基于改进的Mask R-CNN图像篡改取证方法  

A forensic method of image tampering based on improved Mask R-CNN

作　　者：吴云 张玉金 江潇潇[1] 许灵龙 WU Yun;ZHANG Yujin;JIANG Xiaoxiao;XU Linglong(School of Electronic and Electrical Engineering,Shanghai University of Engineering Science,Shanghai 201620,China)

机构地区：[1]上海工程技术大学电子电气工程学院,上海201620

出　　处：《光电子．激光》2023年第4期413-421,共9页Journal of Optoelectronics·Laser

基　　金：上海市自然科学基金项目(17ZR1411900);上海市科委重点项目(18511101600)资助项目。

摘　　要：随着现代科学技术的进步,图像编辑工具的发展极大地降低了篡改所需成本。图像篡改手段有多种,现有的方法往往存在通用性差的问题。同时,这些方法只关注篡改定位而忽略对篡改手段的分类。本文提出一种基于改进的Mask R-CNN两阶段网络模型用于图像篡改取证。在特征提取部分,结合空域富模型(spatial rich model, SRM)和受约束卷积对输入图像进行预处理,再输入到ResNet101前4层中,以建立能够有效体现各种篡改痕迹的统一特征表示。一阶段网络通过注意力区域提议网络(attention region proposal network, A-RPN)检测篡改区域,预测模块实现篡改操作分类和粗略篡改区域定位。继而,一阶段网络得到的定位信息引导二阶段网络学习局部特征以定位出最终的篡改区域。本文所提出的模型能检测3种不同类型的图像篡改操作,包括复制-粘贴、拼接和移除。实验结果表明,本文所提出的方法在NIST16、COVERAGE、Columbia和CASIA数据集的F1值分别达到了0.924、0.761、0.791和0.473,优于传统方法和一些主流深度学习方法。With the progress of modern science and technology,the development of image editing tools has greatly reduced the cost of tampering.There are many methods for image tampering,and the existing methods often have the problem of poor universality.Meanwhile,these methods only focus on tampering location and ignore the classification of tampering means.This paper proposes a two-stage network model based on improved Mask R-CNN for image tampering forensics.In the feature extraction part,the input image is preprocessed with spatial rich model(SRM)and constrained convolution,and then input into the first four layers of ResNet101,so as to establish a unified feature representation that can effectively reflect various tampering traces.The first-stage network detects the tampering area through the attention region proposal network(A-RPN),and the prediction module realizes the classification of tampering operation and the location of rough tampering area.Then,the location information obtained by the first-stage network guides the second-stage network to learn local features to locate the final tampering area.The proposed model can detect three different types of image tampering operations,including copy-paste,splicing and removal.The experimental results show that the F1 values of the proposed method in NIST16,COVERAGE,Columbia and CASIA datasets reach 0.924,0.761,0.791 and 0.473 respectively,which is superior to traditional methods and some state-of-the-art deep learning methods.

关 键 词：数字图像取证 篡改检测 深度学习 两阶段网络 注意力机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]