基于协方差度量矩阵的多尺度融合的小样本学习  

Few-Shot Learning Based on Multi-Scale Fusion of Covariance Metric Matrix

作　　者：莫春晗 陆建峰[1] MO Chunhan;LU Jianfeng(School of Computer Science and Engineering,Nanjing Universitiy of Science and Tecnology,Nanjing 210094)

机构地区：[1]南京理工大学计算机科学与工程学院,南京210094

出　　处：《计算机与数字工程》2024年第5期1373-1376,1405,共5页Computer & Digital Engineering

摘　　要：对于少量训练样本下的图像分类问题,现有的深度模型算法存在着很多问题。比如传统的基于度量学习的算法只是依据定义的距离关系来推断样本类别,易导致模型在类别过多时难以捕获类间差异。同时它们只聚焦于一阶统计量的关系计算和高层语义信息的利用,忽视了低级丰富的原始视觉特征的表达。针对这些问题,论文提出了基于协方差度量矩阵的多尺度融合算法。该算法利用不同尺度信息下的二阶统计量来更新模型参数。实验结果表明该算法能有效地提高小样本图像分类的准确率,具有一定的实用价值。For the image classification problem with limited training samples,the existing deep neural network-based models have exposed many problems.For example,traditional algorithms based on metric learning usually rely on the pre-defined distance measurement,which makes it difficult for the model to capture the differences among classes when there are too many classes to be classified.Moreover,they only focus on the relational calculation of first-order statistics and the utilization of high-level semantic information,and ignore the low-level but abundant visual features.To address above issues,in this paper,a multi-scale fusion al-gorithm is proposed based on the covariance metric matrix.The second-order statistics under different scale is used to update model parameters.Experimental results show that the proposed model can effectively improve the accuracy of few-shot image classifica-tion,which indicates its promising value in real-world applications.

关 键 词：协方差度量矩阵 多尺度融合 小样本学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]