摘要
针对基于特征分布校准的小样本分类算法无法准确揭示新类特征分布的问题,提出一种融合隐空间变换和密度聚类的改进算法,以解决N way-K shot任务模式下的小样本图像分类问题.首先,通过广度残差神经网络提取基类和新类图像的深度特征;其次,采用隐空间变换方法约束新类特征分布,使其更接近正态分布;再次,利用密度聚类方法为新类选取合适基类,将基类统计信息迁移到新类,并通过多元正态分布矩阵实现样本扩充;最后,构建基于集成学习的分类器,完成小样本图像分类任务.实验结果表明,相比于传统特征分布校准方法,该算法的分类准确率更高.
The few-shot classification algorithm based on feature distribution calibration can not accurately reveal the feature distributions of novel classes.To address the issue and solve the N way-K shot few-shot image classification task,an improved algorithm combining latent space transform and density-based spatial clustering is proposed in this paper.Firstly,the deep features of base and novel images are extracted by breadth residual neural network.Then,the latent space transformation method is introduced to con-strain the feature distributions of novel classes so that it is closer to the normal distribution.Mean-while,the density-based spatial clustering method is employed to select a suitable base class for each novel class.Thus the statistical information of the base class is transferred to the novel class,and the sample can be effectively expanded by multivariate normal distribution matrix.Finally,a classifier based on ensemble learning is constructed to accomplish the classification task.The experimental results on two baseline datasets show that the proposed method can further improve the classification accuracy compared with the traditional feature distribution calibration model.
作者
张涛
王波
赵宇
袁运浩
ZHANG Tao;WANG Bo;ZHAO Yu;YUAN Yunhao(School of Information Engineering,Yangzhou University,Yangzhou 225127,China)
出处
《扬州大学学报（自然科学版）》
CAS
2024年第1期56-61,共6页
Journal of Yangzhou University：Natural Science Edition
基金
江苏省自然科学基金资助项目(BK20200921)
江苏省高等学校自然科学基金资助项目(20KJB520007,20KJB510024)。
关键词
小样本学习
图像分类
特征分布校准
隐空间变换
密度聚类
few-shot learning
image classification
feature distribution calibration
latent space transform
density-based spatial clustering
