Robust augmentation-based contrastive clustering with negative data mining

Contrastive learning has been widely applied in unsupervised clustering tasks due to its superior performance in learning discriminative representations. However, the data augmentations commonly used in current contrastive clustering methods tend to combine weak and strong augmentations, which make it challenging to produce truly hard augmented examples due to the property of linear transformations. In addition, the simplicity of negative sample setting in contrastive learning generally leads to false positive samples, which aggravates the clustering performance. To alleviate these concerns, we propose a contrastive clustering method based on robust augmentation and negative data mining. Firstly, a robust augmentation method is introduced, which combines nonlinear augmentations with traditional linear augmentation methods to form contrastive pairs, to compensate for the limitations of linear augmentation in capturing complex features. Subsequently, we propose a hypothesis, i.e., if numerous samples of one class which widely separated from its cluster prototype are closest to the other cluster prototype, the two classes are considered to be the visual confusion categories (VCCs). Obviously, VCCs are particularly prone to boundary overlap in the clustering process. Therefore, we employ a confidence estimation criterion to obtain pseudo labels and increase the weight of negative samples/categories with VCCs labels, thereby enhancing the quality of clustering decision boundaries. Comprehensive experiments performed on four commonly used image benchmarks show that the proposed method is efficient.