Deep robust multi-channel learning subspace clustering networks

Abstract

Subspace clustering methods are now widely used for unsupervised high-dimensional data processing in computer vision and other domains. Deep subspace clustering methods based on auto-encoder networks have made a significant improvement in nonlinear subspace clustering problems in comparison to previous works. However, these methods ignore the valid information lost during feature extraction, resulting in incomplete information and imprecise feature representations for subspace clustering. In addition, the clustering performance of the existing clustering methods is excessively dependent on hyper-parameters, making training difficult and unstable. In this paper, we propose Deep Robust Multi-Channel Learning Subspace Clustering Networks (DRMCLSC), a novel deep subspace clustering network for learning more comprehensive feature representations with good robustness for subspace clustering. The multi-channel learning strategy allows the model to extract, retain and fuse features simultaneously, enabling all valid information from the sample data to be obtained. Moreover, the multi-channel learning structure of the proposed method produces a more stable integration network that is less dependent on hyper-parameters and more resistant to training errors than previous works. Extensive experimental results on four benchmark datasets demonstrate the proposed method is superior and more effective than the state-of-the-art subspace clustering methods.