Discriminant Incoherent Component Analysis.

Area of Science:

• Computer Vision
• Machine Learning
• Pattern Recognition

Background:

• Facial images contain complex information, separable into low-complexity components related to identity, expressions, and facial action units (AUs).
• Low-rank components often represent identity, while sparse components capture dynamic changes like expressions and AU activations.

Purpose of the Study:

• To propose Discriminant Incoherent Component Analysis (DICA) for extracting mutually incoherent, low-complexity components corresponding to facial attributes.
• To develop a method robust to gross sparse errors in training data for attribute extraction.

Main Methods:

• Formulated an optimization problem minimizing nuclear- and l1-norms to find class-specific incoherent components.
• Expressed unseen images as group-sparse linear combinations of these components to reveal attribute classes.

Main Results:

• DICA successfully extracts discriminative facial attribute components.
• The method demonstrates superior performance across various face analysis tasks, including joint recognition and AU detection.
• Achieved state-of-the-art results even with corrupted training data.

Conclusions:

• DICA is an effective method for extracting facial attributes and performing complex face analysis.
• The proposed approach offers robust performance in challenging conditions, outperforming existing methods.