Multi-Objective Matrix Normalization for Fine-grained Visual Recognition.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Bilinear pooling is effective for fine-grained visual recognition (FGVC).
• Existing matrix power normalization methods stabilize second-order information but face issues like redundancy and overfitting.
• There is a need for improved normalization techniques in FGVC.

Purpose of the Study:

• To propose an efficient Multi-Objective Matrix Normalization (MOMN) method.
• To simultaneously normalize bilinear representations using square-root, low-rank, and sparsity regularizers.
• To stabilize second-order information, compact features, and enhance model generalization.

Main Methods:

• Developed MOMN to jointly optimize three non-smooth regularizers.
• Formulated regularization into an augmented Lagrange formula with approximated constraints.
• Introduced auxiliary variables and alternating optimization for efficient solving.
• Employed gradient descent strategies for convergence and implementation.

Main Results:

• MOMN effectively stabilizes and discriminates bilinear features.
• The method achieves superior accuracy and efficiency compared to existing normalization techniques.
• Experiments on five public FGVC benchmarks validate the proposed approach.

Conclusions:

• MOMN offers an efficient and effective solution for normalizing bilinear representations in FGVC.
• The proposed method addresses limitations of previous normalization techniques.
• MOMN enhances model performance through feature stabilization and compaction.