DeepSD:Sinkhorn Distance Calculated by Different Methods for Few-Shot Learning

Few-Shot Learning (FSL) has emerged as an effective approach for tasks with limited or no labeled samples, showing great potential. However, the challenge remains in efficiently extracting and learning valuable information in a timely manner. In this study, we revisit Few-Shot Learning classification methods from an optimization perspective. We leverage the Sinkhorn Distance (SD) as a measure of similarity between computationally intensive sample features. Moreover, we explore the Sinkhorn Stabilized Distance (SSD) and Unbalanced Sinkhorn Distance (UNSD) to optimize the matching flow between two samples, aiming for the best alignment. To address issues stemming from intra-class variations and the challenge of differentiating foreground and background features, we introduce the Polarized Selective Kernel Attention (PSKA) mechanism. Extensive experiments validate the efficacy of our proposed methods, demonstrating competitive performance across multiple benchmark datasets.