Double Discrete Cosine Transform-Oriented Multi-View Subspace Clustering.

Area of Science:

• Machine Learning
• Computer Vision
• Data Mining

Background:

• Low-rank tensor representation using tensor nuclear norm is popular in multi-view subspace clustering (MVSC).
• Existing discrete Fourier transform (DFT)-based MVSC methods suffer from high tubal tensor rank due to complex arithmetic and neglect local structure.

Purpose of the Study:

• To propose a novel double discrete cosine transform (DCT)-oriented multi-view subspace clustering (D2CTMSC) method.
• To address limitations of DFT-based methods by avoiding complex arithmetic and incorporating local structural information.

Main Methods:

• Developed a D2CTMSC method utilizing two DCT transformations.
• The first DCT derives the tensor nuclear norm without complex arithmetic.
• The second DCT explores the local structure of the self-representation tensor.
• An alternating iteration strategy is employed to solve the proposed model.

Main Results:

• The D2CTMSC method effectively exploits low-rankness and sparsity in multi-view features.
• Experimental results on diverse datasets (News, Face, Scene, Generic Objects) show superior performance.
• D2CTMSC outperforms DFT-based and other state-of-the-art clustering methods.

Conclusions:

• The proposed D2CTMSC method offers a more effective approach to multi-view subspace clustering.
• DCT-based tensor nuclear norm and local structure exploration enhance clustering performance.
• D2CTMSC provides a promising alternative to existing DFT-based MVSC techniques.