.png){kind=icon}
.png){kind=icon}
Return
Joint representation and classifier learning for long-tailed image classification
.png){kind=icon}
PRE
AI
DOI：10.1016/j.imavis.2023.104759.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Long-tailed classification with fine-grained appearance, e.g., in chest X-ray images, is very challenging due to the very similar appearance and imbalanced distribution between normal and abnormal samples, which extremely limits the ability of deep networks to learn powerful representations and discriminative classifiers. In this paper, we propose a novel Joint Representation and Classifier Learning (JRCL) framework to achieve the above purposes, simultaneously. In terms of representation learning, we propose a One-to-All supervised contrastive learning strategy to avoid the medium or tail classes mixing in the head classes. For the classifier cleaning, we propose a novel Binary Distribution Consistency (BDC) loss to learn a discriminative classifier that could separate the normal and abnormal samples.The BDC loss measures the binary distribution consistency between the designed multi-class classifier and an auxiliary binary classifier. Consequently, the JRCL framework is optimized with a supervised contrastive learning loss, a binary distribution consistency loss, and a multi-classification loss. We conduct experiments on large-scale, long-tail image datasets, NIH-CXR-LT, MIMIC-CXR-LT, iNaturalist 2018, and Places-LT. Experimental results demonstrate JRCL could improve the discriminate ability of the imbalanced data and thus obtain better classification performance. Compared with the state-of-the-art methods, our proposed JRCL achieves comparable or even better performance. The source codes are available at https://github. com/guanqj932/JRCL.
Keywords:
Long-tailed image classification
Representation learning
Classifier learning
Supervised contrastive learning
Journal
IF:
4.2
Papers:
4.0K
Citations:
6.7K