Beyond Mortality: Advancements in Post-Mortem Iris Recognition through Data Collection and Computer-Aided Forensic Examination
Authors: Rasel Ahmed Bhuiyan, Parisa Farmanifard, Renu Sharma, Andrey Kuehlkamp, Aidan Boyd, Patrick J Flynn, Kevin W Bowyer, Arun Ross, Dennis Chute, Adam Czajka
Organizations: University of Notre Dame, Notre Dame, IN 46556, USA. · Department of Computer Science and Engineering, Michigan State University, East Lansing, MI 48824, USA. · Meta, New York, NY 10003, USA. · Dutchess County Medical Examiner’s Office, Poughkeepsie, NY 12601, USA.
Post-mortem iris recognition brings both hope to the forensic community (a short-term but accurate and fast means of verifying identity) as well as concerns to society (its potential illicit use in post-mortem impersonation). These hopes and concerns have grown along with the volume of research in post-mortem iris recognition. Barriers to further progress in post-mortem iris recognition include the difficult nature of data collection, and the resulting small number of approaches designed specifically for comparing iris images of deceased subjects. This paper makes several unique contributions to mitigate these barriers. First, we have collected and we offer a new dataset of NIR (compliant with ISO/IEC 19794-6 where possible) and visible-light iris images collected after demise from 259 subjects, with the largest PMI (post-mortem interval) being 1,674 hours. For one subject, the data has been collected before and after death, the first such case ever published. Second, the collected dataset was combined with publicly-available post-mortem samples to assess the current state of the art in automatic forensic iris recognition with five iris recognition methods and data originating from 338 deceased subjects. These experiments include analyses of how selected demographic factors influence recognition performance. Thirdly, this study implements a model for detecting post-mortem iris images, which can be considered as presentation attacks. Finally, we offer an open-source forensic tool integrating three post-mortem iris recognition methods with explainability elements added to make the comparison process more human-interpretable.
Figures & tables
Dataset name
Number
Number
Number of cases
Sensors used
Min → Max
Metadata
of subjects
of images
with ante-
PMI
(eyes)
(NIR / RGB)
and post-mortem
(hours)
image pairs
Warsaw-BioBase-Post-Mortem-Iris
17 (34)
574 / 626
0
IriShield M2120U
5 → 814
For all:
v1.0 [ 10 ]
Olympus TG-3
PMI
Warsaw-BioBase-Post-Mortem-Iris
20 (40)
1,094 / 1,023
0
IriShield M2120U
7 → 453
gender
TABLE I: Main novel characteristics of the newly-collected NIJ-2018-DU-BX-0215 post-mortem iris dataset compared with all existing publicly-available research datasets.
Fig. 1: Number of images as a function of subject’s age in the NIJ-2018-DU-BX-0215 dataset.
Fig. 2: Number of images as a function of the PMI in the NIJ-2018-DU-BX-0215 dataset.
Fig. 3: Number of samples for each subject in all acquisition sessions in the NIJ-2018-DU-BX-0215 dataset, shown separately for the left and right eyes.
Fig. 4: Visualization of the progression of post-mortem deformations in time for the case with the longest PMI (1,674 hours) in the NIJ-2018-DU-BX-0215 dataset. Corresponding pictures in lower rows illustrate the iris segmentation results. Numbers below pictures indicate the PMI for a given sample.
Age (years)
Gender
1-33
34-66
67-99
Male
Female
Total Subjects
37
210
81
238
91
Impostor Pairs
8,450
257,319
29,443
455,486
52,691
Genuine Pairs
35,88
29,198
8,743
57,687
21,208
Average PMI
50.7
50.7
50.7
85.3
85.3
Average Age
26.15
53.53
75.56
55.8
57.7
TABLE II: Data statistics after balancing along the subject age and gender dimensions.
Fig. 5: The comparison score distributions for the combined NIJ-2018-DU-BX-0215 + Warsaw NIR datasets , obtained for five different iris matching algorithms. The main performance metrics (d’, Failure to Match, Equal Error Rate and Area Under ROC curve) are also shown, and results are split by the PMI range : 0-24h (top row), 0-72h (second row), 0-240h (third row), and all PMIs combined (bottom row).
Fig. 6: Same as in Fig. 5 in the main paper (the comparison score distributions obtained for five different iris matching algorithms) except that only scored obtained for NIJ-2018-DU-BX-0215 NIR samples are shown, and results are split by the PMI range : 0-24h (top row), 0-72h (second row), 0-240h (third row), and all PMIs combined (bottom row).
Fig. 7: Same as in Fig. 5 in the main paper (the comparison score distributions obtained for five different iris matching algorithms) except that only scored obtained for Warsaw NIR samples are shown, and results are split by the PMI range : up to 24h (top row), up to 72h (second row), up to 240h (third row), and all PMIs combined (bottom row).
Fig. 8: The comparison score distributionss for the combined NIJ-2018-DU-BX-0215 + Warsaw RGB datasets , obtained for five different iris matchers.
Fig. 9: The comparison score distributions for the combined NIJ-2018-DU-BX-0215 + Warsaw datasets obtained in a cross-wavelength experiment.
Fig. 10: The comparison score distributions for the ante-mortem vs post-mortem samples , obtained for five different iris matchers.
Fig. 11: The comparison score distributions for the combined NIJ-2018-DU-BX-0215 and Warsaw datasets , obtained for five different iris matching algorithms, and split by gender : results for male and female subjects are shown in the top and bottom rows, respectively.
ANOVA
Kruskal-Wallis
Method
Statistics
p -value
Statistics
p -value
HDBIF
462.66
2.52e-29
44.26
2.87e-11
USIT
4005.54
3.18e-55
44.26
2.87e-11
OSIRIS
10.48
0.0019
8.83
0.0029
DGR
214.83
3.68e-21
44.26
2.87e-11
VeriEye
22.49
1.41e-05
18.50
1.69e-05
TABLE III: Statistical hypothesis test results for male vs female average d-prime values.
Fig. 12: The comparison score distributions for the combined NIJ-2018-DU-BX-0215 and Warsaw datasets , obtained for five different iris matching algorithms and shown separately for each subject’s age group in the rows, from top to bottom: 1-33 years old, 34-66 years old and 67-99 years old.
ANOVA
Kruskal-Wallis
Method
Statistics
p -value
Statistics
p -value
HDBIF
346.09
3.82e-42
65.44
6.14e-15
USIT
316.21
1.23e-40
68.42
1.38e-15
OSIRIS
1762.93
4.00e-71
79.12
6.59e-18
DGR
1219.85
2.27e-64
79.12
6.59e-18
VeriEye
170.92
7.30e-31
68.83
1.13e-15
TABLE IV: Statistical hypothesis test results for average d-prime values across three age groups.
Fig. 13: The presentation attack detection score distributions obtained for all eight post-mortem detection experiments described in Section V-C . Experiments 1 & 2: Pre-trained model tested on Warsaw and NIJ-2018-DU-BX-0215 without fine-tuning; Experiments 3 & 4: Fine-tuned model tested cross-dataset (Warsaw and NIJ-2018-DU-BX-0215); Experiments 5 & 6: Model fine-tuned with 50 samples and tested cross-dataset (Warsaw and NIJ-2018-DU-BX-0215); Experiments 7 & 8: Model fine-tuned with 5 samples and tested cross-dataset (Warsaw and NIJ-2018-DU-BX-0215) .
Exp.
1-BPCER @ 0.01%
1-BPCER @ 1%
3
98.29%
98.29%
4
95.29%
95.29%
5
90.89%
93.00%
6
83.74%
95.03%
7
43.04%
52.22%
8
41.84%
49.78%
TABLE V: True detection rate (1-BPCER) at APCER of 0.01% and 1% for Experiments 3 to 8 using D-NetPAD. BPCER stands for Bona fide Presentation Classification Error Rate, and APCER stands for Attack Presentation Classification Error Rate [ 51 ] .
Dataset
Min PMI(h)
PA (Min)
Max PMI(h)
PA (Max)
Warsaw
5.0
0.3046
814.0
0.4247
NIJ
1.0
0.6770
1674.0
0.7583
TABLE VI: Minimum and maximum PMI and the corresponding presentation attack (PA) scores for each dataset in Experiments 3 and 4.
Fig. 14: Post-mortem iris images that were misclassified as live by the D-NetPAD model. The top row shows samples from Experiment 4 (tested on Warsaw), and the bottom row shows samples from Experiment 3 (tested on NIJ-2018-DU-BX-0215).
Fig. 15: Four screens of the PMExpert tool illustrating how the forensic experts can evaluate matching results using different post mortem-specific iris recognition methods along with their visual explanations (from left to right: TLPIM, HDBIF and PBM methods), as well as get the ISO metrics for each image (the upmost right screen).
Figure 22Figure 23
Renu Sharma completed her PhD in the Computer Science and Engineering Department at Michigan State University (MSU) in 2022. Currently, she is with Amazon as Applied Scientist. Renu also has 8 years of experience in Research & Development at CDAC Mumbai, India.
Fig. 16: ISO/IEC 29794-6 quality metrics density heatmaps split by gender and type of comparison (genuine and impostor) for the combined NIJ-2018-DU-BX-0215 and Warsaw datasets.
Fig. 17: ISO/IEC 29794-6 quality metrics density heatmaps for three age groups, and split by the type of comparison (genuine and impostor) for the combined NIJ-2018-DU-BX-0215 and Warsaw datasets.