A model with pseudo-label correction and distribution alignment for image clustering

Abstract

Self-supervised learning has recently created a surge of research interest in image clustering with the help of pseudo-labels. However, most of methods have to face the challenge of dealing with noisy labels since the noisy pseudo-labels might provide erroneous guidance for model training and deteriorate the clustering performance. To solve this problem, a Pseudo-label Correction and Distribution Alignment-based deep Clustering (PCDAC) model is proposed in this paper. To correct the noisy pseudo-labels, PCDAC embeds the historical cluster assignment by maintaining a memory bank. Additionally, a distribution alignment operation is employed to impose the cluster probability distribution of strong-augmented instances close to those of weak-augmented instances, to reduce the possibility of generating incorrect pseudo-labels at the initial training phase. Extensive comparisons experiments are conducted and the results illustrate that PCDAC is superior to the SOTA methods. The code is available on GitHub at https://github.com/LiuZiweiAI/PCDAC.