Canonical Correlation Analysis With Low-Rank Learning for Image Representation

Area of Science:

• Computer Vision and Pattern Recognition
• Multivariate Data Analysis

Background:

• Canonical Correlation Analysis (CCA) is a multivariate data analysis tool widely used in computer vision and pattern recognition.
• Traditional CCA is sensitive to noise and outliers due to its reliance on Euclidean distance.
• CCA requires identical training set sizes, limiting its practical application.

Purpose of the Study:

• To develop novel canonical correlation learning methods that address the limitations of traditional CCA.
• To enhance image representation through robust and low-rank learning techniques.
• To enable CCA methods to handle training datasets with different numbers of samples.

Main Methods:

• Proposed two new methods: robust canonical correlation analysis (robust-CCA) and low-rank representation canonical correlation analysis (LRR-CCA).
• Robust-CCA employs low-rank learning to denoise data and extract maximal correlation features, using nuclear and L1 norms as constraints.
• LRR-CCA integrates low-rank representation into CCA to ensure correlative features are derived from a low-rank subspace.

Main Results:

• The proposed methods successfully overcome CCA's sensitivity to noise and sample size limitations by introducing regular matrices.
• Experiments on five public image databases demonstrate superior performance compared to existing CCA-based and low-rank learning methods.
• Robust-CCA effectively extracts correlation features from cleaned data, while LRR-CCA ensures correlative features are captured in a low-rank representation.

Conclusions:

• Robust-CCA and LRR-CCA offer significant improvements over traditional CCA for image representation tasks.
• The novel methods provide a more robust and flexible approach to canonical correlation learning.
• These advancements have the potential to advance applications in computer vision and pattern recognition.