Self-supervised learning method of image maskbased on enhanced chunk embedding andthreshold binarization

Abstract Self-supervised learning is popular for avoiding labeling large-scale datasets to reduce the cost, effectively improving the model's generalization and representation ability . Masking self-supervised learning is the hot spot of self-supervised learning nowadays, most of the existing masking methods are masking the regular chunks after data enhancement chunking, which is not the real sense of random masking, and at the same time, the above masking methods result in a poor local correlation between the chunks of the image. In this paper, we propose a new masking method for self-supervised learning, named TBMP, which masks the image after threshold binarisation before data enhancement, and enhances the chunked embedding between chunks while aggregating the input position information, effectively enhancing the local correlation between image chunks in the model, and solving the problem of mask shape region fixation in the current self-supervised learning. We use the DINO model as a benchmark for which we perform threshold binarisation masking as well as augmented chunk embedding. In this paper, we conduct linear evaluation experiments on several public datasets, and the results show that the TBMP method outperforms all existing mainstream models in self-supervised learning using masks.