结合原型校准分布的小样本学习方法  

Prototype-based calibration distribution for few-shot learning

作　　者：黄友文[1] 胡燕芳 魏国庆 Huang Youwen;Hu Yanfang;Wei Guoqing(School of Information Engineering,Jiangxi University of Science and Technology,Ganzhou 341000,China)

机构地区：[1]江西理工大学信息工程学院,赣州341000

出　　处：《电子测量技术》2022年第5期132-139,共8页Electronic Measurement Technology

基　　金：江西省教育厅科技项目(GJJ180443)资助。

摘　　要：针对小样本学习中样本数量过少难以表述类别特征的问题,提出一种结合原型校准数据分布的小样本学习方法。首先,利用嵌入网络对图像进行预处理,将提取到的新类特征进行幂次变换。随后,通过相似度加权基类来表征新类样本的原型,充分利用已学过的基类知识来减小计算原型与实际原型的偏差。最后,以新类样本特征与其对应的原型构造均匀分布,根据样本数量改变均匀分布边界,使采样集中在置信度高的区域,生成大量的新类特征扩充训练分类器的支持集。在5-way 1-shot和5-way 5-shot设置中,该方法在miniImageNet数据集上的准确率分别为68.94%和84.75%,在CUB数据集上分别为81.75%和91.88%,均优于现有方法的最好结果。因此,所提方法能有效提高模型在小样本图像上的分类性能,获得更高的预测准确率。Aiming at the problem that the number of samples in few-shot learning is too small to represent the characteristics of data categories, a few-shot learning method combining prototype calibration data distribution is proposed. First, the embedded network is used to preprocess the image to obtain new class features, and the extracted new class features are processed by power transformation. Then, the prototype of the new sample is characterized by the weighting of the similarity of the base class, and the learned knowledge of the base class is fully utilized to reduce the deviation between the calculated prototype and the actual prototype. Finally, we sample from the uniform distribution between the features of the new class instance and its prototype representation. It generates a large amount of feature data to expand the support set of the new class. At the same time, we propose a method to change the boundary of the uniform distribution according to the number of samples. As a result, the samples are concentrated in areas with high confidence. The accuracy of 5-way 1-shot and 5-way 5-shot of our method are 68.94% and 84.75% on the miniImageNet dataset, respectively, and the accuracy on the CUB dataset are 81.75% and 91.88%, respectively, which are better than the best results of existing methods. The experimental results show that our method can effectively improve the model’s prediction accuracy in few-shot classification.

关 键 词：小样本学习 图像分类 原型表征 均匀分布 数据生成 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]