Hierarchical multi-view image clustering

Image data have gained significant attention due to its extensive application in real-world scenarios. However, existing multi-view image clustering methods often encounter critical limitations, such as inadequate consideration of denoising and redundancy reduction, as well as limited utilization of multi-order approaches for information extraction and integration. These issues are further exacerbated when dealing with large-scale image data, resulting in inferior clustering results. To address these issues, we propose Hierarchical Multi-view Image Clustering, a framework designed to enhance clustering quality by incorporating a series of advanced modules. Specifically, we introduce a Kernel Laplacian-Based Denoising and Redundancy Reduction Module that refines the initial bipartite graphs by mitigating noise and removing redundancy, thereby preserving essential structural patterns. Building on this, we present a Bi-Order Bipartite Graph Learning Framework that systematically captures structural relationships across multiple levels. The algorithm facilitates sample-level aggregation and discrimination by integrating an Adaptive Feature Fusion Module to maintain structural coherence and a Discriminative Learning Module to enhance the separability of different views by learning view-specific information. Simultaneously, the module uncovers intricate interview relationships through the application of low-rank constraints and the Schatten-p norm, fostering a more comprehensive and effective multi-view information fusion process. Evaluations on five image datasets show that our method outperforms state-of-the-art approaches, achieving superior clustering results and demonstrating the robustness and effectiveness of the proposed framework.