Pseudo-supervised image clustering based on meta-features

Abstract Stable semantics is a prerequisite for achieving excellent image clustering. However, most current methods suffer from inaccurate class semantic estimation, which limits the clustering performance. For sake of addressing issue, we propose pseudo-supervised framework based on meta-features. First, mines meta-semantic features (i.e., meta-features) categories instance-level features, not only preserves information but also ensures robustness Ulteriorly, propagate pseudo-labels to its global neighbor samples with meta-features as center, effectively avoids accumulation errors caused by misclassification at cluster boundary. Finally, exploit cross-entropy loss label smoothing optimize pseudo-label optimization network. This method achieves direct mapping stable labels, suboptimal solutions multi-level optimization. Extensive experiments demonstrate that our significantly outperforms twenty-one competing six challenging datasets.