Skip to main content

Table 4 Available latent fingerprint matching approaches

From: An investigation of latent fingerprinting techniques

Ref. Year Description Database Limitation Results
(Deshpande et al., 2020) 2020 A clustered minutiae-based scale and rotation invariant fingerprint matching method is proposed FVC2004 and NIST SD27 criminal fingerprint databases Matching efficiency is poor in cases where sufficient clustered minutiae set is obtained. 97.5% and 100% of Rank-1 identification accuracy respectively on plain FVC2004 dataset.
Rank-1 identification accuracy of 88.8% and 93.80% respectively on NIST database (LMS, CLMP algorithms)
(Manickam et al., 2019b) 2019 Matching using SIFT feature FVC2004 and IIIT-latent fingerprint Database size is small.
The feature set used is small.
Linear index of fuzziness
IIIT-latent fingerprint = 0.2702
FVC2004 database 1 = 0.1912
FVC2004 database 2 = 0.2008
(Ezeobiejesi & Bhanu, 2018) 2018 Matching is patch-based using a deep learning approach. NIST SD27 The approach does not work well with mixed image resolutions Rank-20 identification rate = 93.65%
(Lin & Kumar, 2018) 2018 Minutia Spherical Coordinate Code is used for matching AFIS data and NIST special data27 There are many redundancies in MCC and MSCC’s feature Rank-1 recognition rate = 49.2%
(Ezhilmaran & Adhiyaman, 2017) 2017 Descriptor-based Hough transform used for matching (NIST SD27 and WVU latent databases Latent matching is slow Rank-1 accuracy = 53.5%
(Zhou et al., 2017) 2017 The fusion of various extended features to improve performance NIST SD4, SD14, and SD27 databases The separation of feature extraction and matching in automatic systems leads to some information loss. Rank-1 identification rate of 74% was achieved
(Medina-Pérez et al., 2016) 2016 Local and global matching NIST SD27(A) Approaches used for level 2 and level 3 matching are different which decreases accuracy Rank-1 identification accuracy of 74%
(Zheng et al., 2015) 2015 CovNet and Dictionary-based approach NIST SD27 and WVU latent databases Recognition performance can be improved. Speed of feature extraction and comparison can be raised Superior performance of texture (virtual minutiae) template on bad and ugly images (i.e. 47.1%; good-quality image is 83%)
(Cao et al., 2014) 2014 Extended features used for performance enhancement NIST SD27 Differences in the approach used by latent experts and automatic matches.
Prone to false minutiae and distortions.
Information loss due to separation of automatic matching and feature extraction.
Identification rate
Good images = 90%
Bad images = 85%
Ugly images = 71%
(Lan et al., 2019). 2014 A new feature Distinctive Ridge Point (DRP) is proposed NIST14 and NIST4 High ridge point dependence with minutiae. Rank-1 accuracy = 70.9%
(Jain & Feng, 2010b) 2014 Algorithm based on directional information FVC2004 DB1, Tsinghua Distorted Fingerprint database, NIST SD27 database and NIST SD30 database. Do not consider the rotation and translation of the whole image Identification rate using
Correlation score = 80%
Verifinger score = 82%
(Feng, 2012) 2012 Descriptor-based Hough transform algorithm NIST SD27 and WVU latent database Do not work well with overlapping fingerprints Identification rate = 67%
(https://www.nist.gov/itl/iad/image-group/nist-special-database-2727a, 2019) 2012 Two minutiae-based descriptors are proposed FingerPass and Multi-Sensor Optical and Latent Fingerprint For different sensor technology, performance is not good
Poor performance when fingerprints were distorted
False matching rate = 1.166%
Equal error rate = 0.41%
(Jain & Feng, 2010b). 2010 Fusion of minutiae NIST SD27 Orientation field reconstruction to be improved Identification rate = 65% (manually marked minutiae)
(Jain & Cao, 2015) 2009 Fusion of plain and rolled fingerprints ELFT-EFS Public Challenge Dataset Does not appear to be a common practice in law enforcement Rank-1 identification rate of 83.0%
(Feng et al., 2009) 2009 Fusion of plain and rolled fingerprints NIST SD27 The distortion between rolled and plain fingerprints is not taken into account.
Manual extraction of level 1 and level 2 features
Rank-1 identification rate = 83.0%
(Feng & Jain, 2008) 2008 Filtering-based approach NIST SD27 Singular point detection is not accurate
More filtering approaches can be used to improve performance. Background database is small
Rank-1 matching accuracy = 73.3%