Skip to main content

Table 2 Available latent fingerprint enhancement approaches

From: An investigation of latent fingerprinting techniques

Ref. Year Description Database Limitation Results
(Joshi et al., 2021) 2021 Direct de-noise the fingerprints and reconstruct the missing ridge structure without explicitly estimating the orientation field using GAN’s IIITD-MOLF
GAN’s are difficult to train as they require a large dataset for accurate results. NFIQ (lower score means better quality) = 2.64
(Agarwal & Bansal, 2021) 2021 The fusion of pores and minutiae at score level is used to re-rank the minutiae-based latent matcher IIITD latent fingerprint database
LivDet 2015 database
Less number of minutiae are used. Additional features such as ridge flow map and ridge quality map can improve the performance. True detection rate RT = 82.89%
Average of the false detection rate RF = 21.2%
(Gupta et al., 2020) 2020 Enhancement and reconstruction of image using the minutiae density and the orientation field direction Fingerprint verification competition 2002 (FVC2002) and fingerprint verification competition 2004 (FVC2004) Only local orientation patterns are considered in the proposed method. Type I attack: TARa = 97.95% on FVC2002 and 94.09% on FVC2004
Type 2 attack: TAR = 49.25% and 50.02% on FVC2002 and FVC2004
(Huang et al., 2020) 2020 A generative adversarial network (GAN) is proposed for the enhancement of latent fingerprint images. NIST SD27 dataset, NIST SD14   Identification rate (%):
Cumulative match characteristics all = 50%
Cumulative match characteristics good = 77%
Cumulative match characteristics bad = 45%
Cumulative match characteristics ugly = 29%
(Horapong et al., 2020) 2020 Two-Stage Spectrum Boosting with Matched Filter and Sparse Autoencoder is used for enhancement IIT-D MOLF latent fingerprint database The proposed method depends on high ridge signal strength initially to boost ridge spectra. Identification rate (%)
Rank 20 = 43%
(Joshi et al., 2021) 2020 A conditional generative adversarial network-based latent fingerprint enhancement algorithm is proposed. IIITD-MOLF and IIITD-MSLF database The proposed algorithm generates spurious features when the ridge information is insufficient NFIQ (lower score means better quality)
= 2.64
(Jhansirani & Vasanth, 2019) 2019 Image enhancement is done using the Gabor function via multiscale patch-based sparse representation NIST SD27 Dictionary creation and lookup is slow The best training performance is 7.8717e obtained at epoch 10.
(Joshi et al., 2019a) 2019 Latent fingerprint enhancement algorithm based on generative adversarial networks is used IIITD-MOLF database and IIITD-MSLFD database Spurious features are generated when the ridge information is insufficient. Matching results: Rank-50 accuracy of 35.66% (DB 1)
30.16% (DB 2)
(Manickam & Devarasan, 2019) 2019 An intuitionistic fuzzy set is used for contrast enhancement of fingerprints Fingerprint verification competition-2004 and IIIT-latent fingerprint database Imperfect matching in case of presence of background noise and nonlinear ridge distortion Matching scores
IIIT-latent fingerprint = 0.2702
FVC2004 database 1 = 0.1912
FVC2004 database 2 = 0.2008
(Manickam et al., 2019a) 2019 Scale-Invariant Feature Transformation (SIFT) is used for the enhancement of an image. FVC2004 and IIIT-latent fingerprint Does not work well with very poor and partial prints Linear index of fuzziness
IIIT-latent fingerprint = 0.2702
FVC2004 database 2 = 0.2008
(Liban & Hilles, 2018) 2018 A hybrid model that is a combination of edge directional total variation model (EDTV) and quality image enhancement with lost minutia reconstruction is used. NIST SD27 database for testing
RMSE, PSNR to measure performance.
Results are not good with ugly images
Overlapping images not considered
RMSE average = 0.018373 (good-quality image)
PSNR average = 82.99068 (good-quality image)
(Chaidee et al., 2018) 2017 The spectral dictionary is used for enhancement NIST SD27 Failure due to the wide bandwidth of filter which leads to noise leakage into enhancement process Identification rate
good-quality print = 76%
bad quality = 59%
ugly quality = 35%
(Liu et al., 2014) 2014 Multiscale Patch Based Sparse Representation used for enhancement NIST SD27 Global ridge structures are ignored
Do not work well for low-quality fingerprints
Identification rate = approx. 64%
(Cao et al., 2014) 2014 Ridge structure dictionary is used for enhancement NIST SD27 and WVU DB Confidence measure is poorly defined for the segmentation and enhancement results. Computational efficiency of the algorithm is low Identification rate
NIST SD27 = 71%
WVU DB = 78%
(Zhang et al., 2013) 2013 Adaptive directional total variation model NIST SD27   Identification accuracy less than 12% (rank20)
(Feng et al., 2012) 2012 Prior knowledge-based approach NIST SD27 The speed of the proposed algorithm is slow with low-quality latents Identification rate:
good-quality print = 60%
bad quality = 24%
ugly quality = 11%
(Yoon et al., 2011) 2011 Enhancement using hypothesized orientation fields NIST SD27 Human markup of minutiae is required.
Performance is poor for bad and ugly-quality latents.
Latent quality assessment is not automatic
Identification rate
good-quality print = 66%
bad quality = 50%
ugly quality = 40%
(Yoon et al., 2010) 2010 Polynomial model and zero pole model NIST SD27 Uses fixed ridge frequency Identification accuracy = 35% (rank1)