Fine-Grained Classification via Categorical Memory Networks.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Fine-grained visual classification (FGVC) requires learning subtle inter-class differences.
• Existing methods often struggle to effectively capture and leverage class-specific feature representations.
• Exploiting shared patterns across categories can enhance discriminative feature learning.

Purpose of the Study:

• To develop a novel class-specific memory module for improved fine-grained feature learning.
• To investigate the utility of combining category similarities as a discriminative cue.
• To enhance classification accuracy by augmenting original features with tailored response features.

Main Methods:

• Proposed a class-specific memory module storing prototypical feature representations as moving averages.
• Utilized an attention mechanism to query similarities with class prototypes.
• Combined original and attention-weighted prototype features to create an augmented feature representation.
• Integrated the module into a standard convolutional neural network (CNN), forming a Categorical Memory Network.

Main Results:

• The Categorical Memory Network significantly improved accuracy compared to baseline CNNs.
• Achieved competitive performance against state-of-the-art methods on four challenging benchmarks.
• Demonstrated the effectiveness of the class-specific memory module in fine-grained feature learning.

Conclusions:

• The proposed class-specific memory module is an effective approach for fine-grained feature learning.
• Attention-based combination of category prototypes provides a powerful discriminative signal.
• The method offers a simple yet effective way to enhance deep learning models for fine-grained classification tasks.