[1]俞圳韬,刘万里,杨晓辉,等.基于贝叶斯优化和迁移学习的 CNN 算法研究[J].计算机技术与发展,2022,32(S2):68-71.[doi:10. 3969 / j. issn. 1673-629X. 2022. S2. 012]
　YU Zhen-tao,LIU Wan-li,YANG Xiao-hui,et al.Research on CNN Algorithm Based on BayesianOptimization and Transfer Learning[J].,2022,32(S2):68-71.[doi:10. 3969 / j. issn. 1673-629X. 2022. S2. 012]
点击复制

基于贝叶斯优化和迁移学习的 CNN 算法研究()
分享到：

《计算机技术与发展》[ISSN:1006-6977/CN:61-1281/TN]

卷:
32
期数:
2022年S2期
页码:
68-71
栏目:
图形与图像
出版日期:
2022-12-11

文章信息/Info

Title:
Research on CNN Algorithm Based on BayesianOptimization and Transfer Learning
文章编号:
1673-629X(2022)S2-0068-04
作者:
俞圳韬1 ; 刘万里2 ; 杨晓辉2 ; 黄玉珍2 ; 徐 雷1 ; 夏吉安3
1. 南京理工大学 计算机科学与工程学院,江苏 南京 210094;
2. 南京市中西医结合医院,江苏 南京 210014;
3. 南京工业职业技术大学,江苏 南京 210023
Author(s):
YU Zhen-tao1 ; LIU Wan-li2 ; YANG Xiao-hui2 ; HUANG Yu-zhen2 ; XU Lei1 ; XIA Ji-an3
1. School of Computer Science and Engineering,Nanjing University of Science and Technology,Nanjing 210094,China;
2. Nanjing Hospital of Integrated Traditional Chinese and Western Medicine,Nanjing 210014,China;
3. Nanjing Vocational University of Industry Technology,Nanjing 210023,China
关键词:
图像识别; 卷积神经网络; 迁移学习; 贝叶斯优化算法; VGG-16
Keywords:
image recognition; convolutional neural network; transfer learning; Bayesian optimization algorithm; VGG-16
分类号:
TP301. 6
DOI:
10. 3969 / j. issn. 1673-629X. 2022. S2. 012
摘要:
随着深度学习的快速发展,深层次的卷积神经网络在图像识别领域获得了广泛的应用,但是在实际应用中,越深的卷积神经网络在拥有强大拟合能力和优越性能的同时,也存在模型难以训练、耗费时间较长的问题。 为了解决这个问题,提出使用贝叶斯优化算法对卷积神经网络的超参数进行优化,在不断的迭代过程中,得到超参数向量的最优解,然后使用最优解参数对神经网络进行训练验证,生成最优分类模型。 另外,为了避免每次迭代对网络权重参数的大幅度调整,网络模型架构底层使用了 VGG-16 进行迁移学习,这大大缩短了一次迭代网络收敛的时间。 实验在数据集 MNIST 和CIFAR-10 上展开,根据实验结果可知,在不牺牲图像分类模型准确率的前提下,使用贝叶斯优化算法只需要进行少数几次迭代就可以搜索到神经网络超参数的最优解,大大节省了生成最优分类模型的时间和 GPU 资源。
Abstract:
With the rapid development of deep learning,deep convolutional neural networks have been widely applied in the field of imagerecognition. However, in practical application, the deeper convolutional neural networks have strong fitting ability and superiorperformance,but also have the problems of difficult model training and it will take long time. In order to solve this problem,we proposeto use Bayesian optimization algorithm to optimize the hyperparameter of convolutional neural network. In the process of continuousiteration,the optimal solution of the hyperparameter vector is obtained,and then the optimal solution parameters are used to train andverify the neural network and generate the optimal classification model. In addition,in order to avoid the drastic adjustment of networkweight parameters in each iteration,VGG-16 is used in the bottom layer of the network model architecture for migration learning,whichgreatly reduces the time of network convergence in one iteration. In this paper, the experiment is based on MNIST and CIFAR - 10datasets,according to the result of experiment, it shows that under the precondition of without sacrificing the accuracy of imageclassification model,using the Bayesian optimization algorithm only needs a few iterations can search the optimal hyperparameter,greatlysaving the time and GPU resources to generate the optimal classification model.

相似文献/References:

[1]王明平 宋丽梅.基于计算机视觉的车架号采集系统[J].计算机技术与发展,2008,(04):239.
　WANG Ming-ping,SONG Li-mei.Vehicle Identify Number Acquisition System Based on Machine-Vision[J].,2008,(S2):239.
[2]徐仕玲 赵敏 徐建波.野外早期火灾图像识别方法研究[J].计算机技术与发展,2008,(06):214.
　XU Shi-ling,ZHAO Min,XU Jian-bo.Research for Early Fire Image Recognition Technology in Outdoors[J].,2008,(S2):214.
[3]张婷 吴元君 黄俊 吴建国.选票选举系统中选票图像的预处理方法研究[J].计算机技术与发展,2007,(04):225.
　ZHANG Ting,WU Yuan-jun,HUANG Jun,et al.Research of Image Pre- Processing in Vote Processing System[J].,2007,(S2):225.
[4]吴忠 朱国龙 黄葛峰 吴建国.基于图像识别技术的手写数字识别方法[J].计算机技术与发展,2011,(12):48.
　WU Zhong,ZHU Guo-long,HUANG Ge-feng,et al.Handwritten Digit Recognition Based on Image Recognition System[J].,2011,(S2):48.
[5]汪晨 张涛 林为民 邓松 时坚 李伟伟.图像识别综述及在电力信息安全中的应用研究[J].计算机技术与发展,2012,(04):161.
　WANG Chen,ZHANG Tao,LIN Wei-min,et al.Image Recognition Review and Application Research in Electric Power Information Security[J].,2012,(S2):161.
[6]王珂雅 邱力军.一种新的脑部CT图像异常检测算法[J].计算机技术与发展,2012,(05):185.
　WANG Ke-ya,QIU Li-jun.A New Anomaly Detection Algorithm for Brain CT Image[J].,2012,(S2):185.
[7]周诚诚 张代远[].利用图像识别技术过滤海量可疑钓鱼网站[J].计算机技术与发展,2012,(11):246.
　ZHOU Cheng-cheng,ZHANG Dai-yuan.Using Image Recognition Technology to Filter Mass Suspicious Phishing Sites[J].,2012,(S2):246.
[8]崔凤焦.表情识别算法研究进展与性能比较[J].计算机技术与发展,2018,28(02):145.[doi:10．3969/j．issn．1673－629X．2018．02．031]
　CUI Feng-jiao.Ｒesearch and Performance Comparison of Facial Expression Ｒecognition Algorithm[J].,2018,28(S2):145.[doi:10．3969/j．issn．1673－629X．2018．02．031]
[9]张丹丹,李雷. 基于PCANet-RF的人脸检测系统[J].计算机技术与发展,2016,26(02):31.
　ZHANG Dan-dan,LI Lei. Face Detection System Based on PCANet-RF[J].,2016,26(S2):31.
[10]陈强锐,谢世朋.基于深度学习的肺部肿瘤检测方法[J].计算机技术与发展,2018,28(04):201.[doi:10.3969/ j. issn.1673-629X.2018.04.043]
　CHEN Qiang-rui,XIE Shi-peng.Lung Cancer Detection Method Based on Deep Learning[J].,2018,28(S2):201.[doi:10.3969/ j. issn.1673-629X.2018.04.043]
[11]王山海,刘谦,马鑫鑫.基于图像识别的人工影响天气业务的研究[J].计算机技术与发展,2019,29(05):172.[doi:10. 3969 / j. issn. 1673-629X. 2019. 05. 036]
　WANG Shan-hai,LIU Qian,MA Xin-xin.Research on Weather Modification Based on Image Recognition[J].,2019,29(S2):172.[doi:10. 3969 / j. issn. 1673-629X. 2019. 05. 036]
[12]孙娇娇,龚 安,史海涛.基于卷积神经网络的低剂量 CT 图像肺结节检测[J].计算机技术与发展,2019,29(11):173.[doi:10. 3969 / j. issn. 1673-629X. 2019. 11. 035]
　SUN Jiao-jiao,GONG An,SHI Hai-tao.Detection of Low-dose CT Pulmonary Nodule Based on Convolutional Neural Network[J].,2019,29(S2):173.[doi:10. 3969 / j. issn. 1673-629X. 2019. 11. 035]
[13]徐明远,崔 华,张立恒.基于改进 CNN 的公交车内拥挤状态识别[J].计算机技术与发展,2020,30(05):32.[doi:10. 3969 / j. issn. 1673-629X. 2020. 05. 007]
　XU Ming-yuan,CUI Hua,ZHANG Li-heng.Recognition of Crowded State in Bus Based on Improved Convolution Neural Network[J].,2020,30(S2):32.[doi:10. 3969 / j. issn. 1673-629X. 2020. 05. 007]
[14]张素智,吴玉红,常 俊.基于改进 AlexNet 卷积神经网络的轮胎图像识别[J].计算机技术与发展,2021,31(07):182.[doi:10. 3969 / j. issn. 1673-629X. 2021. 07. 030]
　ZHANG Su-zhi,WU Yu-hong,CHANG Jun.Tire Damage Image Recognition Based on Improved AlexNetConvolutional Neural Network[J].,2021,31(S2):182.[doi:10. 3969 / j. issn. 1673-629X. 2021. 07. 030]

更新日期/Last Update: 2022-10-10