Deep Image Clustering Based on Label Similarity and Maximizing Mutual Information across Views

Abstract

Most existing deep image clustering methods use only class-level representations for clustering. However, the class-level representation alone is not sufficient to describe the differences between images belonging to the same cluster. This may lead to high intra-class representation differences, which will harm the clustering performance. To address this problem, this paper proposes a clustering model named Deep Image Clustering based on Label Similarity and Maximizing Mutual Information Across Views (DCSM). DCSM consists of a backbone network, class-level and instance-level mapping block. The class-level mapping block learns discriminative class-level features by selecting similar (dissimilar) pairs of samples. The proposed extended mutual information is to maximize the mutual information between features extracted from views that were obtained by using data augmentation on the same image and as a constraint on the instance-level mapping block. This forces the instance-level mapping block to capture high-level features that affect multiple views of the same image, thus reducing intra-class differences. Four representative datasets are selected for our experiments, and the results show that the proposed model is superior to the current advanced image clustering models.

Keywords

• image clustering
• extended mutual information
• unsupervised learning