Research Article
Semantic Modeling and Pixel Discrimination for Image Manipulation Detection
Table 1
The pixel-level F1 score/AUC comparison on four standard datasets.
| | Pixel-level | Forgery object-level | Dual branch | “AND” | “OR” | “MRF” | F1 | AUC | F1 | AUC | F1 | AUC | F1 | AUC | F1 | AUC |
| NIST′16 [16] | 0.780 | 0.796 | 0.717 | 0.912 | 0.739 | 0.854 | 0.828 | 0.892 | 0.807 | 0.929 | CASIA [17, 18] | 0.751 | 0.713 | 0.688 | 0.844 | 0.744 | 0.743 | 0.742 | 0.851 | 0.789 | 0.862 | Coverage [19] | 0.551 | 0.723 | 0.433 | 0.802 | 0.531 | 0.808 | 0.561 | 0.828 | 0.563 | 0.828 | Columbia [20] | 0.831 | 0.641 | 0.398 | 0.756 | 0.529 | 0.809 | 0.776 | 0.803 | 0.833 | 0.801 |
|
|