Topology preserving non-negative matrix factorization for face recognition.

Area of Science:

• Computer Vision
• Machine Learning
• Pattern Recognition

Background:

• Non-negative Matrix Factorization (NMF) is widely used for dimensionality reduction.
• Traditional NMF methods may not effectively capture the intricate local topology of face patterns.
• Existing methods like PCA and LDA focus on Euclidean structures, potentially overlooking crucial manifold information.

Purpose of the Study:

• To propose a novel Topology Preserving Non-negative Matrix Factorization (TPNMF) method for improved face recognition.
• To leverage gradient distance for revealing latent manifold structures in high-dimensional face data.
• To enhance face representation by preserving local topological information such as edges and textures.

Main Methods:

• Derivation of the TPNMF model by incorporating local topology preservation into the original NMF framework.
• Minimization of constraint gradient distance in high-dimensional space to capture manifold structures.
• Transformation of high-dimensional face space into a local topology-preserving subspace via TPNMF decomposition.

Main Results:

• TPNMF effectively preserves local topology, including edges and textures, unlike PCA, LDA, and standard NMF.
• Experimental validation on three diverse databases with over 12,000 images demonstrates superior performance.
• The proposed TPNMF approach yields higher face recognition rates compared to the original NMF.

Conclusions:

• TPNMF offers a more robust representation of face patterns by preserving local topology.
• The gradient distance metric is effective in revealing underlying manifold structures for pattern recognition.
• TPNMF represents a significant advancement over existing methods for face recognition tasks.