A migration learning based binary classification method for small datasets

With the rapid development of digital technology, people increasingly use images to express and convey information, and image classification, as a core problem in the image field, has shown its great potential and value. Therefore, the use of deep learning techniques to achieve accurate and fast classification of images is a problem that needs to be solved urgently nowadays. However, deep learning methods at the current stage rely too much on huge data samples and detailed data annotations. In the above context, this paper proposes a binary classification method based on migration learning for small datasets. In the construction of the migration model, the Swin MAE model is adopted for the upstream task, and the strategy of the Swin Transformer model is adopted for the downstream task with global identification. Comparing with the mainstream deep learning network at the current stage, the experimental results show that on the relevant small dataset, the method proposed in this paper has a better classification effect, with an accuracy of 97.98%, a sensitivity of 96.2%, and a specificity of 95.6%, which is a certain advantage compared with the mainstream deep learning classification method at the current stage.