[1]白罡旭,杨海峰,蔡江辉,等.基于混合采样策略与 Fixmatch 的图像分类方法[J].计算机技术与发展,2023,33(11):72-77.[doi:10. 3969 / j. issn. 1673-629X. 2023. 11. 011]
　BAI Gang-xu,YANG Hai-feng,CAI Jiang-hui,et al.An Image Classification Method Based on Hybrid Sampling Strategy and Fixmatch[J].,2023,33(11):72-77.[doi:10. 3969 / j. issn. 1673-629X. 2023. 11. 011]

点击复制

基于混合采样策略与 Fixmatch 的图像分类方法()

分享到：

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

卷:

33

期数:

2023年11期

页码:

72-77

栏目:

媒体计算

出版日期:

2023-11-10

文章信息/Info

Title:

An Image Classification Method Based on Hybrid Sampling Strategy and Fixmatch

文章编号:

1673-629X(2023)11-0072-07

作者:

白罡旭¹ ; 杨海峰¹ ; 蔡江辉¹; 2 ; 王玉鹏¹

1. 太原科技大学 计算机科学与技术学院,山西 太原 030024;

2. 中北大学,山西 太原 030051

Author(s):

BAI Gang-xu1 ; YANG Hai-feng1 ; CAI Jiang-hui1; 2 ; WANG Yu-peng1

1. School of Computer Science and Technology,Taiyuan University of Science and Technology,Taiyuan 030024,China;

2. North University of China,Taiyuan 030051,China

关键词:

半监督学习; 聚类; 采样策略; 图像分类; 数据增强

Keywords:

semi-supervised learning; clustering; sampling strategy; image classification; data augmentation

分类号:

TP181

DOI:

10. 3969 / j. issn. 1673-629X. 2023. 11. 011

摘要:

RUC( Unsupervised Image Clustering with Robust Learning) 是一种为改善聚类性能而提出的图像分类方法,但是由于它的协同训练仅适用于双视图数据集,并且没有考虑到数据

之间相似性对其伪标签采样策略的影响, 此外其使用的Mixmatch 只是单纯地进行 k 次随机增强求均值却没有考虑到强增强与弱增强的联系。 为了解决这些问题,该文提出了HFC

( classification method based on Hybrid sampling strategy and Fixmatch) 。 首先,设计了一种置信度与距离的伪标签采样策略,联合两种策略以提高筛选到正确标签的概率;其次,

使用 Tri-training 取代 Co-training,即通过两个分类器指导第三个分类器进行训练,使得模型不再受限于双视图数据集;最后,采用目前较好的 Fixmatch 的数据增强方法取代 RUC 中Mixmatch 随机增强,以突出强增强与弱增强的联合作用。 HFC 在 CIFAR-10、CIFAR-100 和 STL-10 数据集上进行实验,取得了较好的结果,验证了该方法的有效性。

Abstract:

RUC is an image classification method proposed to improve clustering performance, but its cooperative training is onlyapplicable to two-view data sets. Moreover,the influence?

of the similarity between the data on the pseudo-label sampling strategy wasnot taken into account. In addition,the Mixmatch used simply carried out k sub-random enhancement to find the mean value without considering the relationship between strong enhancement and weak enhancement. To solve these problems, we propose an HFC( classification method based on Hybrid sampling strategy and Fixmatch) . Firstly,a pseudo-label sampling strategy with confidence anddistance is designed,and the two strategies are combined to improve the probability of filtering to the correct label. Secondly,the Tri-training is used instead of Co-training,that is,the third classifier is guided by two classifiers for training,so that the model is no longerlimited by the two-view dataset. Finally,we replace the random Mixmatch augmentation in RUC with the Fixmatch data augmentationmethod to highlight the combined effect of strong enhancement and weak enhancement. Experiments with HFC on CIFAR-10,CIFAR-100 and STL-10 datasets have obtained good results,which verify the effectiveness of the proposed method.

常用功能