基于ResNet的唐卡检索  

Thangka Retrieval Based on ResNet

作　　者：闫龙泉 骆沛然 史伟 刘秀全 Yan Longquan;Luo Peiran;Shi Wei;Liu Xiuquan(School of Information Engineering,Ningxia University,Yinchuan 750021,China)

机构地区：[1]宁夏大学信息工程学院,宁夏银川750021

出　　处：《宁夏大学学报（自然科学版）》2021年第3期257-262,269,共7页Journal of Ningxia University(Natural Science Edition)

基　　金：国家自然科学基金资助项目(62166030,12061055,61662060);宁夏大学-中国西部一流大学科技创新项目(ZKZD2017005).

摘　　要：唐卡是独具特色的藏族艺术资源,因其内容复杂、种类多样、分辨率高,传统的图像检索方法无法实现对唐卡图像进行快速而精准的检索.针对上述问题,基于残差神经网络(ResNet)进行唐卡图像的检索,首先使用ResNet50网络结构进行唐卡图像的深层语义信息提取,然后采用余弦相似度比对算法对提取的唐卡特征进行准确分析.实验表明,相较于SIFT方法,ResNet50检索方法的查准率提升了40%,召回率提升了25%;相较于VGG16,ResNet50检索方法的查准率提升了3%,召回率提升了5%.Thangka is a kind of unique Tibetan art resources.Traditional methods cannot achieve rapid and accurate retrieval of Thangka images because of its complex content,diverse types,and high resolution.Accordingly,this paper suggests a method to retrieve Thangka images based on the residual neural network(ResNet).The ResNet50 network structure is used to extract the deep semantic information of the Thangka images and then the extracted Thangka features are accurately analyzed by using the cosine-similarity-comparison algorithm.Experiments show that compared with the SIFT method,the accuracy of ResNet50 retrieval method increased by 40%,and the recall rate increased by 25%;and compared with VGG16,the accuracy of ResNet50 retrieval increased by 3%,and the recall rate increased by 5%.

关 键 词：唐卡图像 CNN 图像检索 ResNet 特征提取 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]