Bilaterally Normalized Scale-Consistent Sinkhorn Distance for Few-Shot Image Classification.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Few-shot image classification leverages limited data for recognizing novel classes.
• Current methods struggle with object and scale mismatches due to limited training examples.
• Misaligned positions and background clutter further complicate feature matching.

Purpose of the Study:

• To address object and scale mismatch issues in few-shot image classification.
• To propose a novel approach enhancing feature comparison between support and query images.
• To improve the accuracy and robustness of few-shot learning models.

Main Methods:

• Introduced the bilaterally normalized scale-consistent Sinkhorn distance (BSSD).
• Utilized Sinkhorn distance for optimal image matching, mitigating positional misalignment.
• Incorporated intra-image and inter-image attentions for background clutter robustness.
• Employed multiscale pooling to enhance local features for scale consistency.

Main Results:

• The proposed BSSD method effectively resolves object and scale mismatch issues.
• Achieved state-of-the-art performance on three standard few-shot image classification benchmarks.
• Demonstrated significant improvements in recognizing unseen classes with few examples.

Conclusions:

• BSSD offers a robust solution for few-shot image classification challenges.
• The method enhances transferable feature learning by addressing feature mismatch problems.
• This work advances the field by improving accuracy in low-data regimes.