基于非局部自相关的复制粘贴检测算法

doi:10.16180/j.cnki.issn1007-7820.2022.10.010

摘要/Abstract

摘要：

针对数字图像复制-粘贴篡改无法区分目标与来源的问题,文中通过改进相似度匹配算法和利用非局部自注意力机制,在定位出篡改区域的前提下,解决了篡改源区域和目标区域的分类问题。总体框架为双分支检测网络,主分支采用经典U-net分割篡改区域像素,副分支通过孪生网络进行特征提取并计算自相关性,从而分割出篡改目标与源区域像素。将双分支融合后进行端到端训练,最终网络预测出三分类结果。实验结果表明,文中算法检测定位目标区域时的像素级分类精确率达到了80.47%,且F1值及准确度均优于对比算法。可视化结果和鲁棒性实验也表明文中算法具有良好的泛化性能。

关键词: 孪生网络, 图像被动取证, 复制粘贴篡改, 相似度距离, 非局部算子, 自相关性, 编码器-解码器, 端到端

Abstract:

On account of the problem that forgery target and source of digital image copy-move manipulation cannot be distinguished, this study improves the similarity matching algorithm and uses non-local self-attention mechanism to solve the classification problem of copy-move forgery source and target areas, under the premise that manipulated regions are detected. The overall framework is a dual-branch detection network. The main branch uses the classic U-net to segment the pixel of forgery regions, and the auxiliary branch uses the siamese network to extract features and calculate the autocorrelation to separate the forgery targets and source area pixels. Finally, three-categories results can be predicted by end-to-end training after fusing two branches. The experiment result shows that the pixel-level classification accuracy of the proposed algorithm when detecting the localized target area reaches 80.47%, and the F1 value and accuracy are better than the compared algorithm. The visualization results and robustness experiments also show that the proposed algorithm has excellent generalization performance.

Key words: siamese network, image passive forensic, copy-move forgery, similarity distance, non-local operator, self-correlation, encoder-decoder, end-to-end

中图分类号:

TP391.41

吴旭,刘翔. 基于非局部自相关的复制粘贴检测算法[J]. 电子科技, 2022, 35(10): 59-64.

WU Xu,LIU Xiang. Copy-Move Forgery Detection Algorithm Based on Non-Local Self-Correlation[J]. Electronic Science and Technology, 2022, 35(10): 59-64.

图/表 9

图1

图2

图3

图4

图5

表1

表2

图6

表3

参考文献 21

[1]	周琳娜. 数字图像盲取证技术研究[M]. 北京: 北京邮电大学出版社, 2007.
	Zhou Linna. Research on blind forensics technology of digital image[M]. Beijing: Beijing University of Posts and Telecommunications Press, 2007.
[2]	Fridrich A J, Soukal B D, Luká A J. Detection of copy-move forgery in digital images[C]. Cleveland: Proceedings of the Digital Forensic Research Workshop, 2003.
[3]	Zhao J, Guo J. Passive forensics for copy-move image forgery using a method based on DCT and SVD[J]. Forensic Science International, 2013, 233(1-3):158-166. doi: 10.1016/j.forsciint.2013.09.013 pmid: 24314516
[4]	Wang X, Liu Y, Xu H, et al. Robust copy-move forgery detection using quaternion exponent moments[J]. Pattern Analysis and Applications, 2018, 21(2):451-467. doi: 10.1007/s10044-016-0588-1
[5]	Lowe D G. Distinctive image features from scale-invariant keypoints[J]. International Journal of Computer Vision, 2004, 60(2):91-110. doi: 10.1023/B:VISI.0000029664.99615.94
[6]	Li K, Li H, Yang B, et al. Detection of image forgery based on improved PCA-SIFT[J]. Lecture Notes in Electrical Engineering, 2014, 277(1):679-686.
[7]	Bay H, Ess A, Tuytelaars T, et al. Speeded-up robust features[J]. Computer Vision and Image Understanding, 2008, 110(3):346-359. doi: 10.1016/j.cviu.2007.09.014
[8]	Xu B, Wang J W, Liu G J, et al. Image copy-move forgery detection based on SURF[C]. Nanjing: Proceedings of the International Conference on Multimedia Information Networking and Security, 2010.
[9]	Rao Y, Ni J. A deep learning approach to detection of splicing and copy-move forgeries in images[C]. Abu Dhabi: Proceedings of the IEEE International Workshop on Information Forensics and Security, 2016.
[10]	Ouyang J, Liu Y, Liao M. Copy-move forgery detection based on deep learning[C]. Shanghai: Proceedings of the Tenth International Congress on Image and Signal Processing, BioMedical Engineering and Informatics,IEEE, 2017.
[11]	Liu Y, Guan Q, Zhao X. Copy-move forgery detection based on convolutional kernel network[J]. Multimedia Tools and Applications, 2018, 77(14):18269-18293. doi: 10.1007/s11042-017-5374-6
[12]	Wu Y, Abd-Almageed W, Natarajan P. Image copy-move forgery detection via an end-to-end deep neural network[C]. Lake Tahoe: Proceedings of the IEEE Winter Conference on Applications of Computer Vision,IEEE, 2018.
[13]	Zhong J L, Pun C M. An end-to-end dense-inceptionNet for image copy-move forgery detection[J]. IEEE Transactions on Information Forensics and Security, 2019(15):2134-2146.
[14]	Wu Y, Abd-Almageed W, Natarajan P. BusterNet: Detecting copy-move image forgery with source/target localization[C]. Munich: Proceedings of the European Conference on Computer Vision, 2018.
[15]	李应灿, 杨建权, 丁峰, 等. 区分来源和目标区域的图像 Copy-move伪造检测方法[J]. 信号处理, 2020, 36(9):1533-1543.
	Li Yingcan, Yang Jianquan, Ding Feng, et al. Copy-move detection method for distinguishing between source and target regions[J]. Journal of Signal Processing, 2020, 36(9):1533-1543.
[16]	Dong J, Wang W, Tan T. Casia image tampering detection evaluation database[C]. Beijing: Proceedings of the IEEE China Summit and International Conference on Signal and Information Processing,IEEE, 2013.
[17]	Tralic D, Zupancic I, Grgic S, et al. CoMoFoD-new database for copy-move forgery detection[C]. Zadar: Proceedings of the ELMAR,IEEE, 2013.
[18]	Zagoruyko S, Komodakis N. Learning to compare image patches via convolutional neural networks[C]. Boston: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2015.
[19]	许凤翔. 一种改进相似度的协同过滤算法实现[J]. 电子科技, 2020, 33(2):54-59.
	Xu Fengxiang. Implementation of a collaborative filtering algorithm based on improved similarity[J]. Electronic Science and Technology, 2020, 33(2):54-59.
[20]	Wang X, Girshick R, Gupta A, et al. Non-local neural networks[C]. Salt Lake City: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018.
[21]	郑萌. 基于改进注意力机制模型的智能英语翻译方法研究[J]. 电子科技, 2020, 33(11):84-87.
	Zheng Meng. Research on intelligent English translation method based on improved attention mechanism model[J]. Electronic Science and Technology, 2020, 33(11):84-87.

评估指标	区域	文献[14]	文献[15]	本文算法
Precision	原始	90.23%	90.91%	91.78%
	来源	25.46%	30.37%	35.58%
	目标	13.27%	33.93%	34.81%
Recall	原始	97.48%	97.22%	97.38%
	来源	14.41%	22.59%	25.83%
	目标	5.89%	23.77%	24.94%
F1-Score	原始	93.71%	93.69%	94.50%
	来源	18.40%	25.91%	29.93%
	目标	8.16%	27.96%	29.06%

评估指标	区域	BusterNet	cGANs	本文算法
Precision	原始	93.96%	95.53%	97.66%
	来源	66.93%	70.42%	78.92%
	目标	48.67%	72.26%	80.47%
Recall	原始	97.14%	95.48%	96.40%
	来源	50.40%	58.84%	63.18%
	目标	21.74%	49.76%	60.79%
F1-Score	原始	95.52%	95.50%	97.03%
	来源	57.50%	64.11%	70.18%
	目标	30.05%	58.94%	69.26%

攻击手段		BusterNet	cGANs	本文算法
JPEG压缩	60%	0.47	0.70	0.83
	70%	0.41	0.66	0.79
	80%	0.35	0.61	0.70
亮度	线性变换	0.49	0.73	0.81
	对数变换	0.46	0.71	0.80
	指数变换	0.45	0.71	0.79
模糊	高斯	0.43	0.72	0.82
	均值	0.40	0.74	0.82
	中值	0.42	0.75	0.80
尺度变换	缩小0.5倍	0.46	0.73	0.80
	放大2倍	0.47	0.75	0.81
	放大1倍	0.47	0.75	0.83
噪声	高斯	0.37	0.65	0.74
	平滑	0.36	0.63	0.75
	伽马	0.34	0.66	0.71