摘要
针对卷积神经网络(CNN)全连接层得到的是图像类别的全局语义信息,无法有效抑制背景噪声以及表示图像局部的细节信息,导致细粒度图像检索任务中负样本靠前的问题,提出了一种选择性加权来聚合卷积特征并利用k相互最近邻(k-reciprocal nearest neighbor,k-RNN)重排的图像检索方法。该方法主要是通过提取并筛选CNN最后一层特征来聚合形成单维全局特征向量,再引入k相互最近邻算法对检索出的结果进行重排。在细粒度基准数据集CUB-200-2011、室内场景数据集Indoor和普通类别数据集Caltech-101进行验证评估。实验结果表明该方法能够有效改善检索出负样本靠前的问题,相比SCDA方法,该方法检索精度及召回率有显著提升。
Aiming at the problem that the full connection layer of convolutional neural network(CNN)obtains the global semantic information of image categories,which cannot effectively suppress the background noise and represent the local detail information of the image,leading to the negative samples in the fine-grained image retrieval task,this paper proposed a selective weighting method to aggregate the convolution features and rearrange the k-reciprocal nearest neighbor.This method mainly extracted and filtered the last layer features of CNN to aggregate into a single-dimensional global feature vector,and then introduced the k-nearest neighbor algorithm to rearrange the retrieved results.It carried out validation and evaluation in fine-grained benchmark dataset CUB-200-2011,indoor scene dataset Indoor,and common category dataset Caltech-101.The experimental results show that this method can effectively improve the retrieval of negative samples.Comparing with SCDA method,the retrieval accuracy and recall rate of this method are significantly improved.
作者
苟光磊
朱东旭
杨雨
Gou Guanglei;Zhu Dongxu;Yang Yu(School of Computer Science&Engineering,Chongqing University of Technology,Chongqing 400054,China)
出处
《计算机应用研究》
CSCD
北大核心
2022年第4期1259-1264,共6页
Application Research of Computers
基金
重庆市基础科学与前沿技术研究项目(cstc2017jcyjAX0144)
重庆理工大学研究生创新计划资助项目(clgycx20202089)。
关键词
图像检索
卷积特征聚合
细粒度
k相互近邻
image retrieval
convolutional features aggregation
fine-grained
k-reciprocal nearest neighbor
