通道注意力双线性度量网络  

Channel attention bilinear metric network

作　　者：李晓旭[1] 安文娟 武继杰 李真 张珂[2,3] 马占宇[4] LI Xiao-xu;AN Wen-juan;WU Ji-jie;LI Zhen;ZHANG Ke;MA Zhan-yu(School of Computer and Communication,Lanzhou University of Technology,Lanzhou 730050,China;Department of Electronics and Communication Engineering,North China Electric Power University,Baoding 071003,China;Hebei Key Laboratory of Power Internet of Things Technology,North China Electric Power University,Baoding 071003,China;Laboratory of Pattern Recognition and Intelligent System,School of Artificial Intelligence,Beijing University of Posts and Telecommunications,Beijing 100876,China)

机构地区：[1]兰州理工大学计算机与通信学院,兰州730050 [2]华北电力大学电子与通信工程系,保定071003 [3]华北电力大学河北省电力物联网技术重点实验室,保定071003 [4]北京邮电大学人工智能学院模式识别与智能系统实验室,北京100876

出　　处：《吉林大学学报（工学版）》2024年第2期524-532,共9页Journal of Jilin University:Engineering and Technology Edition

基　　金：国家自然科学基金项目(62176110,U19B2036,62225601,U23B2052,62076093);甘肃省重点研发计划项目(22YF7GA130);河北省省级科技计划项目(SZX2020034);北京邮电大学优秀青年团队项目(2023QNTD02);兰州理工大学红柳杰出青年基金项目。

摘　　要：针对小样本图像分类任务中模型对不同类的相似图片进行度量时,由于缺少对样本局部重要特征的关注且难以捕捉相似图片间的细微差别,导致出现部分查询样本与正确类原型的分类边界较模糊的问题,提出了一种通道注意力双线性度量网络(CABMN)。该网络首先增加模型对图片局部重要区域的关注度,然后利用双线性哈达玛积操作挖掘该重要区域的深层次二阶特征信息,使模型对图片局部关键区域的定位更精准。对比实验结果表明:本文提出的CABMN在各数据集上的分类性能均有提高,尤其在细粒度数据集CUB-200-2011和Stanford-Cars上达到86.19%和81.51%的分类准确率。In few-shot image classification tasks,when the model measures similar images of different classes,due to the lack of attention to local important features of the sample and the difficulty in capturing the subtle differences between similar images,the classification boundary between query samples and correct class prototype is fuzzy.Based on this,this paper proposes a Channel Attention Bilinear Metric Network(CABMN),which firstly increases the attention of the model to the local important region of the image,and then uses the bilinear Hadamard product operation to mine the deep second-order feature information of the important region,so that the model can locate the local key region of the image more accurately.Comparative experimental results show that the proposed CABMN has improved the classification performance on all datasets,especially on the fine-grained datasets CUB-200-2011 and Stanford-Cars,reaching 86.19%and 81.51%classification accuracy.

关 键 词：小样本学习 细粒度图像分类 度量学习 注意力机制 哈达玛积 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]