Learnable low-rank latent dictionary for subspace clustering

Highlights：

• We propose a unified framework to integrate linear or nonlinear feature extraction and subspace clustering.

• Adopting a linear or nonlinear projection strategy can be viewed as linear or nonlinear dimension reduction since the strategy satisfies that the dimension of the latent dictionary is lower than the dimension of the observed original data.

• The neural network is utilized to implement the projection transformation for improving the nonlinear coding ability, and also guarantees that the extracted features have a low-rank structure. This is the first time to combine the traditional low-rank coding with the neural network for learning the latent dictionary in subspace clustering, which not only can effectively extract features but also guarantee the data structure of extracted features.