基于混合注意力的布朗距离协方差小样本图像分类算法  

Brownian Distance Covariance Few-shot Image Classification Algorithm Based on Hybrid Attention

作　　者：包春梅 王前 陈望 李志玲 王彬 王林 BAO Chunmei;WANG Qian;CHEN Wang;LI Zhiling;WANG Bin;WANG Lin(School of Data Science and Information Engineering,Guizhou Minzu University,Guiyang 550025,China;Key Laboratory of Pattern Recognition and Intelligent System,Guiyang 550025,China)

机构地区：[1]贵州民族大学数据科学与信息工程学院,贵阳550025 [2]贵州省模式识别与智能系统重点实验室,贵阳550025

出　　处：《湖北民族大学学报（自然科学版）》2024年第4期521-527,共7页Journal of Hubei Minzu University:Natural Science Edition

基　　金：贵州省科技计划项目(黔科合基础-ZK[2022]一般195,黔科合平台人才-ZCKJ[2021]007);贵州省青年科技人才成长项目(黔教合KY字[2021]104);贵州省模式识别与智能系统重点实验室开放课题(GZMUKL[2022]KF01,GZMUKL[2022]KF05);贵州民族大学基金科研项目(GZMUZK[2023]YB14),黔教技([2024]063号)。

摘　　要：针对小样本图像分类中主干卷积神经网络所提取特征缺乏关联性,以及在将其表示为布朗距离协方差(Brownian distance covariance,BDC)矩阵时易出现通道特征信息丢失的问题,提出了基于混合注意力的布朗距离协方差(hybrid attention-Brownian distance covariance,HA-BDC)小样本图像分类算法。该算法首先在主干卷积神经网络中引入加权的非局部注意力机制来增强特征提取能力,然后采用跨空间学习的高效多尺度注意力模块重塑嵌入特征,使得新的特征张量能够保留每个通道上的信息。再用BDC度量模块将新的特征张量表示为BDC矩阵,并通过全连接层得到BDC矩阵的权重矩阵,最后采用逻辑回归模型进行分类。在大规模视觉识别挑战的子数据集(mini ImageNet large scale visual recognition challenge,miniImageNet)和分层大规模视觉挑战的子数据集(tiered ImageNet large scale visual recognition challenge,tieredImageNet)上进行分类实验,结果表明,HA-BDC算法相较于简单迁移学习深度布朗协方差(simple transfer learning deep Brownian distance covariance,STL DeepBDC)算法在5-类别1-样本任务中的分类准确率分别提升了2.83%和0.77%。该研究能有效应用于数据量较小的濒危动物识别、罕见疾病识别等领域。In order to solve the problems that the extracted feature of backbone convolutional neural network lacked correlation in few-shot image classification and the channel feature information was easily lost when it was expressed as Brownian distance covariance(BDC),and a image classification algorithm based on hybrid attention-Brownian distance covariance(HA-BDC)was proposed.Firstly,the backbone convolutional neural network was introduced into a weighted nonlocal attention mechanism,which enhanced the feature extraction ability.Secondly,an efficient multiscale attention module with cross-space learning was used to reshape the embedded features so that the new feature tensor could retain the information on each channel.The BDC module was used to represent the new feature tensor as a BDC matrix,the weight matrix of the BDC matrix was obtained through the fully connected layer and finally a logistic regression model was employed for classification.Classification experiments were performed on the mini ImageNet large scale visual recognition challenge(miniImageNet)and tiered ImageNet large scale visual recognition challenge(tieredImageNet)datasets.The results showed that the classification accuracy of the proposed algorithm was improved by 2.83%and 0.77%respectively in the 5-way 1-shot task compared with the simple transfer learning deep Brownian distance convariance(STL DeepBDC)model.The research can be used in the identification of endangered animals and rare diseases with a small amount of data.

关 键 词：机器学习 小样本学习 度量学习 布朗距离协方差 混合注意力机制 特征重塑 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]