Image Manipulation Detection Algorithm Based on Improved RGB-N

Fig. 6. Schematic diagram of positive and negative sample selection method. (a) Spliced image; (b) source image of mosaic target; (c) manipulation target label

Download full size

Fig. 7. Training samples and labeled samples. (a) Target splicing; (b) target erasure and repair

Download full size

Fig. 8. Visual effects of training different filter parameters. (a) None; (b) KB kernel; (c) second order linear kernel; (d) KB kernel and second order linear kernel

Download full size

Fig. 9. Visual outputs of model. (a) Target splicing; (b) target erasure and repair; (c) normal images

Download full size

Table 1. Results of each algorithm
View table
Table 1. Results of each algorithm
Algorithm F1 score M /%
ELA 0.235 1.8
DCT 0.434 2.3
NOI1 0.287 1.9
Faster-RCNN 0.570 16.1
EXIF-Consistency 0.683 3.7
RGB-N 0.722 16.7
Proposed algorithm 0.759 0.2

Table 2. F1 score of each algorithm when using different quality factors for compression

View table

Table 2. F1 score of each algorithm when using different quality factors for compression

Algorithm	QF 100	QF 90	QF 80	QF 70	QF 60	QF 50
ELA	0.235	0.231	0.229	0.223	0.215	0.207
DCT	0.434	0.205	0.198	0.185	0.103	0.096
NOI1	0.287	0.285	0.281	0.274	0.258	0.235
Faster-RCNN	0.570	0.570	0.567	0.564	0.559	0.550
EXIF-Consistency	0.683	0.678	0.677	0.671	0.661	0.653
RGB-N	0.722	0.722	0.719	0.716	0.713	0.708
Proposed algorithm	0.759	0.759	0.759	0.754	0.742	0.738

Table 3. Results of training different filter parameters
View table
Table 3. Results of training different filter parameters
Training kernel F1 score M /%
None 0.718 17.9
KB kernel 0.722 16.7
Second order kernel 0.684 18.5
KB and second order kernel 0.659 19.3

Table 4. Results of ablation experiment

View table

Table 4. Results of ablation experiment

	Self-attention module	Average pooling	Authenticity judgement module		Sample selection	F1 score	M /%
	Self-attention module	Average pooling	Loss	Loss and output	Sample selection	F1 score	M /%
Faster-RCNN						0.570	16.1
Faster-RCNN					√	0.572	11.6
Proposed algorithm						0.722	16.7
	√					0.753	16.9
	√	√				0.754	16.0
	√	√	√			0.759	11.5
	√	√	√	√		0.758	3.5
	√	√	√	√	√	0.759	0.2

Table 5. Results of noise network pooling method
View table
Table 5. Results of noise network pooling method
Average pooling layer F1 score M /%
None 0.753 16.9
U1 0.754 16.1
U1+U2 0.754 16.0
U1+U2+U3 0.746 17.1
U1+U2+U3+U4 0.741 17.8

Table 6. Results of different loss functions
View table
Table 6. Results of different loss functions
Loss function F1 score M /%
BCE 0.753 0.8
Dice loss 0.757 0.2
BCE+Dice Loss 0.759 0.2

Tools

Get Citation

Copy Citation Text

Haoyue Liu, Wenwei Ma, Xiao Fu, Chengxiu Shen, Yaling Wang. Image Manipulation Detection Algorithm Based on Improved RGB-N[J]. Laser & Optoelectronics Progress, 2021, 58(22): 2211003

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites