基于深度特征加权的图像表示方法  

Deep Feature Weighting Based Image Representation

作　　者：朱杰[1] 赵相坤[2] 谢博鋆[3,4] 吴树芳[5] ZHU Jie;ZHAO Xiangkun;XIE Bojun;WU Shufang(Department of Information Management,the National Police University for Criminal Justice,Baoding 071000,China;School of Biomedical Engineering,Capital Medical University,Beijing 100069 China;Hebei Key Laboratory of Machine Learning and Computational Intelligence,Baoding 071002,China;College of Mathematics and Information Science,Hebei University,Baoding 071002,China;College of Management,Hebei University,Baoding 071002,China)

机构地区：[1]中央司法警官学院信息管理系,河北保定071000 [2]首都医科大学生物医学工程学院,北京100069 [3]河北省机器学习与计算智能重点实验室,河北保定071002 [4]河北大学数学与信息科学学院,河北保定071002 [5]河北大学管理学院,河北保定071002

出　　处：《郑州大学学报（理学版）》2020年第1期47-53,共7页Journal of Zhengzhou University:Natural Science Edition

基　　金：河北省自然科学基金项目(F2018511002);河北大学高层次创新人才科研启动经费项目;河北省高等学校科学技术研究项目(QN2018251,QN2018084,Z2019037);中央司法警官学院项目(XYZ201602);河北省高等学校科学技术研究项目(Z2019037);首都医科大学基础-临床科研合作基金(17JL86)

摘　　要：卷积神经网络可以在图像检索中为图像内容提供有效的表示,基于该理论提出一种基于深度特征加权的图像表示方法,此方法通过对深度特征加权,突出图像表示中对象的内容,并降低背景信息的影响。首先,通过预训练卷积神经网络提取出图像的特征映射,然后根据不同特征映射的特点,计算出深度特征的位置重要性、区域重要性和通道重要性,并根据3种重要性对深度特征进行加权,最后通过池化与深度特征聚合的方式生成图像表示。实验结果表明,与其他图像表示方法相比,提出的方法在Holiday、Oxford和Paris图像库中取得了更好的检索效果。Convolutional Neural Networks could provide effective representations for image contents in image retrieval.A deep feature weighting based image representation method was proposed,which could highlight the content of objects and reduce the impact of background information by weighting the deep features.Firstly,the feature maps were extracted by a pre-trained Convolutional Neural Network.Then,according to the characteristics of these features maps,the position weight,region weight and channel weight were computed and used to weight the deep features.Finally,the image representation was generated by pooling and aggregation of the weighted deep features.Experiments demonstrated that compared with some state-of-the-art image representation methods,the proposed method could achieve better retrieval results in the Holiday,Oxford and Paris image databases.

关 键 词：卷积神经网络 图像检索 特征加权 池化 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]