基于元度量的小样本学习方法

A Metametric Based Few Shot Learning Method

摘要: 针对小样本图像分类任务中样本量不足导致的提取特征不精确、网络性能下降等问题，基于元度量学习方法提出一个用于处理小样本图像分类的学习方法。该方法通过对同一类别的每个样本特征赋予权值，利用加权求和的方式得到该类的原型。权值由样本间的曼哈顿距离、欧氏距离和切比雪夫距离综合决定.最后采用三元损失函数使网络拉近同种类之间的距离，增加不同种类之间的距离。实验结果表明，该方法在5-way 1-shot和5-way 5-shot的设置下，在MiniImageNet数据集上实验准确率为68.33%和81.75%，在CUB数据集上实验准确率为73.76%和87.35%。对比结果显示，该方法能有效处理小样本问题。

Abstract: Aiming at the problems of inaccurate extraction features and reduced network performance caused by insufficient sample size in the few-shot image classification task, a learning method for pro-cessing few-shot image classification is proposed based on the metametric learning method. This method gives weights to each sample feature of the same category and obtains the prototype of the class by weighted summation. The weight is determined by the comprehensive distance between samples, the distance between the Manhattan, the Euclidean distance and the Chebyshev distance. Finally, a ternary loss function is used to bring the network closer to the distance between the same type and increase the distance between different types. Experimental results show that under the setting of 5-way 1-shot and 5-way 5-shot, the experimental accuracy of this method is 68.33% and 81.75% on the MiniImageNet data set, and the experimental accuracy on the CUB data set is 73.76% and 87.35%. The comparison results show that this method can effectively deal with few-shot learning.