Counterfactual Contrastive Learning for Fine Grained Image Classification

Chenke Yin, Jia Wang*, Haichao Zhang, Kaiyue Feng, Lin Shi, Qianyi Ma

*Corresponding author for this work

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

Abstract

In the realm of fine-grained image classification, discerning subtle distinctions between closely related categories remains a challenge. However, these approaches typically fall short in addressing the deeper causal relationships that underlie the visible features, leading to potential biases and limited generalizability. This paper presents a fine-grained causal contrastive network (FCCN), a novel architecture that integrates causal inference with contrastive learning to address the intricacies of fine-grained visual classification. Unlike traditional approaches that predominantly rely on feature correlations, FCCN deeply studies the causal relationship between salient features and labels, which greatly enhances the model’s ability to discriminate fine-grained images. A key innovation of FCCN is the introduction of a backdoor adjustment technique for feature decoupling, effectively minimizing the impact of irrelevant context feature and purifying the feature space for more precise classification. We validate our model on CUB-200-2011, Stanford Cars, and WM-811K datasets. Both accuracy and robustness are significantly improved, which demonstrate notable improvements in accuracy and robustness.

Original languageEnglish
Title of host publicationArtificial Neural Networks and Machine Learning – ICANN 2024 - 33rd International Conference on Artificial Neural Networks, Proceedings
EditorsMichael Wand, Jürgen Schmidhuber, Michael Wand, Kristína Malinovská, Jürgen Schmidhuber, Igor V. Tetko, Igor V. Tetko
PublisherSpringer Science and Business Media Deutschland GmbH
Pages169-183
Number of pages15
ISBN (Print)9783031723407
DOIs
Publication statusPublished - 2024
Event33rd International Conference on Artificial Neural Networks, ICANN 2024 - Lugano, Switzerland
Duration: 17 Sept 202420 Sept 2024

Publication series

NameLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume15019 LNCS
ISSN (Print)0302-9743
ISSN (Electronic)1611-3349

Conference

Conference33rd International Conference on Artificial Neural Networks, ICANN 2024
Country/TerritorySwitzerland
CityLugano
Period17/09/2420/09/24

Keywords

  • Causal inference
  • Explainable AI
  • Fine-grained visual classification

Fingerprint

Dive into the research topics of 'Counterfactual Contrastive Learning for Fine Grained Image Classification'. Together they form a unique fingerprint.

Cite this