DeepEMD: Differentiable Earth Mover's Distance for Few-Shot Learning.

Area of Science:

• Computer Vision
• Machine Learning
• Pattern Recognition

Background:

• Few-shot image classification is challenging due to limited training data.
• Existing methods struggle with background clutter and intra-class variations.
• Structural information in images is crucial for accurate classification.

Purpose of the Study:

• To develop a novel few-shot image classification method based on optimal region matching.
• To introduce Earth Mover's Distance (EMD) as a metric for structural image comparison.
• To improve classification accuracy by effectively handling background clutter and appearance variations.

Main Methods:

• Utilizing Earth Mover's Distance (EMD) to compute structural distances between dense image representations.
• Designing a cross-reference mechanism to generate element weights for EMD, mitigating background and intra-class issues.
• Proposing a structured fully connected layer for direct classification using EMD, enabling end-to-end network training.

Main Results:

• The proposed method significantly outperforms state-of-the-art approaches on five benchmark datasets: miniImageNet, tieredImageNet, FC100, CUB, and CIFAR-FS.
• Demonstrated effectiveness in few-shot image classification tasks.
• Showcased applicability and success in image retrieval tasks.

Conclusions:

• The optimal matching perspective using EMD offers a powerful new direction for few-shot image classification.
• The developed method effectively addresses challenges posed by limited data, background clutter, and intra-class variations.
• This approach provides a robust and accurate solution for few-shot learning and image retrieval applications.