基于粒子群的卷积神经网络细粒度搜索方法  

Fine-Grained Search Method Convolutional Neural Network for Classification Based on Particle Swarm Optimization

作　　者：程吉祥 肖舒 王圳鹏 李志丹 CHENG Ji-xiang;XIAO Shu;WANG Zhen-peng;LI Zhi-dan(School of Electrical Information,Southwest Petroleum University,Chengdu Sichuan 610500,China)

机构地区：[1]西南石油大学电气信息学院,四川成都610500

出　　处：《计算机仿真》2023年第11期300-305,489,共7页Computer Simulation

基　　金：国家自然科学基金资助项目(61603319,61601385);西南石油大学智能控制与图像处理青年科技创新培育团队资助项目(2017CXTD010)。

摘　　要：针对图像分类任务设计高效紧致的卷积神经网络(Convolutional Neural Networks,CNN)是极具挑战性的问题,其设计过程极大依赖专家的经验和不断的试错。综合考虑CNN中卷积层、池化层、全连接层和激活函数等要素,提出了一种可变规模的CNN细粒度搜索方法。方法设计了一种细粒度搜索空间,增加搜索参数的类别;提出粒子群可变长编码映射方法,解决候选网络编码冗余性问题;通过创建代理数据集评估方式,降低CNN评估时间;提出进化CNN方法,实现高效的训练网络模型。在7个图像分类任务中,实验结果表明,所提方法获得模型与大量经验设计模型和其它搜索方法获得模型相比分类准确率和模型大小有较强的优势和竞争力。Designing fficient and compact convolutional neural networks for image classification tasks is a very challenging problem,and its design process relies heavily on the experience of experts and continuous trial and error.In this paper,a fine-grained search method of CNN with variable scale is proposed,considering the convolution layer,pooling layer,full connection layer and activation function in CNN.This method designs a fine-grained search space and increases the categories of search parameters.A particle swarm variable-length coding mapping method is pro-posed to solve the redundancy problem of candidate network coding.By creating proxy dataset evaluation methods,the CNN evaluation time is reduced.An evolutionary CNN method is proposed to achieve efficient training of network models.In seven image classification tasks,the experimental results show that the proposed model has stronger advan-tages and competitiveness in classification accuracy and model size compared with a large number of empirical design models and other search methods.

关 键 词：图像分类 卷积神经网络 可变长编码映射 代理数据集 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]