Deep Non-Negative Matrix Factorization Architecture Based on Underlying Basis Images Learning.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Non-negative Matrix Factorization (NMF) represents images as linear combinations of basis images, aligning with human 'parts-to-whole' perception.
• Existing deep NMF methods factorize coefficient matrices, but basis images are derived from a single factorization of original images.

Purpose of the Study:

• To propose a novel deep NMF architecture for learning underlying basis images, enhancing feature extraction for deep localization characteristics.
• To develop interpretable deep factorization architectures for image representation.

Main Methods:

• Introduced a deep NMF architecture based on learning underlying basis images through deep factorization of the basis images matrix.
• Proposed a deep non-negative basis matrix factorization algorithm for obtaining underlying basis images.
• Developed a regularized deep non-negative basis matrix factorization algorithm with a regularization term to constrain basis images for local characteristics.
• Also proposed a regularized deep nonlinear NMF algorithm for complex pattern recognition tasks.

Main Results:

• The proposed deep NMF architecture effectively learns underlying basis images with good local characteristics.
• Theoretical convergence of the proposed algorithms was proven.
• Experimental results demonstrated superior recognition performance compared to state-of-the-art methods.

Conclusions:

• The novel deep NMF architecture based on underlying basis images learning offers improved image recognition.
• The proposed algorithms provide an interpretable and effective approach for feature extraction and pattern recognition in complex image data.