Information maximization clustering via multi-view self-labelling

Image clustering is a particularly challenging computer vision task, which aims to generate annotations without human supervision. Recent advances focus on the use of self-supervised learning strategies in image clustering, by first valuable semantics and then representations. These multiple-phase algorithms, however, involve several hyper-parameters transformation functions, are computationally intensive. By extending grouping based approach, this work proposes novel single-phase method that simultaneously learns meaningful representations assigns corresponding annotations. This achieved integrating discrete representation into paradigm through classifier net. Specifically, proposed objective employs mutual information maximise dependency integrated probability distribution. The distribution derived means process compares learnt latent with set trainable prototypes. To enhance performance classifier, we jointly apply across multi-crop views. Our empirical results show framework outperforms state-of-the-art techniques an average accuracy 89.1%, 49.0%, 83.1%, 27.9%, respectively, baseline datasets CIFAR-10, CIFAR-100/20, STL10 Tiny-ImageNet/200. Finally, also demonstrates attractive robustness parameter settings, large number classes, making it ready be applicable other datasets.