Abstract
Semi-supervised learning mitigates the problem of labeled data scarcity by utilizing unlabeled data, but the generalization performance of existing methods usually degrades significantly when the unlabeled data is small in size or poor in quality. To this end, this paper proposes a semi-supervised image classification method based on multi-mode augmentation, which mitigates the effects of insufficient quality and limited scale of unlabeled data by simultaneously improving the sample completeness within and between classes. Specifically, the model's prediction confidence and bias are used for uncertainty-based screening to improve pseudo-label quality, while retaining as many unlabeled samples as possible to fully exploit their potential information. Secondly, a multi-modal data augmentation strategy combining intra-class random augmentation and inter-class mixed augmentation is designed to enhance the diversity of the data and the feature expression capability. Finally, a pseudo-label consistency metric is introduced to further improve the model's generalization ability. The experimental results on STL-10 and CIFAR-10 datasets show that the generalization performance of the proposed method is significantly better than the existing mainstream methods in the scenarios of small unlabeled data and mismatched samples.