Low-Rank Preserving Projections.

Area of Science:

• Computer Vision
• Pattern Recognition
• Machine Learning

Background:

• Locality Preserving Projections (LPP) is a popular dimensionality reduction technique.
• Real-world data often contains noise, which can degrade LPP's performance.
• Noisy data may cause samples from the same class to be non-adjacent, reducing LPP effectiveness.

Purpose of the Study:

• To propose a novel dimensionality reduction method, Low-Rank Preserving Projections (LRPP), for image classification.
• To address the challenge of gross data corruption with sparse noise.
• To develop a method that preserves the global data structure despite noise.

Main Methods:

• LRPP learns a low-rank weight matrix by projecting data onto a low-dimensional subspace.
• Utilizes the L21 norm as a sparse constraint on the noise matrix.
• Employs the nuclear norm as a low-rank constraint on the weight matrix.

Main Results:

• LRPP effectively reduces the disturbance of noise in dimensionality reduction.
• The method learns a robust subspace from corrupted data.
• Experimental results demonstrate the effectiveness and feasibility of LRPP compared to state-of-the-art methods.

Conclusions:

• LRPP is a robust dimensionality reduction technique for noisy image data.
• The method maintains data integrity and improves classification performance.
• LRPP offers a promising solution for image classification in the presence of significant noise.