Research On Algorithm Of Convolutional Neural Network Suitable For Engineering Implementation

In the field of artificial intelligence,deep learning is a very important component.In recent years,the development of deep learning technology has been very rapid,and it has been widely applied to many fields in real life,such as face recognition and voice assistance.The main technical principle of deep learning is to combine various specific features of the bottom layer to generate more advanced representations and to extract useful features among them.Deep learning is a type of artificial neural network.It has deeper layers,larger scales,higher complexity,and greater training difficulty.Convolutional neural networks are representative of deep learning.The convolutional neural network is mainly composed of three characteristics,namely local receptive fields,shared weights,and temporal or spatial subsampling,and it has the recognition invariance of deformation,such as movement,scaling,and distortion.For purpose of achieving better results in picture classification tasks,the convolutional neural network has a gradually deeper level,the structure becomes increasingly complex,leading to an increase in storage capacity and computational memory,which hinders the widespread application of convolutional neural networks in the industry.In order to reduce these requirements and realize the deployment of convolutional neural networks on hardware platform(FPGA/ASIC chip),this paper focuses on the convolutional neural network algorithm suitable for engineering implementation.The main work and contributions in this paper are as follows:(1)This paper studies the quantization algorithm of training convolutional neural networks using fixed-point number,compares and analyzes the existing different fixed-point quantization methods and binarization methods for convolutional neural networks.(2)This paper innovatively proposes a quantization method of quantifying the pre-training floating-point convolutional neural network to a fixed-point network,and compares it with the quantization method using a fixed-point number training convolutional neural network.(3)This paper studies the representation format of fixed-point number used in the fixedpoint quantization of convolutional neural networks.(4)According to the memory requirements of the hardware platform,the time cost,and the requirements for network test accuracy,this paper points out the most suitable quantization algorithm and representation format of fixed-point number for the convolutional neural network deployed on the hardware platform.This paper first introduces the development and background of convolutional neural networks and the research status of convolutional neural networks based on engineering improvement algorithms.Then it elaborates the principle of convolutional neural networks,including the basic knowledge of artificial neural networks,the structure of the convolutional neural network,the structural characteristics of the convolutional layer,the training process of the convolutional neural network,and techniques to prevent overfitting.Next,this paper introduces the different fixed-point quantization algorithms for convolutional neural networks and compares their performance by testing these algorithms.Then it validate the number format of fixed-point number used for fixed-point quantization.This paper points out the fixed-point quantization algorithm and the representation format of fixed-point number that are suitable for implementation on the hardware platform.Finally,it summarizes and analyzes the tasks accomplished in the full text,and proposes the inadequacies and follow-up tasks.