Label contrastive learning for image classification

Image classification is one of the most important research tasks in computer vision. Current image classification methods with supervised learning have achieved good classification accuracy. However, supervised image classification methods mainly focus on the semantic differences at the class level, while lacking attention to the instance level. The core idea of contrastive learning is to compare positive and negative samples in the feature space to learn the feature representation, and the focus on instance-level information can make up for the lack of supervised learning. To this end, in this paper, we combine supervised learning and contrastive learning to propose labeled contrastive learning (LCL). Here, the supervised learning component ensures the distinguishability of different classes, the contrastive learning component enhances the compactness within classes and the separability between classes. In the contrastive learning component, instances with the same label are set as positive samples and instances with different labels are set as negative samples, which avoids the problem of false negative samples (positive samples are mislabeled as negative samples). Also, we applied a dynamic label memory bank and a momentum updated encoder. The experimental results show that LCL can further improve the accuracy of image classification compared with some supervised learning method.