北京航空航天大学学报 ›› 2020, Vol. 46 ›› Issue (5): 1039-1044.doi: 10.13700/j.bh.1001-5965.2019.0583

• 论文 • 上一篇    

基于深度学习的图像拼接篡改检测

边亮1, 罗霄阳2, 李硕2   

  1. 1. 北京航空航天大学 电子信息工程学院, 北京 100083;
    2. 中国人民公安大学 信息技术与网络安全学院, 北京 100038
  • 收稿日期:2019-11-13 发布日期:2020-05-19
  • 通讯作者: 边亮 E-mail:askquestionbl@163.com
  • 作者简介:边亮 男,硕士研究生。主要研究方向:图像信息融合与处理;罗霄阳 男,硕士研究生。主要研究方向:机器学习、图像取证;李硕 男,硕士研究生。主要研究方向:视频图像侦查。

Image mosaic tampering detection based on deep learning

BIAN Liang1, LUO Xiaoyang2, LI Shuo2   

  1. 1. School of Electronic and Information Engineering, Beihang University, Beijing 100083, China;
    2. School of Information Technology and Network Security, People's Public Security University of China, Beijing 100038, China
  • Received:2019-11-13 Published:2020-05-19

摘要: 传统图像拼接检测算法通过研究人员手动构造拼接特征,随着科技的进步以及图像处理技术的不断发展,手动构造特征的局限性逐渐体现出来,鲁棒性较弱,位置不易确定等。为了解决这些问题,构建了一种卷积神经网络(CNN),将卷积核前置并固定,自主学习相关特征从而检测拼接篡改的图像区域。经过一系列研究,发现拼接篡改图像的拼接篡改区域特征可以被CNN模型学习。在CNN模型之前,卷积核使用高通滤波器,激活函数采用指数线性单元(ELU),使得CNN模型具有识别拼接篡改图像边缘痕迹等特征的能力。检测结果表明:在IEEE IFS-TC图像拼接取证竞赛训练集上对拼接篡改图像拼接篡改区域定位的准确率为84.3%,对拼接篡改区域判定的真负类率为96.18%。

关键词: 卷积神经网络(CNN), 图像拼接取证, 深度学习, 指数线性单元(ELU), 损失函数

Abstract: The traditional image stitching detection algorithm manually constructs the stitching features by researchers. With the advancement of technology and the continuous development of image processing technology, the limitations of the features of manual construction, such as weak robustness and difficult positioning, are gradually manifested. Aimed at this kind of problem, this paper proposes to construct a Convolutional Neural Network (CNN) by means of fixed pre-convolution kernel, and detect the image tampering area by feature self-learning. Through experiments and research, it is found that the features of the mosaic tampering area of the spliced tamper image can be learned by the CNN model. Prior to the CNN model, the convolution kernel uses a high-pass filter and the activation function uses an Exponential Linear Unit (ELU), which makes the CNN model be capable of identifying features such as splicing and tampering with image edge traces. The detection results show that the positioning accuracy for the falsification image’s tampering area is 84.3% in the IEEE IFS-TC image forensics training set and the detection true negative rate of the tampering area is 96.18%.

Key words: Convolutional Neural Network (CNN), image mosaic forensics, deep learning, Exponential Linear Unit (ELU), loss function

中图分类号: 


版权所有 © 《北京航空航天大学学报》编辑部
通讯地址:北京市海淀区学院路37号 北京航空航天大学学报编辑部 邮编:100191 E-mail:jbuaa@buaa.edu.cn
本系统由北京玛格泰克科技发展有限公司设计开发