BSNet: Bi-Similarity Network for Few-shot Fine-grained Image Classification

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Few-shot learning is crucial for fine-grained image classification.
• Metric-based methods are state-of-the-art but often rely on a single similarity measure.
• A single feature space may limit discriminative power in few-shot scenarios.

Purpose of the Study:

• To improve few-shot fine-grained image classification by learning more discriminative features.
• To address the limitations of single similarity measures in metric-based few-shot learning.
• To enhance model generalization ability in fine-grained image recognition tasks.

Main Methods:

• Proposed the Bi-Similarity Network (BSNet) incorporating a single embedding module and a bi-similarity module.
• Utilized two similarity measures with diverse characteristics to learn feature maps.
• Applied the BSNet to convolution-based embedding of support and query images.

Main Results:

• BSNet learns less similarity-biased and more discriminative features from limited data.
• The approach significantly improves model generalization ability.
• Experiments demonstrated substantial improvements on several fine-grained image benchmark datasets.

Conclusions:

• The Bi-Similarity Network effectively enhances few-shot fine-grained image classification.
• Employing multiple diverse similarity measures leads to more compact and discriminative feature spaces.
• The proposed method offers a promising direction for advancing metric-based few-shot learning.