Image Forgery Detection Using Tamper-Guided Dual Self-Attention Network with Multiresolution Hybrid Feature

Table 1

Summary of existing image forgery detection methods.

Method	Backbone	Forensic clue	Fusion method	Training data	Localization
ELA [4]	—	Error level analysis	—	Authentic, tamper	Block-level
NOI [5]	—	Noise-inconsistency	—	Authentic, tamper	Block-level
CFA [6]	—	Local CFA inconsistency	—	Authentic, tamper	Block-level
J-LSTM [7]	Patch-LSTM	RGB	—	Tamper	Pixel-level
RGB-N [8]	Faster R-CNN	RGB, noise-inconsistency	Bilinear pooling	Tamper	Object-level
ManTra-net [9]	Wider VGG	RGB, noise-inconsistency	Feature concatenation	Tamper	Pixel-level
FCN [3]	-	RGB	—	Tamper	Pixel-level
CR-CNN [10]	Mask R-CNN	Noise-inconsistency	—	Tamper	Pixel-level
GSR-net [11]	Deeplabv2	RGB	—	Tamper	Pixel-level
Ours	HRNet-48	RGB, noise-inconsistency	Multi-resolution concatenation, tamper-guided dual self-attention	Authentic, tamper	Image-level
					Pixel-level