Redundancy-Aware External Signal Learning for Image Clustering
提出冗余感知外部信号学习框架，利用信息瓶颈机制压缩预训练模型文本特征，去除冗余语义，提升图像聚类性能，并设计多簇头结构解决跨模态语义不一致问题。
The emergence of contrastive language–image pretraining (CLIP)-like models has significantly advanced image clustering by providing additional semantic information to guide the clustering process. However, existing approaches often incorporate redundant semantic features, which may adversely affect clustering performance. To address this issue, we propose a redundancy-aware external signal learning (RESL) framework based on the information bottleneck (IB) mechanism to compress and refine text features provided by large pretrained models. By leveraging the IB principle, RESL dynamically strips task-irrelevant and redundant semantics, significantly enhancing feature quality for clustering tasks. Furthermore, we design a multicluster-head architecture to address the issue of cross-modal semantic inconsistency. Distinct cluster heads collaboratively enhance each other through a cross-head distillation mechanism, enabling reliable semantic alignment while preserving modality-specific characteristics. Extensive experimental results show that RESL not only outperforms traditional methods and existing advanced methods based on external signals on multiple public datasets but also further verifies the significant advantages of the proposed framework in effectively filtering redundant information through redundancy analysis and clustering visualization. Its demonstrated ability to handle numerous fine-grained categories and high-resolution images makes it a promising approach for large-scale, real-world clustering scenarios where the data are complex and unlabeled. The source code is available athttps://github.com/Rhett-Lin/RESL