【摘要】 小样本图像分类因标注样本稀缺而具有挑战性，度量学习作为该领域的主流方法，其通常采用欧几里得距离来衡量查询样本与支持样本之间的差异以实现类别判别．然而，欧几里得距离对极端值高度敏感，易导致模型产生误判．提出一种基于布雷柯蒂斯（Bray–Curtis）距离的优化度量方法，以替代传统的欧几里得距离．布雷柯蒂斯距离具有较强的鲁棒性，不易受极端值干扰，但其在小样本度量学习中的原始计算数值范围较窄，使类别间差异难以得到充分表征，制约了模型的分类性能．针对该问题，引入尺度变换因子，对布雷柯蒂斯距离进行数值调整，以提升其在小样本图像分类中的适用性与判别力．基于这一方法，分别在ProtoNet与Meta DeepBDC框架上构建了ProtoNet＿Bray与Meta DeepBDC＿Bray两种变体网络．通过在MiniImageNet,TieredImageNet和CUB-200-2011三个数据集上的广泛实验，证明提出的方法能够有效地提高模型性能．更多还原

【Abstract】 Few-shot image classification remains a challenging task due to the scarcity of annotated samples. Metric learning has been widely adopted in this field,with the Euclidean distance commonly used to quantify the difference between query and support samples for category discrimination. However,the Euclidean distance is highly sensitive to outliers,which may result in misclassification. To address this limitation,this paper proposes a metric learning approach that leverages an optimized form of the Bray-Curtis distance as an alternative to the standard Euclidean distance. The Bray-Curtis distance is more robust against extreme values,but its original numerical range in few-shot learning is relatively narrow,limiting its ability to effectively capture inter-class differences. To overcome this drawback,we introduce a scaling factor to rescale the Bray-Curtis distance,thereby enhancing its applicability and discriminative capability in few-shot classification. Based on this idea,two network variants,ProtoNet＿Bray and Meta DeepBDC＿Bray,are constructed by integrating the optimized Bray-Curtis distance into ProtoNet and Meta DeepBDC,respectively. Extensive experiments on the MiniImageNet,TieredImageNet,and CUB-200-2011 datasets demonstrate that the proposed approach yields consistent performance gains over baseline methods.更多还原