Learning deep discriminative representations with pseudo supervision for image clustering

Abstract

Image clustering is a crucial but challenging task in machine learning and computer vision. Its performance highly depends on the quality of image feature representations. Recently, deep joint clustering which combines representation learning with clustering has presented a promising performance. However, existing joint methods suffer from two severe problems. That is, the learned representations lack discriminability especially for intricate images, and the performance often encounters a bottleneck due to the lack of supervision information. To address these problems, we propose a pseudo-supervised joint method for image clustering, i.e., Discriminative Pseudo Supervision Clustering (DPSC). Our key idea is to discover and utilize the pseudo supervision information to provide supervisory guidance for discriminative representation learning. With the aid of pseudo supervision, the representations can be continuously refined to facilitate inter-cluster separability and intra-cluster compactness, thereby leading to more discriminative representations and correctly separated clusters. To fully benefit from joint learning, we further introduce a self-evolution training algorithm to jointly optimize the DPSC model, in which the learned representations and clustering results boost each other progressively as more reliable pseudo supervision information is discovered during the iteration. Experimental results show that DPSC significantly outperforms state-of-the-art methods on various image datasets. Moreover, the learned feature representations generalize well across various algorithms.