Improve Contrastive Clustering Performance by Multiple Fusing-Augmenting ViT Blocks

The recent study on improving image clustering performance introduces a novel method utilizing multiple fusing-augmenting ViT blocks (MFAVBs). Traditional contrastive learning networks often fail to fully leverage the complementarity of positive pairs, which this new approach addresses by explicitly fusing features from these pairs. By feeding augmented positive pairs into shared-weight Vision Transformers (ViTs) and subsequently fusing their outputs, the method enhances feature extraction. This innovation is crucial as it aims to maximize the similarity between positive pairs while minimizing the dissimilarity of negative pairs, potentially leading to significant improvements in clustering performance. The reliance on the excellent feature learning capabilities of Vision Transformers underlines the method's promise in advancing the field of image clustering.