Self-supervised learning can assist neural network models in automatically learning knowledge and patterns from a large amount of unlabeled data. Recently, self-supervised learning has garnered significant attention and has made substantial progress. Contrastive learning, as an effective method within self-supervised learning, enhances the model’s ability to model training images by increasing inter-class distances and decreasing intra-class distances. In this work, we propose a novel self-supervised deep learning paradigm with similarity-based sample judgement to choose reliable positive/negative samples, thereby avoiding cases where samples of the same class are mistakenly classified as negative sample. Specifically, we use a teacher-student network to generate multi-view features for self-supervised contrastive learning and make robust positive/negative sample judgments based on the similarity information between samples. Finally, we conduct training based on the results of reliable sample judgments. Extensive experimental results demonstrate that the proposed method achieves satisfactory performance in terms of classification accuracy on the ImageNet dataset and various downstream tasks.