基于多尺度特征先验的图像拼接篡改检测  

Image splicing tamper detection based on multi-scale feature priori

作　　者：吕建凯 卢望龙 王敏[1] 刘影 史开杰 黄辉[1] 赵汉理[1] LÜ Jiankai;LU Wanglong;WANG Min;LIU Ying;SHI Kaijie;HUANG Hui;ZHAO Hanli(College of Computer Science and Artificial Intelligence,Wenzhou University,Wenzhou,Zhejiang 325035,China;School of Information Engineering,Wenzhou Business College,Wenzhou,Zhejiang 325035,China)

机构地区：[1]温州大学计算机与人工智能学院,浙江温州325035 [2]温州商学院信息工程学院,浙江温州325035

出　　处：《中国科技论文》2022年第11期1267-1275,共9页China Sciencepaper

基　　金：国家自然科学基金资助项目(62072340);浙江省自然科学基金资助项目(LZ21F020001,LSZ19F020001);浙江大学CAD&CG国家重点实验室开放课题资助项目(A2220)。

摘　　要：针对现有图像拼接篡改算法存在识别准确度不高的问题,提出了一种基于多尺度特征先验的图像拼接篡改检测框架。首先,训练一个环状残差网络对数据集的深度特征进行学习,将训练得到的模型作为后续网络学习的先验知识。然后,开发一个由多个外部注意力块和环状残差块组合的分支网络,该分支网络有效地利用环状残差网络训练模型中的多尺度特征先验知识,进一步学习到精确的拼接篡改图像区域信息。为了进一步解决数据不均衡问题,结合二值交叉熵损失和Dice损失来设计模型训练目标函数,最终较为精确地检测出篡改区域图。大量实验对比数据表明,所提方法可以有效检测出不同场景下的拼接篡改图像,与对比方法相比有更高的检测准确率。为了进一步检验模型的鲁棒性,对压缩和噪声图像进行了实验测试,结果表明所提方法在鲁棒性方面也能取得较好的结果。Aiming at the defect as low recognition accuracy of current splicing tamper detection algorithm,an image splicing tamper detection framework based on the multi-scale feature priori was proposed.Firstly,an annular residual network was trained to learn deep features of the training dataset and the trained model was used as the priori knowledge of the subsequent network learning.Then,a branch network composed of multiple external attention blocks and annular residual blocks was developed.The branch network effectively utilized the priori knowledge of multi-scale features in the annular residual network training model to further learn the accurate information of splicing tampered region.In order to solve the issue of data imbalance,the objective function of model training was designed by combining binary cross entropy loss and Dice loss,and the tampered area map was finally detected more accurately.A large number of experimental comparative data show that the proposed method can effectively detect splicing tampered images in various scenes and has a higher detection accuracy than the comparison methods.To further verify the robustness of the model,experimental tests were carried out on compressed and noisy images.Experimental results show that the proposed method is able to achieve satisfying results in robustness.

关 键 词：图像篡改检测 图像分割 伪造检测 残差网络 深度学习 

分 类 号：TP317.4[自动化与计算机技术—计算机软件与理论]