一种基于原型校正的小样本学习算法  

Prototype Calibration Based Few-shot Learning Algorithm

作　　者：梁辰晨 江宸逸 王室栋 张浩峰[1] LIANG Chenchen;JIANG Chenyi;WANG Shidong;ZHANG Haofeng(School of Computer Science and Engineering,Nanjing University of Science and Technology,Nanjing 210094,China;School of Computing,Newcastle University,Newcastle NE17RU,UK)

机构地区：[1]南京理工大学计算机科学与工程学院,南京210094 [2]纽卡斯尔大学计算学院,纽卡斯尔NE17RU

出　　处：《小型微型计算机系统》2025年第3期520-527,共8页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(62371235,62072246)资助.

摘　　要：现有的少样本学习(Few-Shot Learning)方法通常使用基于任务(episode)的训练策略训练模型,但在随机采样构建任务的过程中不可避免地存在采样偏差问题,质量较差的支持样本一定程度上导致模型难以收敛.本文提出了交互视图原型校正网络通过两种策略缓解这个问题,首先,模型利用全局信息来纠正每个任务中的支持样本特征,从而缓解随机采样导致的原型偏移问题;其次,文中提出了实例原型生成算法,该算法利用支持样本特征的局部视图原型集根据不同的查询样本实例生成对应的实例级原型,从而优化了使用支持样本全局特征作为类原型区分度不足的问题.模型在多个基准数据集上的进行了充分实验,实验结果验证了方法的有效性及鲁棒性.Existing Few-Shot Learning(FSL)methods commonly use episode-based training strategy to train models,but this inevitably has a sample bias problem that makes it difficult for the model to maintain good performance with poor-quality support samples.This paper proposes a Mutual-view prototype calibration network to alleviate this problem in two ways.First,the model utilizes global information to correct support features in each episode,thereby alleviating the prototype offset problem caused by random sampling;Second,this paper proposes an instance prototype generation algorithm,which uses the local-view prototype set of support features and generates query instance-level prototypes based on different query images,thus eliminating the disadvantage of using the global feature of support images as class prototypes.Extensive experiments on multiple benchmark datasets verify the state-of-the-art effectiveness of our method and confirm its robustness.

关 键 词：小样本学习 原型校正 图像分类 原型集生成 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]