北京邮电大学学报

  • EI核心期刊

北京邮电大学学报 ›› 2019, Vol. 42 ›› Issue (1): 68-73.doi: 10.13190/j.jbupt.2018-244

• 论文 • 上一篇    下一篇

基于改进DeepLabv3+的拼接篡改定位检测技术

张继威, 牛少彰, 曹志义, 王心怡   

  1. 北京邮电大学 智能通信软件与多媒体北京市重点实验室, 北京 100876
  • 收稿日期:2018-09-24 出版日期:2019-02-28 发布日期:2019-03-08
  • 通讯作者: 牛少彰(1963-),男,教授,博士生导师,E-mail:szniu@bupt.edu.cn. E-mail:szniu@bupt.edu.cn
  • 作者简介:张继威(1989-),男,博士生.
  • 基金资助:
    国家自然科学基金项目(U1536121,61370195)

Locating Image Splicing by Improved DeepLabv3+

ZHANG Ji-wei, NIU Shao-zhang, CAO Zhi-yi, WANG Xin-yi   

  1. Beijing Key Laboratory of Intelligent Telecommunication Software and Multimedia, Beijing University of Posts and Telecommunications, Beijing 100876, China
  • Received:2018-09-24 Online:2019-02-28 Published:2019-03-08

摘要: 目前大部分的拼接检测算法主要检测图片是否经历了拼接篡改,而不是对拼接区域进行定位检测,对此,提出了一种基于改进DeepLabv3+的拼接区域定位检测技术.首先,改变原DeepLabv3+网络的分类数;其次,通过改造图像训练库,在训练库中加入含有人物的原图,对原图和篡改图的标签进行区别设置,引导改进的DeepLabv3+网络去学习原图人物和篡改人物特征的区别.实验结果显示,基于改进DeepLabv3+的拼接区域定位检测技术在CASIA数据库上取得了更好的拼接区域定位效果.

关键词: 改进DeepLabv3+, 图像拼接, 区域定位, 轮廓学习, 边缘特征

Abstract: As the existing splicing detection algorithms mainly detect whether the image has undergone image splicing, rather than regional localization of the spliced area. An improved DeepLabv3+ network that changes the categories of original DeepLabv3+ network is proposed to locate image regional forgery. In addition, the original images are added to the training dataset, and the ground truth of the original images are set to be black images, which will guide the network to learn the difference between the original images and the spliced images. Experimental results show that the improved DeepLabv3+ network has achieved a finer effect for locating image regional forgery than the existing algorithms on the CASIA database.

Key words: improved DeepLabv3+, image splicing, regional localization, contour learning, edge features

中图分类号: