LSPC: Exploring contrastive clustering based on local semantic information and prototype

Recently years, several prominent contrastive learning algorithms, a kind of self-supervised learning methods, have been extensively studied that can efficiently extract useful feature representations from input images by means of data augmentation techniques. How to further partition the representations into meaningful clusters is the issue that deep clustering is addressing. In this work, a deep clustering algorithm based on local semantic information and prototype is proposed referring to LSPC that aims at learning a group of representative prototypes. Rather than learning the distinguishing characteristics between different images, more attention is given to the essential characteristics of images that are maybe from a potential category. On the training framework, contrastive learning is skillfully combined with k-means clustering algorithm. The prediction is transformed into soft assignments for end-to-end training. In order to enable the model to accurately capture the semantic information between images, we mine similar samples of training samples in the embedded space as local semantic information to effectively increase the similarity between samples belonging to the same cluster. Experimental results show that our algorithm achieves state-of-the-art performance on several commonly used public datasets, and additional experiments prove that this superior clustering performance can also be extended to large datasets such as ImageNet.