Overcomplete Deep Low-Rank Subspace Clustering

Abstract

Aiming at the fact that when the input data in the deep subspace clustering networks (DSC) has noise, its robustness is poor, the performance is significantly degraded, and the method has too many learnable parameters, we suggest an overcomplete deep low-rank subspace clustering (ODLRSC). The technique is easy to use, efficient, and has shown to be a great fit for subspace clustering. By inserting a fully connected linear layer and its transposition between the encoder and decoder in our suggested technique, we may automatically put rank restrictions on the learnt representations. Additionally, in order to obtain a more reliable representation of the input data for clustering, the characteristics of the under- and over-complete auto-encoder networks are fused in the encoder. Our technique beats DSC and other clustering algorithms in the field of clustering error, and can sustain high level of quality throughout a broad range of LRRs, according to experimental findings on benchmark datasets.