基于偏序结构的图像标注  

Image Annotation Based on Partial Order Structure

作　　者：孙雅倩 顾广华[1,2] 赵志明 卢辉斌 Sun Yaqian;Gu Guanghua;Zhao Zhiming;Lu Huibin(School of Information Science and Engineering,Yanshan University,Qinhuangdao,Hebei 066000,China;Hebei Key Laboratory of Information Transmission and Signal Processing(Yanshan University),Qinhuangdao,Hebei 066000,China)

机构地区：[1]燕山大学信息科学与工程学院,河北秦皇岛066000 [2]河北省信息传输与信号处理重点实验室,河北秦皇岛066000

出　　处：《信号处理》2020年第9期1574-1581,共8页Journal of Signal Processing

基　　金：国家自然科学基金(61303128);河北省自然科学基金(F2017203169,F2018203239);河北省高等学校科学研究项目重点项目(ZD2017080)。

摘　　要：图像标注旨在为图像分配一系列的语义标签描述图像的内容。针对高级语义与低级特征之间的语义鸿沟问题,本文提出了基于偏序结构的图像标注方法。首先,利用卷积神经网络VGG-19模型提取图像特征。然后,利用提取的图像特征计算训练图像与测试图像的相似性得分,得到测试图像的初始邻近集及邻近标签。最后,通过构建的属性偏序结构,获得邻近标签的相关语义,提高标签的丰富度;利用构建的对象偏序结构,得到最终的标注候选集。为了提高标注的准确率,设置一个频率阈值筛选出频率较高的标签作为最终的关键词。通过实验证明,实验结果有效地提高了标注的准确率和召回率。Image annotation aims at assigning a set of semantic labels to describe the content of the image.Aiming at the gap of high-level semantics and low-level features in image annotation,this paper proposed an image annotation methodology based on partial order structure.At first,VGG-19 structure in convolutional neural network was used to extract image features.Then,calculated the visual score with extracted image features and obtained the initial neighbor set and adjacent labels.Finally,through the attribute partial order structure diagram,the method can get the related semantics of the adjacent labels,and the related semantics were used to construct the object partial order structure(OPOS)diagram in the purpose of obtaining the final semantic neighbor set.Technically,set a threshold of the word frequency to select the labels with the higher frequency as the final keywords.More remarkably,the experimental results show that the method effectively improves precision and recall rate of the annotation.

关 键 词：图像标注 属性偏序结构 对象偏序结构 邻近集 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]