A Study Of Convolutional Neural Network Based On Improved Particle Swarm Optimization And Its Application

In the context of the rapid development of computer technology and the Internet,neural networks are currently a hot research field.Domestic and foreign universities and major technology giants are investing more and more resources in researching neural networks,building higher-level networks and researching more efficient training methods.At present,the method used in neural network training is mainly the Back Propagation algorithm(BP algorithm).In addition,Particle Swarm Optimization(PSO),as a classical group intelligent optimization method,is widely used in various fields of optimization.With the development of big data,the nature of optimization problems gradually moves from low dimensions to high dimensions.In this paper,the high-dimensional data similarity measurement is first introduced to measure the distance between particles in high-dimensional PSO,and the improved PSO is applied to the design of convolutional neural network optimization algorithm.Finally,it is applied to traffic sign recognition.The specific research work is as follows :(1)This paper presents an improved high-dimensional PSO algorithm.For the PSO to solve the high-dimensional optimization problem,the PSO algorithm suitable for low-dimensional optimization problems is prone to early convergence and poor global searching ability,etc.Through literature review and simulation,this paper firstly verifies that using the traditional iteration method to find the global optimal individual and historical optimal individual information is not very efficient.Therefore,based on the high-dimensional data similarity measurement method,this paper redefines a distance paradigm between particles and embeds it into the PSO iterative process.For the CEC test function set example,the simulation results show that the improved PSO improves the iterative effect of the particle to some extent and improves the convergence speed and global search ability in the high-dimensional optimization problem.(2)By analyzing the weight parameter optimization problem of convolutional neural network training,this paper designs a traditional gradient reduction hybrid algorithmcombining the improved PSO algorithm and BP algorithm.The hybrid algorithm takes the loss function as target function and optimizes target function based on the mechanism of PSO algorithm and the gradient information of the loss function of convolutional neural network.The hybrid algorithm utilizes the gradient information of the loss function and the cooperative search mechanism of the PSO algorithm to effectively improve the convergence speed and global search ability of the algorithm while ensuring that the computational load is acceptable.In addition,the design of the algorithm solves some problems in the implementation of PSO algorithm in convolutional neural network optimization.(3)At the end of this paper,the traffic sign recognition task is selected to analyze the performance of the algorithm.Based on the PyTorch neural network framework of Python language,this paper constructs a LeNet network for handwritten digit recognition and analyzes the performance of the proposed algorithm.Based on AlexNet network,the improved multi-layer convolutional neural network is applied to traffic sign recognition.The experimental results demonstrate the effectiveness of the improved optimization algorithm.