Abstract: In order to improve the accuracy of deep learning network, it is often necessary to use a large number of labeled samples for training, and sample labeling work consumes a lot of manpower and material resources. In response to the above problems, a small sample classification method is proposed, which combines the feature extraction module of swin transformer with contrastive learning. To a certain extent, it reduces the use of labeled samples during the training process while maintaining a high classification accuracy. The established model is divided into two stages:the self-supervised learning stage and the supervised learning stage. During the self-supervised learning stage, the model learns the commonalities among similar defects and the characteristics among different types of defects. After the learning stage, it is readjustment to distinguish various defect types of leather. The experimental results show that this method can reduce labeled training samples to a certain extent and maintain a high classification accuracy, and has great application potential in the field of industrial inspection.