Semantic-Enhanced Image Clustering

Image clustering is an important and open challenging task in computer vision. Although many methods have been proposed to solve the image task, they only explore images uncover clusters according features, thus being unable distinguish visually similar but semantically different images. In this paper, we propose investigate of with help visual-language pre-training model. Different from zero-shot setting, which class names are known, know number setting. Therefore, how map a proper semantic space cluster both spaces two key problems. To above problems, novel method guided by model CLIP, named Semantic-Enhanced Clustering (SIC). new method, given first efficient generate pseudo-labels relationships between semantics. Finally, perform consistency learning space, self-supervised fashion. The theoretical result convergence analysis shows that our can converge at sublinear speed. Theoretical expectation risk also reduce improving neighborhood consistency, increasing prediction confidence, or reducing imbalance. Experimental results on five benchmark datasets clearly show superiority method.