Image Inpainting Forensics Algorithm Based on Deep Neural Network

Xinshan Zhu; Yongjun Qian; Biao Sun; Chao Ren; Ya Sun; Siru Yao

doi:10.3788/AOS201838.1110005

Journals >Acta Optica Sinica >Volume 38 >Issue 11 >Page 1110005 > Article

Acta Optica Sinica
Vol. 38, Issue 11, 1110005 (2018)

Image Inpainting Forensics Algorithm Based on Deep Neural Network

Xinshan Zhu^1、2、*, Yongjun Qian¹, Biao Sun^1、*, Chao Ren¹, Ya Sun¹, and Siru Yao¹

Author Affiliations

¹ School of Electrical and Information Engineering, Tianjin University, Tianjin 300072, China

² State Key Laboratory of Information Security, Institute of Information Engineering, Chinese Academy of Sciences, Beijing 100093, China

show less

DOI: 10.3788/AOS201838.1110005 Cite this Article Set citation alerts

Xinshan Zhu, Yongjun Qian, Biao Sun, Chao Ren, Ya Sun, Siru Yao. Image Inpainting Forensics Algorithm Based on Deep Neural Network[J]. Acta Optica Sinica, 2018, 38(11): 1110005 Copy Citation Text

show less

Image inpainting process by Criminisi algorithm. (a) Determining the inpainting block; (b) determining the matching block; (c) inpainting the image block; (d) renewal the margin

Fig. 1. Image inpainting process by Criminisi algorithm. (a) Determining the inpainting block; (b) determining the matching block; (c) inpainting the image block; (d) renewal the margin

Download full size

Fig. 2. Image inpainting forensics. (a) Original image; (b) falsified image; (c) image after CNN inpainting forensics

Download full size

Fig. 3. Structural diagram of network

Download full size

Fig. 4. Structural diagram of network

Download full size

Fig. 5. Partial classical images in testing set

Download full size

Fig. 6. Falsified images with different area ratios of falsification regions. (a) 5%; (b) 10%; (c) 20%; (d) 0-5%; (e) 10%-30%; (f) 30%-50%

Download full size

Fig. 7. Detection results of inpainting regions. (a) Original image; (b) mask image; (c) inpainted image; (d) detection results by algorithm in Ref. [14]; (e) detection results by algorithm in Ref. [15]; (f) detection results by proposed method

Download full size

Type of network	Encoder network					Decoder network
Number offeature maps	64	128	256	512	512	256	256	128	64	48
Feature size	256×256	128×128	64×64	32×32	16×16	16×16	32×32	64×64	128×128	256×256

Table 1. Structural parameters of convolution-wide network

Image No.	Method in Ref. [14]		Method in Ref. [15]		Proposed method
Image No.	TPR	FPR	TPR	FPR	TPR	FPR
Fig. 5(a)	66.55	1.24	88.53	0.2	95.03	4.87
Fig. 5(b)	90.42	4.13	85.01	0	98.03	0.16
Fig. 5(c)	86.04	30.09	83.15	0.86	96.78	3.22
Fig. 5(d)	92.42	6.75	92.16	0.27	98.99	0.09
Fig. 5(e)	0	10.09	89.29	14.42	97.69	0.15

Table 2. TPR and FPR of classical images by different algorithms (unit: %)

Image Num.	Method in Ref. [14]		Method in Ref. [15]		Proposed method
Image Num.	TPR	FPR	TPR	FPR	TPR	FPR
200	87.35	14.7	94.75	10.8	95.6	0.8

Table 3. TPR and FPR of images in UCID dataset by different algorithms (unit: %)

Parameter	Mask size /%	TPR /%	FPR /%	T /s
Regularregion	5	98.64	0.34	1.8
	10	97.84	0.24	1.9
	20	96.90	0.15	1.8
Irregularregion	0-10	90.43	0.4	2.1
	10-30	94.88	1.08	2.0
	30-50	96.96	1.74	1.8
Mean	—	95.94	0.674	1.9

Table 4. Detection results of images inpainted by Criminisi algorithm

Parameter	Mask size /%	TPR /%	FPR /%	T /s
Regularregion	5	93.29	0.29	1.9
	10	92.48	0.19	1.8
	20	89.95	0.25	2
Irregularregion	0-10	80.09	0.32	2.1
	10-30	80.36	2.18	2.2
	30-50	73.79	6.17	2.0
Mean	—	84.99	1.57	2

Table 5. Detection results of images inpainted by Shift-map algorithm

Xinshan Zhu, Yongjun Qian, Biao Sun, Chao Ren, Ya Sun, Siru Yao. Image Inpainting Forensics Algorithm Based on Deep Neural Network[J]. Acta Optica Sinica, 2018, 38(11): 1110005

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information