Maximum Entropy Subspace Clustering Network

Deep subspace clustering networks have attracted much attention in clustering, which an auto-encoder non-linearly maps the input data into a latent space, and fully connected layer named self-expressiveness module is introduced to learn affinity matrix via typical regularization term (e.g., sparse or low-rank). However, adopted terms ignore connectivity within each subspace, limiting their performance. In addition, framework suffers from coupling issue between module, making network training non-trivial. To tackle these two issues, we propose novel deep method Maximum Entropy Subspace Clustering Network (MESC-Net). Specifically, MESC-Net maximizes entropy of promote its elements corresponding same are uniformly densely distributed. Furthermore, design explicitly decouple module. We also theoretically prove that learned satisfies block-diagonal property under independent subspaces. Extensive quantitative qualitative results on commonly used benchmark datasets validate significantly outperforms state-of-the-art methods.