|
| Author | Year | Distortions (alterations) | Features | Limitations |
|
| Yu et al. [19] | 2008 | Rayleigh-like distribution, dithering patterns | Physical modeling of a recapturing process | The model performance is not good in planar surfaces in generic scenes |
| Gao et al. [12] | 2010 | Colour distortions | Statistics-based features, chromaticity covariance matrix | Degradation process directly by the application of mathematical models |
| Cao and Kot [7] | 2010 | Aliasing-like distortions’ colour distortion, blurriness | Blurriness, texture, noise, and colour features | The network performance is not good enough |
| Ke et al. [42] | 2013 | Texture, colour noise, difference histogram | Histogram of image local difference, colour moments | Feature extraction becomes more and more cumbersome |
| Muammar and Dragotti [20] | 2013 | Noise residual features | Used 2DFT of noise residual and theory of cyclostationarity | This method works only for aliasing-free image datasets |
| Zhai et al. [8] | 2013 | Texture features | Based on texture features | Only texture feature has been considered |
| Mahdian et al. [43] | 2015 | Aliasing-like distortions | Used 2DFT of noise residual and theory of cyclostationarity | Images with dark content create detection challenge |
| Thong et al. [14] | 2015 | Aliasing and blurriness | Dictionary approximation error of edge profiles | Performance is degrading with Kodak camera images |
| Ni et al. [44] | 2015 | Colour distortions | Features of colour moments and DCT coefficients | This method is only effective for JPEG images that are compressed images |
| Yang et al. [31] | 2016 | Extract features from dataset using a CNN | Used CNN algorithm | Laplacian filter layer removes valuable information such as colour, which is very helpful in image recapturing detection |
| Samaraweera et al. [45] | 2016 | Texture, HSV colour, and blurriness | Used SVM classifier | Only used images taken by back-facing camera |
| Yang et al. [16] | 2017 | Quality-aware feature and histogram feature | Compression artefacts have been used | This method only works for JPEG compressed images |
| Li et al. [18] | 2017 | Co-occurrence matrices extracted from the residual image are used for detection | Support vector machine and ensemble classifier | Proposed network performance is not very high |
| Choi et al. [29] | 2017 | Extract features from dataset using a CNN | CNN-based approach | Proposed network use only fewer convolutional layers |
| Li et al. [32] | 2017 | Aliasing distortion and noise artefacts | Convolution and recurrent neural network-based | Large fully connected layers increase the model complexity |
| Sun et al. [9] | 2018 | Wavelet characteristics and noise characteristics | Mean value, variance, and skewness, extract noise image using LBP | The images have not been tested on different image formats |
| Zhu and Li [17] | 2018 | Structure distortions | Enhanced residual-based correlation coefficients | Only work for JPEG compressed images |
| Zhu et al. [33] | 2019 | LBP-coded maps | LBP features | Fixed filter layer at the input may ignore some useful information |
| Anjum and Islam [1] | 2019 | Exploit recapture and original image edge profile to classify it | An image’s edge profile can be used to determine its structural level | Only consider edge profile and proposed method cannot identify images obtain from different media |
|
