Robust Discriminant Subspace Learning With α-Divergence for Image Classification

Area of Science:

• Machine Learning
• Data Science
• Computer Vision

Background:

• Fisher Discriminant Analysis (FDA) is crucial for subspace learning.
• Classical FDA is sensitive to outliers, limiting its real-world applicability.
• Robustness in subspace learning is essential for reliable data analysis.

Purpose of the Study:

• To develop a novel robust Fisher Discriminant Analysis (FDA) method.
• To enhance discriminative subspace learning in the presence of data outliers.
• To provide a flexible and adaptive approach to outlier mitigation.

Main Methods:

• Proposed a robust FDA model based on the maximum-likelihood perspective and α-divergence.
• Introduced an adaptive redescending weighting scheme controlled by α.
• Implemented a two-fold iterative optimization for class-modeling and projection learning.

Main Results:

• The proposed method effectively suppresses the influence of outliers.
• Classical FDA is recovered when α = 1.
• Demonstrated superior performance over existing robust FDA variants on synthetic and image datasets.

Conclusions:

• The novel robust FDA offers an effective and efficient solution for subspace learning with outliers.
• The adaptive weighting scheme provides tunable robustness.
• The method shows significant improvements in handling contaminated data across various settings.