GATCluster: Self-supervised Gaussian-Attention Network for Image Clustering

We propose a self-supervised Gaussian ATtention network for image Clustering (GATCluster). Rather than extracting intermediate features first and then performing traditional clustering algorithms, GATCluster directly outputs semantic cluster labels without further post-processing. We give a Label Feature Theorem to guarantee that the learned features are one-hot encoded vectors and the trivial solutions are avoided. Based on this theorem, we design four self-learning tasks with the constraints of transformation invariance, separability maximization, entropy analysis, and attention mapping. Specifically, the transformation invariance and separability maximization tasks learn the relations between samples. The entropy analysis task aims to avoid trivial solutions. To capture the object-oriented semantics, we design a self-supervised attention mechanism that includes a Gaussian attention module and a soft-attention loss. Moreover, we design a two-step learning algorithm that is memory-efficient for clustering large-size images. Extensive experiments demonstrate the superiority of our proposed method in comparison with the state-of-the-art image clustering benchmarks.