Robust Representation Learning for Image Clustering

Unsupervised image clustering is a challenging task with many practical applications, especially when data labels are scarce or costly to obtain. To address this problem, we propose a novel deep clustering method, named Robust Representation Learning Clustering (RRLC), that leverages contrastive learning to learn discriminative and robust representations for complex visual data. RRLC measures the similarities among both instances and features, and encourages the representations of similar instances to be close in the feature space. Moreover, It adopts a self-enhancing contrastive learning strategy to enhance the robustness of the representations by maximising the agreement between different augmented views of the same instance. We conduct extensive experiments on several image benchmark datasets and show that RRLC outperforms existing state-of-the-art deep clustering methods. The representations learned by RRLC can significantly boost the performance of traditional clustering algorithms on various images.