Multiple Contrast Clustering Network

In this paper, we propose a deep clustering framework, called Multiple Contrast Clustering Network (MCC-Net), which is used to learn similarity representations and implement clustering of samples. Based on the framework proposed by Contrastive Clustering (CC), MCC-Net generates positive and negative class instance pairs by contrast instance generation network (CIG) and learns feature representations of these instance pairs. Then, these features are mapped to another feature space by the instance contrastive learning network (ICL) and cluster contrastive learning network (CCL), respectively, and the positive pairs are clustered while the negative pairs are dispersed. Compared with the CC framework, MCC-Net performs more accurate positive and negative instance sample matching and thus more effective feature representation by introducing negative class instance samples. In addition, this paper proposes a new contrast loss function that exploits the similarity and difference of feature characteristics in instance and cluster contrastive learning to effectively reduce the distance between similar samples and increase the distance between negative samples. The experimental results show that the clustering accuracy of the method on the computer vision benchmark dataset is improved than CC by 24.2%, 24.3%, 5.7%, and 7.7% on CIFAR-10, CIFAR-100, STLIO and ImageNet-10, respectively.