GoMIC: Multi-view image clustering via self-supervised contrastive heterogeneous graph co-learning

Abstract Graph learning is being increasingly applied to image clustering reveal intra-class and inter-class relationships in data. However, existing graph learning-based focuses on grouping images under a single view, which under-utilises the information provided by To address that, we propose self-supervised multi-view technique contrastive heterogeneous learning. Our method computes affinity for It conducts Local Feature Propagation (LFP) reasoning over local neighbourhood of each node executes an Influence-aware (IFP) from its influential intention. The proposed framework pioneeringly employs two objectives. first targets contrast fuse multiple views overall LFP embedding, second maximises mutual between IFP representations. We conduct extensive experiments benchmark datasets problem, i.e. COIL-20, Caltech7 CASIA-WebFace. evaluation shows that our outperforms state-of-the-art methods, including popular techniques MVGL, MCGC HeCo.