Manuscript received March 11, 2025; revised April 25, 2025; accepted June 17, 2025; published October 17, 2025.

Abstract—The core challenge of few-shot image classification lies in efficiently learning and accurately classifying with limited labeled data. Current metric-based meta-learning methods primarily rely on computing the structural distance between the query and support sets for matching. However, traditional metric learning typically employs Euclidean distance or standard Sinkhorn Distance (SD) for feature matching, assuming that the total mass of the source and target distributions is equal. This assumption overlooks the imbalance in data distribution in real-world tasks. To overcome these challenges, this study systematically introduces Unbalanced Sinkhorn Distance (USD) into Few- Shot Learning (FSL) to enhance the model’s ability to match features map of query set and support set under imbalanced distributions. USD allows dynamic adjustment of matching distributions during metric computation. Moreover, this method effectively reduces the interference of background noise when matching features between the support and query sets while maintaining low computational costs. Experimental results demonstrate that our method achieves state-of-the-art performance on four FSL benchmark datasets, significantly outperforming existing Optimal Transport (OT)-based methods.

Keywords—few-shot learning, Unbalanced Sinkhorn Distance, optimal transmission theory, meta-learning

Cite: Yun Pang and Hayati A. Rahman, "Efficient Few-Shot Image Classification with Unbalanced Sinkhorn Distance for Robust Feature Alignment and Large-Scale Applications," Journal of Image and Graphics, Vol. 13, No. 5, pp. 570-578, 2025.

Copyright © 2025 by the authors. This is an open access article distributed under the Creative Commons Attribution License (CC-BY-4.0), which permits use, distribution and reproduction in any medium, provided that the article is properly cited, the use is non-commercial and no modifications or adaptations are made.