Fine-Grained 3D-Attention Prototypes for Few-Shot Learning.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Limited labeled data in fine-grained image classification hinders effective class representation.
• Fine-grained images exhibit subtle inter-class variations and larger intra-class variations, posing classification challenges.

Purpose of the Study:

• To propose an end-to-end attention-based model for fine-grained few-shot image classification (AFG).
• To address the limitations of insufficient labeled data and subtle visual differences in fine-grained datasets.

Main Methods:

• Developed an attention-based model (AFG) utilizing an episode training strategy.
• Incorporated a 3D-Attention mechanism in the feature learning module to capture spatial and channel attention.
• Employed an image reconstruction module with a novel loss function for auxiliary supervised learning and a label distribution module for classification.

Main Results:

• The proposed AFG model achieved superior performance compared to existing methods on Mini-ImageNet and three fine-grained datasets.
• The 3D-Attention mechanism effectively learned discriminative local features, enhancing class representation.
• The auxiliary supervised learning approach reduced the need for extra annotations.

Conclusions:

• The AFG model offers an effective solution for fine-grained few-shot image classification.
• Attention mechanisms and auxiliary learning strategies are crucial for improving performance with limited data in fine-grained recognition.