neural network, its hardware implementation can make the entire network processing speed greatly improved.Then, using the Verilog hardware description language to have a hardware implementation with the designed neural network. After that, the Modelsim SE 6.2 simulation software and Xilinx company’s Vivado integrated development tool are used for the design of functional simulation and integrated optimization. In the error range, the hardware function is correct. Moreover, the FPGA resource utilization of the design is also low, which is under 10%. Finally, through the board level verification, the highest frequency of the clock can reach to 70 MHz, which can meet the requirements of real-time recognition of 26 English capital letters. Therefore, the design of the artificial neural network based on FPGA in this paper achieve the original intention.