基于增强引力搜索和神经网络的图像分类  

Image classification based on enhanced gravitational search and neural networks

作　　者：侯小毛 马凌 赵月爱[2] HOU Xiao-mao;MA Ling;ZHAO Yue-ai(School of Electronic Information,Hunan Institute of Information Technology,Changsha 410151,China;Department of Computer Science,Taiyuan Normal University,Jinzhong 030619,China)

机构地区：[1]湖南信息学院电子信息学院,湖南长沙410151 [2]太原师范学院计算机系,山西晋中030619

出　　处：《计算机工程与设计》2020年第12期3495-3502,共8页Computer Engineering and Design

基　　金：湖南省应用特色学科建设基金项目(湘教通〔2018〕469号);湖南省教育厅科学研究基金项目(18B571)。

摘　　要：深度神经网络对于图像分类问题具有较好的准确性,但深度卷积神经网络的参数繁多且难以确定,针对这种情况,提出基于增强引力搜索算法和卷积神经网络的图像分类算法。为引力搜索算法引入对数引力常量衰减函数、交叉算子和变异算子,增强引力搜索的全局搜索能力。设计直接的深度神经网络编码形式,有利于加快引力搜索的计算速度,给出agent各个属性的更新方法。实验结果表明,该方法在保持较高图像分类准确率的情况下,成功加快了深度神经网络参数的学习速度。Deep convolutional neural networks can realize good classification accuracy for the image classification problem,but it suffers from a large amount of parameters and it is difficult to search the parameters,in view of this,an image classification algorithm based on enhanced gravitational search algorithm and convolutional neural networks was proposed.The logarithmic decreasing function,crossover operation and mutation operation were introduced to gravitational search algorithm,these modifications enhanced the global search performance of gravitational search algorithm.A direct encoding representation of deep neural networks was designed,which helped to improve the computational speed of gravitational search algorithm,and the respective updating method of each property of agents was presented.Experimental results show that the proposed method realizes good image classification accuracy,and it speedups the learning process of deep neural networks.

关 键 词：深度学习 卷积神经网络 重引力搜素算法 超参数 图像分类 数据挖掘 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]