相似性保持实例检索方法  

Similarity retention instance retrieval method

作　　者：赵宏伟[1,2] 王鹏[1] 范丽丽[1] 胡黄水 刘萍萍[1] ZHAO Hong-wei;WANG Peng;FAN Li-li;HU Huang-shui;LIU Ping-ping(College of Computer Science and Technology,Jilin University,Changchun 130012,China;State Key Laboratory of Applied Optics,Chinese Academy of Science,Changchun 130033,China;School of Computer Science and Engineering,Changchun University of Technology,Changchun 130012,China)

机构地区：[1]吉林大学计算机科学与技术学院,长春130012 [2]中国科学院应用光学国家重点实验室,长春130033 [3]长春工业大学计算机科学与工程学院,长春130012

出　　处：《吉林大学学报（工学版）》2019年第6期2045-2050,共6页Journal of Jilin University:Engineering and Technology Edition

基　　金：吉林省省级科技创新专项资金项目（20190302026GX）;吉林省发改委产业技术研究与开发项目（2019C054-4）;应用光学国家重点实验室开放基金项目（20173660）

摘　　要：针对多输入样本的网络训练,提出了相似性保持实例检索方法。首先,经深度网络除卷积结构提取输入图像的特征并加以池化操作。然后,依照基准排序,修正低相关性图像与查询图像的相似度关系,获得低相关性图像的对比损失系数,保留损失基准值内的损失值,进行基于相似性保持的对比损失训练。最后,使用训练后网络提取图像特征,用于进行实例级图像检索。实验结果表明:基于相似性保持的损失对比函数具有可行性,该方法显著提高了实例级图像的检索精确度。Aiming at the network training of multi-input samples,a similarity retention instance retrieval method is proposed.Firstly,the input image features are extracted by the convolution structure in the depth network and pooled.Then,according to the benchmark order,the similarity relationship between the low correlation image and the query image is corrected,and the low correlation image contrast loss coefficient is obtained,and the loss value within the loss reference value is retained.The loss value is performed to maintain contrast loss training based on similarity.Finally,the post-training network is used to extract image features for instance-level image retrieval.The experimental results show that the loss comparison function based on similarity is feasible,and the method significantly improves the accuracy of instancelevel image retrieval.

关 键 词：计算机应用 图像检索 卷积神经网络 损失函数 排序学习 相似性保持 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]