Self-supervised deep subspace clustering with entropy-norm

Guangyi Zhao, Simin Kou, Xuesong Yin, Guodao Zhang, Yigang Wang

Abstract: Auto-Encoder based Deep Subspace Clustering (DSC) has been widely applied in computer vision, motion segmentation and image processing. However, existing DSC methods suffer from two limitations: (1) they ignore the rich useful relational information and the connectivity within each subspace due to the reconstruction loss; (2) they design convolutional networks individually according to specific datasets. To address the above problems and improve the performance of DSC, we propose a novel algorithm called Self-Supervised deep Subspace Clustering with Entropy-norm(S\(^{3}\)CE) in this paper. Firstly, S\(^{3}\)CE introduces self-supervised contrastive learning to pre-train the encoder instead of requiring a decoder. Besides, the trained encoder is used as a feature extractor to segment subspace by combining self-expression layer and entropy-norm constraint. This not only preserves the local structure of data, but also improves the connectivity between data points. Extensive experimental results demonstrate the superior performance of S\(^{3}\)CE in comparison to the state-of-the-art approaches.