电子学报 ›› 2020, Vol. 48 ›› Issue (4): 790-799.DOI: 10.3969/j.issn.0372-2112.2020.04.021

• 学术论文 • 上一篇    下一篇

基于改进的有效区域基因选择与跨模态语义挖掘的图像属性标注

张红斌1, 蒋子良1, 熊其鹏1, 武晋鹏1, 邬任重1, 袁天2, 姬东鸿3   

  1. 1. 华东交通大学软件学院, 江西南昌 330013;
    2. 华东交通大学信息工程学院, 江西南昌 330013;
    3. 武汉大学国家网络安全学院, 湖北武汉 430072
  • 收稿日期:2018-07-02 修回日期:2019-12-20 出版日期:2020-04-25 发布日期:2020-04-25
  • 作者简介:张红斌 男,1979年10月生,江苏如皋人.华东交通大学软件学院副教授、硕士生导师.主要研究方向:计算机视觉、自然语言处理、推荐系统.E-mail:zhanghongbin@whu.edu.cn;蒋子良 男,1997年4月生,江西婺源人.华东交通大学软件学院硕士研究生.主要研究方向:机器学习、图像材质识别.E-mail:ziiliangjiang@163.com
  • 基金资助:
    国家自然科学基金(No.61762038,No.61861016);教育部人文社会科学研究规划基金项目(No.17YJAZH117);江西省自然科学基金(No.20171BAB202023);江西省科技厅重点研发计划(No.20171BBG70093,No.20192BBE50071);江西省教育厅科技项目(No.GJJ190323)

Image Attribute Annotation via a Modified Effective Range Based Gene Selection and Cross-Modal Semantics Mining

ZHANG Hong-bin1, JIANG Zi-liang1, XIONG Qi-peng1, WU Jin-peng1, WU Ren-zhong1, YUAN Tian2, JI Dong-hong3   

  1. 1. School of Software, East China Jiaotong University, Nanchang, Jiangxi 330013, China;
    2. School of Information Engineering, East China Jiaotong University, Nanchang, Jiangxi 330013, China;
    3. School of Cyber Science and Engineering, Wuhan University, Wuhan, Hubei 430072, China
  • Received:2018-07-02 Revised:2019-12-20 Online:2020-04-25 Published:2020-04-25

摘要: 图像属性标注是一种更细化的图像标注,它能缩小认知与特征间"语义鸿沟".现有研究多基于单特征且未挖掘属性蕴含的深层语义,故无法准确刻画图像内容.改进有效区域基因选择算法融合图像特征,并设计迁移学习策略,实现材质属性标注;基于判别相关分析挖掘特征间跨模态语义,以改进相对属性模型,标注材质属性蕴含的深层语义-实用属性.实验表明:材质属性标注精准度达63.11%,较最强基线提升1.97%;实用属性标注精准度达59.15%,较最强基线提升2.85%;层次化的标注结果能全面刻画图像内容.

关键词: 图像标注, 有效区域基因选择, 相对属性, 迁移学习, 跨模态语义, 判别相关分析

Abstract: Image attribute annotation is a refined method of image annotation.It can narrow the "semantic gap" between cognition and features.However,a single feature is used to characterize images and the deep-level semantics are not fully explored.So annotations cannot depict images comprehensively.The traditional effective range based gene selection algorithm is modified to complete feature fusion.And transfer learning strategy is designed to complete material annotation.The cross-modal semantics among features are mined by the discriminant correlation analysis algorithm.So the relative attribute model is optimized to complete deep-level semantics (practical attributes) annotation.Experimental results demonstrate:Material attributes annotation accuracy reaches 63.11%,which is improved by 1.97% compared with baseline.Practical attributes annotation accuracy reaches 59.15%,which is improved by 2.85% compared with baseline.The proposed hierarchical annotation mechanism can more comprehensively depict images.

Key words: image annotation, effective range based gene selection, relative attribute, transfer learning, cross-modal semantics, discriminant correlation analysis

中图分类号: