A Two-Stage Active Learning Method for Image Classification

Feiyue Wang, Xu Li, Yifan Zhang, Baoguo Wei, Lixin Li

Research output: Chapter in Book/Report/Conference proceedingConference contributionpeer-review

Abstract

Recent successes in learning-based image classification rely heavily on a large number of annotated training samples, which often require considerable human effort. In this paper, we propose a new two-stage active learning (AL) method for image classification with a query strategy considering both uncertainty and diversity. In the first stage, the uncertainty is used to determine the candidate set. In the second stage, the candidate set is clustered, and the nearest sample from the cluster center is selected to increase the diversity of samples. Our method uses Poly-1 loss as the classification loss and Binary Cross Entropy (BCE) as the binary loss to distinguish between the labeled samples and the unlabeled samples. We train the two classifiers in a joint way and evaluate our method on the CIFAR10 and CIFAR100 datasets. The rich experimental results show that the proposed method outperforms the state-of-the-art AL methods in image classification tasks.

Original languageEnglish
Title of host publicationICIEA 2022 - Proceedings of the 17th IEEE Conference on Industrial Electronics and Applications
EditorsWenxiang Xie, Shibin Gao, Xiaoqiong He, Xing Zhu, Jingjing Huang, Weirong Chen, Lei Ma, Haiyan Shu, Wenping Cao, Lijun Jiang, Zeliang Shu
PublisherInstitute of Electrical and Electronics Engineers Inc.
Pages1134-1139
Number of pages6
ISBN (Electronic)9781665409841
DOIs
StatePublished - 2022
Event17th IEEE Conference on Industrial Electronics and Applications, ICIEA 2022 - Chengdu, China
Duration: 16 Dec 202219 Dec 2022

Publication series

NameICIEA 2022 - Proceedings of the 17th IEEE Conference on Industrial Electronics and Applications

Conference

Conference17th IEEE Conference on Industrial Electronics and Applications, ICIEA 2022
Country/TerritoryChina
CityChengdu
Period16/12/2219/12/22

Keywords

  • active learning
  • image classification
  • joint learning
  • query strategy

Fingerprint

Dive into the research topics of 'A Two-Stage Active Learning Method for Image Classification'. Together they form a unique fingerprint.

Cite this