Laser & Optoelectronics Progress, Volume. 62, Issue 8, 0815006(2025)
Improved RRU-Net for Image Splicing Forgery Detection
Figures & Tables(16)
Fig. 7. Generated partial forgery plots and ground truth. (a1)‒(a3) Tamper images; (b1)‒(b3) tamper images corresponds to ground truth
Fig. 8. Partial forgery of images with different algorithm visualizations. (a) Forgery images; (b) ground truth; (c) DeepLab V3 visualization of inspection map; (d) ManTra-Net visualization of inspection map;(e) visualization of detection plot of PSCC-Net; (f) MVSS-Net++ visual detection map; (g) U-Net visualization of inspection map;(h) RRU-Net visualization of inspection map;(i) this algorithm visualizes detection map
Fig. 9. Results of different models for different datasets under Gaussian noise attack. (a) Homemade; (b) CASIA 1.0; (c) IMD 2020
Fig. 10. Results of different models for different datasets under JPEG compression attack. (a) Homemade; (b) CASIA 1.0; (c) IMD 2020
Fig. 11. Results of different models for different datasets under darkening attack. (a) Homemade; (b) CASIA 1.0; (c) IMD 2020
Table 1. Dataset classification
View tableTable 1. Dataset classification
Dataset Homemade CASIA 1.0 IMD 2020 AU Splic AU Splic Training sets 18861 – – – – Validation sets 990 – – – – Testing sets 990 462 461 253 300
Table 2. Ablation experiments
View tableTable 2. Ablation experiments
Method Loss F1 RRU-Net 0.1646 0.7615 RRU-Net+F 0.0872 0.8046 RRU-Net+A2-Nets 0.1304 0.7816 RRU-Net+F+A2-Nets 0.0390 0.8112 Ours=RRU-Net+N+F+A2-Nets 0.0185 0.9456
Table 3. F1 and accuracy of different algorithms on homemade dataset
View tableTable 3. F1 and accuracy of different algorithms on homemade dataset
Method Accuracy F1 DeepLab V3 0.7328 0.7392 ManTra-Net 0.9643 0.7731 PSCC-Net 0.8182 0.8906 MVSS-Net++ 0.8733 0.7909 U-Net 0.6196 0.6758 RRU-Net 0.7351 0.7615 Ours 0.9063 0.9456
Table 4. F1 and accuracy of different algorithms on common datasets
View tableTable 4. F1 and accuracy of different algorithms on common datasets
Method CASIA 1.0 IMD 2020 Accuracy F1 Accuracy F1 DeepLab V3 0.6755 0.5411 0.6403 0.6670 ManTra-Net 0.8692 0.1343 0.8843 0.1123 PSCC-Net 0.8759 0.5457 0.9291 0.6779 MVSS-Net++ 0.8676 0.7499 0.7420 0.6829 U-Net 0.7154 0.6124 0.8200 0.5580 RRU-Net 0.7366 0.7611 0.8411 0.7142 Ours 0.9115 0.8583 0.9177 0.8117
Table 5. Robustness F1 variance of different algorithms on different datasets
View tableTable 5. Robustness F1 variance of different algorithms on different datasets
Methods Homemade CASIA 1.0 IMD 2020 Noise Var JPEG Var Bright Var Noise Var JPEG Var Bright Var Noise Var JPEG Var Bright Var DeepLab V3 3.8×10-5 1.2×10-5 2.8×10-3 1.6×10-3 4.4×10-6 3.9×10-4 1.4×10-5 2.4×10-6 7.3×10-3 ManTra-Net 1.5×10-2 1.7×10-3 2.9×10-3 4.1×10-4 9.1×10-5 1.1×10-3 3.3×10-5 8.9×10-5 1.1×10-3 PSCC-Net 7.6×10-5 3.1×10-5 4.6×10-4 1.8×10-4 7.3×10-5 1.2×10-4 1.0×10-3 3.7×10-4 1.8×10-4 MVSS-Net++ 1.4×10-4 1.6×10-4 5.9×10-4 1.9×10-4 7.3×10-5 1.8×10-4 6.3×10-5 2.1×10-5 8.4×10-5 U-Net 3.8×10-3 9.9×10-5 4.4×10-3 1.3×10-2 2.6×10-3 9.4×10-4 1.3×10-3 9.3×10-5 9.0×10-4 RRU-Net 8.2×10-5 6.0×10-7 3.6×10-3 1.5×10-4 7.0×10-4 1.4×10-3 9.3×10-5 4.3×10-4 1.7×10-3 Ours 2.3×10-4 1.1×10-4 2.0×10-4 3.0×10-4 7.9×10-5 9.2×10-5 3.9×10-4 2.9×10-5 7.0×10-4
Tools
Get Citation
Copy Citation Text
Ying Ma, Yilihamu Yaermaimaiti, Shuoqi Cheng, Yazhou Su. Improved RRU-Net for Image Splicing Forgery Detection[J]. Laser & Optoelectronics Progress, 2025, 62(8): 0815006
Paper Information
Category: Machine Vision
Received: Jul. 9, 2024
Accepted: Oct. 8, 2024
Published Online: Mar. 21, 2025
The Author Email:
CSTR:32186.14.LOP241655
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20