Overview of Image Inpainting and Forensic Technology

Table 5

Comparison of deep neural network-based forensics.

Category	Source	Method	Detection type	Perf.
Image generating feature-based method	[122]	Histograms of discrete cosine transform (DCT) coefficients	Double JPEG compression	0.98 (AUC)
	[123]	Resampling features, deep learning, and LSTM	Image manipulation	0.9138 (AUC)
	[124]	Camera models	Image authenticity.	0.908 (AUC)
	[125]	Camera response functions	Splicing and copy-move operations	0.97 (accuracy)
	[126]	Residual-based local descriptors	Image manipulation	100.00% (accuracy)
	[127]	Tampered artifacts and noise features	Image manipulation	0.95 (AUC)
	[128]	LSTM and CNN; similar patches	Image object removal	97.89% (accuracy)
Operation feature-based method	[129]	Median filtering residual; Modified CNN;	Compressed image	96.84% (accuracy)
	[130]	Hybrid CNN-LSTM; resampling features	Copy-clone, object splicing, and removal	94.8% (accuracy)
	[131]	Block-like features; self-correlations between different blocks,	Copy-move	0.7572 (F-score)
	[132]	BusterNet; visual artifacts; visual similarities	Copy-move	80.48% (accuracy)
	[133]	Similarity between feature vectors of image overlapped subblocks	Copy-move	0.93 (F-measure)
	[134]	Dense-InceptionNet	Copy-move	0.7058 (precision)
	[135]	Adaptive attention and residual refinement Network；Atrous spatial pyramid pooling； correlation maps	Copy-move	0.8488 (AUC)
CNN structure or loss function changed method	[136]	CNN with weighted cross-entropy loss function	Patch-based operation.	98.3% (AP)
	[137]	Decoder network and feature pyramid network	Patch-based operation	98.99% (TPR)
	[138]	Encoder-decoder network structure; using a label matrix and the weighted cross-entropy as the loss function	Recapture image	99.74% (accuracy)
	[139]	CNNs and the segmentation-based multiscale analysis	Splicing and copy-move	0.4063 (F1-score)
	[140]	Segmentation-based key point distribution strategy and adaptive over segmentation method	Copy-move	0.7627 (F1-measure)
	[141]	CNN-based framework; full-resolution information	Splicing, copy-move	0.886 (AUC)
	[142]	Modified CNN by demosaicing algorithms; mosaic inconsistencies	Splicing, inpainting, or copy-move	0.926 (AUC)
	[143]	Constrained R-CNN	Splicing, copy-move, and removal.	0.992 (AUC)
	[144]	Dense fully convolutional encoder-decoder architecture with dense connections and dilated convolutions	Commonly used editing tools and operations in photoshop	0.99 (AUC)
	[145]	ResNet; image residuals	Remove; CNN inpainting	97.97 (precision)
	[146]	Multibranch CNN architecture	Copy-move	0.920 (F1-score)
	[147]	RGB-N, MSCNNs, DCNNs	Splicing, copy-move	0.7328 (precision)
	[148]	Feature pyramid network, stagewise-weighted cross-entropy; Loss function	JPEG compression, scaling	0.9967 (F1-score)
	[149]	CNN with CRF-based attention model	Splicing, copy-move	0.804 (F1-score)
	[150]	Dense self-attention encoders	Copy-move	0.883 (AUC)