Deep subspace clustering using dual self-expressiveness and convolutional fusion

Deep subspace clustering methods use deep neural networks to project input data into the latent space, leveraging the inherent self-expressiveness (SE) properties of the data as a similarity metric to handle high-dimensional data effectively. However, existing methods focus solely on the SE relationships within the latent space, which constrains their capacity to capture subspace structures. To overcome this limitation, we introduce a novel deep subspace clustering method using dual self-expressiveness and convolutional fusion (DSCDC), which computes SE relationships in both the latent and input spaces. This dual-focus approach captures multi-source SE relationships, enhancing the quality of the SE matrix. Additionally, we designed a convolutional fusion module that effectively integrates the multiple SE matrices through a learnable fusion approach. Experimental results across various datasets validate the superiority of our DSCDC compared to competing methods. Ablation studies further confirm the effectiveness of the proposed modules.