Semantic-consistency multi-view deep subspace clustering network with frequency branches

Multi-view subspace clustering (MvSC) can utilize high-dimensional data from different perspectives of the same target to segment them into multiple low-dimensional subspaces. However, existing methods primarily focus on the spatial information of the data, often neglecting frequency information that typically contains crucial features for data identification and characterization. Additionally, retaining sufficient consistent information across different views while removing view-specific redundant information is a significant challenge in MvSC. In this paper, we propose a semantic-consistency multi-view deep subspace clustering network to address these issues. The proposed model endows an encoder with a frequency branch to capture both spatial and frequency domain information, enriching hidden layer features. To obtain semantically consistent representations across views, we employ a feature integration module with mutual information maximization to enable the model to learn a better self-representation matrix. In addition, we introduce a multi-view information bottleneck loss to suppress unique information from individual views, thereby improving clustering performance. Our experiments demonstrate the effectiveness of the proposed model, showing superior performance compared to mainstream methods.