Enhancing Multi-View Deep Image Clustering via Contrastive Learning for Global and Local Consistency

Multi-view clustering (MVC) is a data clustering method with many applications, including but not limited to image and video analysis, text and language processing, bioinformatics, and signal processing. The objective of multi-view deep clustering is to enhance the efficacy of clustering algorithms by integrating data from disparate views. However, discrepancies and inconsistencies between different views frequently reduce the precision of the clustering outcomes. In the recent popular comparative learning, it has been observed that the processing of positive and negative samples does not consider the multi-view consistency information, ultimately resulting in a decline in clustering accuracy. In this paper, we put forth a global and local consistency-based contrast learning framework to enhance the efficacy of multi-view deep clustering. First, a global consistency constraint is designed to ensure that the global representations of different views can be aligned to capture the data's main features. Secondly, we introduce a local consistency mechanism, which aims to preserve the unique local information in each view and obtain efficient, positive samples to improve the complementarity and robustness of the inter-view representations through comparative learning. The experimental results demonstrate that the proposed method markedly enhances the clustering performance on several real benchmark datasets, mainly when dealing with multi-view data with incompleteness.