Decoupled Learning for Long-Tailed Oracle Character Recognition

Jing Li, Bin Dong, Qiu Feng Wang*, Lei Ding, Rui Zhang, Kaizhu Huang

*Corresponding author for this work

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

1 Citation (Scopus)

Abstract

Oracle character recognition has recently made significant progress with the success of deep neural networks (DNNs), but it is far from being solved. Most works do not consider the long-tailed distribution issue in oracle character recognition, resulting in a biased DNN towards head classes. To overcome this issue, we propose a two-stage decoupled learning method to train an unbiased DNN model for long-tailed oracle character recognition. In the first stage, we optimize the DNN under instance-balanced sampling, obtaining a robust backbone but biased classifier. In the second stage, we propose two strategies to refine the classifier under class-balanced sampling. Specifically, we add a learnable weight scaling module which can adjust the classifier to respect tail classes; meanwhile, we integrate the KL-divergence loss to maintain attention to head classes through knowledge distillation from the first stage. Coupling these two designs enables us to train an unbiased DNN model in oracle character recognition. Our proposed method achieves new state-of-the-art performance on three benchmark datasets, including OBC306, Oracle-AYNU and Oracle-20K.

Original languageEnglish
Title of host publicationDocument Analysis and Recognition – ICDAR 2023 - 17th International Conference, Proceedings
EditorsGernot A. Fink, Rajiv Jain, Koichi Kise, Richard Zanibbi
PublisherSpringer Science and Business Media Deutschland GmbH
Pages165-181
Number of pages17
ISBN (Print)9783031416842
DOIs
Publication statusPublished - 2023
Event17th International Conference on Document Analysis and Recognition, ICDAR 2023 - San José, United States
Duration: 21 Aug 202326 Aug 2023

Publication series

NameLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume14190 LNCS
ISSN (Print)0302-9743
ISSN (Electronic)1611-3349

Conference

Conference17th International Conference on Document Analysis and Recognition, ICDAR 2023
Country/TerritoryUnited States
CitySan José
Period21/08/2326/08/23

Keywords

  • Decoupled learning
  • Knowledge distillation
  • Long tail
  • Oracle character recognition

Fingerprint

Dive into the research topics of 'Decoupled Learning for Long-Tailed Oracle Character Recognition'. Together they form a unique fingerprint.

Cite this