Table 3 Available latent fingerprinting reconstruction approaches
Ref. | Year | Description | Database | Limitation | Results |
|---|---|---|---|---|---|
(Wong & Lai, 2020) | 2020 | CNN-based fingerprint reconstruction from the corrupted image | MOLF, FVC2002 DB1 and FVC2004 DB1 | Unsuccessful in extremely low contrast and noisy images | Accuracy = 84.10% |
(Lee et al., 2020) | 2020 | Deep Neural Network–based approach for recovery of latent fingerprints | NIST Special Database 4 |  | At reconstruction weight = 150 FMR0.01% = 66% FMR0.1% = 93% FMR 1% = 100% |
(Xu et al., 2020) | 2020 | Generative adversarial network (GAN) based data augmentation scheme to improve the reconstruction | NIST SD14 and MOLF DB 1,2,3 were used at the augmentation stage. | Â | Matching accuracy (%) NIST SD27: Rank25 = 82.17% IIITD: Rank25 = 95.12% MOLF DB4: Rank25 = 45.88% |
(Joshi et al., 2019b) | 2019 | Reconstruction is done using generative convolutional networks. | Gallery datasets like Lumidigm, Secugen, Crossmatch are used | False minutiae generation is a challenge | Rank 25 Lumidigm = 16.14% Secugen = 13.27% Crossmatch = 12.66% |
(Dabouei et al., 2018) | 2018 | ID preserving generative adversarial network is used for partial latent fingerprint reconstruction | IIIT-Delhi latent fingerprint database and IIIT-Delhi MOLF database | Minutiae are not directly extracted from the latent input fingerprints. | Rank 10 accuracy = 88.02% (IIIT-Delhi latent fingerprint database) rank 50 accuracy = 70.89% IIIT-Delhi MOLF matching |
(Li et al., 2018) | 2018 | Multiscale dictionaries with texture components are used. | NIST SD27 | Computation for false minutiae removal and repetitive minutiae removal is very high. | The average orientation estimation error (in degrees) is 16.38 |
(Kaushal et al., 2016) | 2016 | An analytical framework is proposed | NIST SD-27 | Different filter used for different images | False acceptance rate = 27% |
(Zhou et al., 2016) | 2016 | Partial fingerprint indexing–based algorithm is proposed | FVC 2000 DB2a, FVC2002 DB1a and NIST SD 14 | Indexing is difficult to apply on a very large database | Average penetration rate on FVC2002 DB1a when hit rate is 100% = 3.51% |
(Cao & Jain, 2015) | 2015 | ConvNet–based approach for latent orientation field estimation | NIST SD27 | When latent overlaps with strong background noise, global orientation patch dictionary and ridge structure dictionary approaches do not work well | The average root-mean-square deviation (RMSD) is 13.51 as compared with other algorithms. |
(http://www.ijirset.com/upload/2017/may/269_Criminal.pdf, 2019) | 2015 | Dictionary-based approach | FVC2002, NIST SD4, | Dictionary lookup is a slow process | Improvement in reconstructed image (visual inspection) |
(Zhou et al., 2013) | 2013 | Reconstruction of partial fingerprints | Self-created images | Tested on few images only that are of good quality | Improvement in reconstructed image (visual inspection) |
(Wang et al., 2007) | 2007 | FOMFE-based approach is proposed | FVC2002 Db1a database and NIST Special Database 14 (SDB14) | Â | At feature vector length = 15 Penetration rate = 0.21 |