A group-based siamese self-supervised learning method

Abstract

The application discloses a kind of Siamese self-supervised learning methods based on group, the method uses Siamese network structure, two relevant views after data enhancement are randomly cropped into N fixed size overlap image blocks, after data enhancement, image block is respectively input to feature encoder and obtains multiple characteristic values.The feature encoder is composed of a ResNet-18 network and two linear layers.Calculate cross-correlation matrix loss using the feature set generated by the two branches of the Siamese network, and use rank-k triplet loss to avoid misclassification of two samples from the same class as negative feature pair groups.This method can solve the limitations of single-view features in self-supervised learning, and can achieve significant convergence in a small number of iteration rounds, making the pre-trained ResNet-18 have good classification effect, and outperforming the state-of-the-art self-supervised learning techniques on multiple datasets.

Images