Estimating sex and age for forensic applications using machine learning based on facial measurements from frontal cephalometric landmarks

TL;DR

This paper presents a machine learning approach using facial landmarks to estimate age and sex for forensic identification, achieving high accuracy on a large Brazilian dataset.

Contribution

It introduces a neural network classifier trained on a new extensive facial landmark dataset for age and sex estimation in forensic applications.

Findings

Sex estimation accuracy > 0.85 for individuals over 14 years old

Age estimation accuracy of 0.72 for 5-year intervals

High accuracy in age group classification over 14 and 18 years

Abstract

Facial analysis permits many investigations some of the most important of which are craniofacial identification, facial recognition, and age and sex estimation. In forensics, photo-anthropometry describes the study of facial growth and allows the identification of patterns in facial skull development by using a group of cephalometric landmarks to estimate anthropological information. In several areas, automation of manual procedures has achieved advantages over and similar measurement confidence as a forensic expert. This manuscript presents an approach using photo-anthropometric indexes, generated from frontal faces cephalometric landmarks, to create an artificial neural network classifier that allows the estimation of anthropological information, in this specific case age and sex. The work is focused on four tasks: i) sex estimation over ages from 5 to 22 years old, evaluating the interference of age on sex estimation; ii) age estimation from photo-anthropometric indexes for four age intervals (1 year, 2 years, 4 years and 5 years); iii) age group estimation for thresholds of over 14 and over 18 years old; and; iv) the provision of a new data set, available for academic purposes only, with a large and complete set of facial photo-anthropometric points marked and checked by forensic experts, measured from over 18,000 faces of individuals from Brazil over the last 4 years. The proposed classifier obtained significant results, using this new data set, for the sex estimation of individuals over 14 years old, achieving accuracy values greater than 0.85 by the F_1 measure. For age estimation, the accuracy results are 0.72 for measure with an age interval of 5 years. For the age group estimation, the measures of accuracy are greater than 0.93 and 0.83 for thresholds of 14 and 18 years, respectively.

Tables

Table 1. TABLE I: The detailed multilayer perceptron structure used to execute the tests.

# Input layer	209 (208 PAIs + sex attribute)
# Hidden layer	1
# Neuron in hidden layer	128
# Epochs	500
Activation function in hidden layer	Sigmoid
Activation function in output layer	Softmax
Learning rate	0.01
Momentum	0.9
Optimizer	Adamax

Table 2. TABLE II: Group A: Structure of tests for sex estimation.

Test	Target	Age as input	Qty. Test
1	Sex	Yes	17
2	Sex	No	1

Table 3. TABLE III: Group B: Structure of tests for age estimation.

Test	Target	Sex as input
1	6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22	No/Female/Male
2	6,8,10,12,14,16,18 ,20,22	No/Female/Male
3	6,10,14,18,22	No/Female/Male
4	5,10,15,20	No/Female/Male

Table 4. TABLE IV: Group C: Age group, structure of tests for threshold of 14 and 18 years.

Test	Target	Sex as input
1	Older/under 14 years	No/Female/Male
2	Older/under 18 years	No/Female/Male

Table 5. TABLE V: Description of the 208 photo-anthropometric indexes (PAIs).

PAI	Landmarks	Description
PAI-0	al_r-al_l $|$ al_l-al_r	Nose width
PAI-1	al_r-ch_r $|$ al_l-ch_l	Wing of the nose - Labial Commissure (Same side)
PAI-2	al_r-ch_l $|$ al_l-ch_r	Wing of the nose - Labial Commissure (Different side)
PAI-3	al_r-cph_r $|$ al_l-cph_l	Wing of the nose - Crista philtri (Same side)
PAI-4	al_r-cph_l $|$ al_l-cph_r	Wing of the nose - Crista philtri (Different side)
PAI-5	al_r-ec_r $|$ al_l-ec_l	Wing of the nose - Ectocanthion (Same side)
PAI-6	al_r-ec_l $|$ al_l-ec_r	Wing of the nose - Ectocanthion (Different side)
PAI-7	al_r-en_r $|$ al_l-en_l	Wing of the nose - Endocanthion (Same side)
PAI-8	al_r-en_l $|$ al_l-en_r	Wing of the nose - Endocanthion (Different side)
PAI-9	al_r-g $|$ al_l-g	Wing of the nose - Glabella
PAI-10	al_r-gn $|$ al_l-gn	Wing of the nose - Chin
PAI-11	al_r-go_r $|$ al_l-go_l	Wing of the nose - Gonion (Same side)
PAI-12	al_r-go_l $|$ al_l-go_r	Wing of the nose - Gonion (Different side)
PAI-13	al_r-il_r $|$ al_l-il_l	Wing of the nose - Lateral iris (Same side)
PAI-14	al_r-il_l $|$ al_l-il_r	Wing of the nose - Lateral iris (Different side)
PAI-15	al_r-im_r $|$ al_l-im_l	Wing of the nose - Medial iris (Same side)
PAI-16	al_r-im_l $|$ al_l-im_r	Wing of the nose - Medial iris (Different side)
PAI-17	al_r-li $|$ al_l-li	Wing of the nose - Lower lip
PAI-18	al_r-ls $|$ al_l-ls	Wing of the nose - Upper lip
PAI-19	al_r-mid $|$ al_l-mid	Wing of the nose - Midnasal
PAI-20	al_r-n $|$ al_l-n	Wing of the nose - Nasion
PAI-21	al_r-pu_r $|$ al_l-pu_l	Wing of the nose - Pupil (Same side)
PAI-22	al_r-pu_l $|$ al_l-pu_r	Wing of the nose - Pupil (Different side)
PAI-23	al_r-sn $|$ al_l-sn	Wing of the nose - Base of the nose
PAI-24	al_r-sto $|$ al_l-sto	Wing of the nose - Stomion
PAI-25	al_r-zy_r $|$ al_l-zy_l	Wing of the nose - Zygion (Same side)
PAI-26	al_r-zy_l $|$ al_l-zy_r	Wing of the nose - Zygion (Different side)
PAI-27	ch_r-ch_l $|$ ch_l-ch_r	Mouth width
PAI-28	ch_r-cph_r $|$ ch_l-cph_l	Labial Commissure - Crista philtri (Same side)
PAI-29	ch_r-cph_l $|$ ch_l-cph_r	Labial Commissure - Crista philtri (Different side)
PAI-30	ch_r-ec_r $|$ ch_l-ec_l	Labial Commissure - Ectocanthion (Same side)
PAI-31	ch_r-ec_l $|$ ch_l-ec_r	Labial Commissure - Ectocanthion (Different side)
PAI-32	ch_r-en_r $|$ ch_l-en_l	Labial Commissure - Endocanthion (Same side)
PAI-33	ch_r-en_l $|$ ch_l-en_r	Labial Commissure - Endocanthion (Different side)
PAI-34	ch_r-g $|$ ch_l-g	Labial Commissure - Glabella
PAI-35	ch_r-gn $|$ ch_l-gn	Labial Commissure - Chin
PAI-36	ch_r-go_r $|$ ch_l-go_l	Labial Commissure - Gonion (Same side)
PAI-37	ch_r-go_l $|$ ch_l-go_r	Labial Commissure - Gonion (Different side)
PAI-38	ch_r-il_r $|$ ch_l-il_l	Labial Commissure - Lateral iris (Same side)
PAI-39	ch_r-il_l $|$ ch_l-il_r	Labial Commissure - Lateral iris (Different side)
PAI-40	ch_r-im_r $|$ ch_l-im_l	Labial Commissure - Medial iris (Same side)
PAI-41	ch_r-im_l $|$ ch_l-im_r	Labial Commissure - Medial iris (Different side)
PAI-42	ch_r-li $|$ ch_l-li	Labial Commissure - Lower lip
PAI-43	ch_r-ls $|$ ch_l-ls	Labial Commissure - Upper lip
PAI-44	ch_r-mid $|$ ch_l-mid	Labial Commissure - Midnasal
PAI-45	ch_r-n $|$ ch_l-n	Labial Commissure - Nasion
PAI-46	ch_r-pu_r $|$ ch_l-pu_l	Labial Commissure - Pupil (Same side)
PAI-47	ch_r-pu_l $|$ ch_l-pu_r	Labial Commissure - Pupil (Different side)
PAI-48	ch_r-sn $|$ ch_l-sn	Labial Commissure - Base of the nose
PAI-49	ch_r-sto $|$ ch_l-sto	Labial Commissure - Stomion
PAI-50	ch_r-zy_r $|$ ch_l-zy_l	Labial Commissure - Zygion (Same side)
PAI-51	ch_r-zy_l $|$ ch_l-zy_r	Labial Commissure - Zygion (Different side)
PAI-52	cph_r-cph_l $|$ cph_l-cph_r	Width of the crista philtri
PAI-53	cph_r-ec_r $|$ cph_l-ec_l	Crista philtri - Ectocanthion (Same side)
PAI-54	cph_r-ec_l $|$ cph_l-ec_r	Crista philtri - Ectocanthion (Different side)
PAI-55	cph_r-en_r $|$ cph_l-en_l	Crista philtri - Endocanthion (Same side)
PAI-56	cph_r-en_l $|$ cph_l-en_r	Crista philtri - Endocanthion (Different side)
PAI-57	cph_r-g $|$ cph_l-g	Crista philtri - Glabella
PAI-58	cph_r-gn $|$ cph_l-gn	Crista philtri - Chin
PAI-59	cph_r-go_r $|$ cph_l-go_l	Crista philtri - Gonion (Same side)
PAI-60	cph_r-go_l $|$ cph_l-go_r	Crista philtri - Gonion (Different side)
PAI-61	cph_r-il_r $|$ cph_l-il_l	Crista philtri - Lateral iris (Same side)
PAI-62	cph_r-il_l $|$ cph_l-il_r	Crista philtri - Lateral iris (Different side)
PAI-63	cph_r-im_r $|$ cph_l-im_l	Crista philtri - Medial iris (Same side)
PAI-64	cph_r-im_l $|$ cph_l-im_r	Crista philtri - Medial iris (Different side)
PAI-65	cph_r-li $|$ cph_l-li	Crista philtri - Lower lip
PAI-66	cph_r-ls $|$ cph_l-ls	Crista philtri - Upper lip
PAI-67	cph_r-mid $|$ cph_l-mid	Crista philtri - Midnasal
PAI-68	cph_r-n $|$ cph_l-n	Crista philtri - Nasion
PAI-69	cph_r-pu_r $|$ cph_l-pu_l	Crista philtri - Pupil (Same side)
PAI-70	cph_r-pu_l $|$ cph_l-pu_r	Crista philtri - Pupil (Different side)
PAI-71	cph_r-sn $|$ cph_l-sn	Crista philtri - Base of the nose
PAI-72	cph_r-sto $|$ cph_l-sto	Crista philtri - Stomion
PAI-73	cph_r-zy_r $|$ cph_l-zy_l	Crista philtri - Zygion (Same side)
PAI-74	cph_r-zy_l $|$ cph_l-zy_r	Crista philtri - Zygion (Different side)
PAI-75	ec_r-ec_l $|$ ec_l-ec_r	Width of the ectocanthion
PAI-76	ec_r-en_r $|$ ec_l-en_l	Eye width
PAI-77	ec_r-en_l $|$ ec_l-en_r	Ectocanthion - Endocanthion (Different side)
PAI-78	ec_r-g $|$ ec_l-g	Ectocanthion - Glabella
PAI-79	ec_r-gn $|$ ec_l-gn	Ectocanthion - Chin
PAI-80	ec_r-go_r $|$ ec_l-go_l	Ectocanthion - Gonion (Same side)
PAI-81	ec_r-go_l $|$ ec_l-go_r	Ectocanthion - Gonion (Different side)
PAI-82	ec_r-il_r $|$ ec_l-il_l	Ectocanthion - Lateral iris (Same side)
PAI-83	ec_r-il_l $|$ ec_l-il_r	Ectocanthion - Lateral iris (Different side)
PAI-84	ec_r-im_r $|$ ec_l-im_l	Ectocanthion - Medial iris (Same side)
PAI-85	ec_r-im_l $|$ ec_l-im_r	Ectocanthion - Medial iris (Different side)
PAI-86	ec_r-li $|$ ec_l-li	Ectocanthion - Lower lip
PAI-87	ec_r-ls $|$ ec_l-ls	Ectocanthion - Upper lip
PAI-88	ec_r-mid $|$ ec_l-mid	Ectocanthion - Midnasal
PAI-89	ec_r-n $|$ ec_l-n	Ectocanthion - Nasion
PAI-90	ec_r-pu_r $|$ ec_l-pu_l	Ectocanthion - Pupil (Same side)
PAI-91	ec_r-pu_l $|$ ec_l-pu_r	Ectocanthion - Pupil (Different side)
PAI-92	ec_r-sn $|$ ec_l-sn	Ectocanthion - Base of the nose
PAI-93	ec_r-sto $|$ ec_l-sto	Ectocanthion - Stomion
PAI-94	ec_r-zy_r $|$ ec_l-zy_l	Ectocanthion - Zygion (Same side)
PAI-95	ec_r-zy_l $|$ ec_l-zy_r	Ectocanthion - Zygion (Different side)
PAI-96	en_r-en_l $|$ en_l-en_r	Inter-canthion width
PAI-97	en_r-g $|$ en_l-g	Endocanthion - Glabella
PAI-98	en_r-gn $|$ en_l-gn	Endocanthion - Chin
PAI-99	en_r-go_r $|$ en_l-go_l	Endocanthion - Gonion (Same side)
PAI-100	en_r-go_l $|$ en_l-go_r	Endocanthion - Gonion (Different side)
PAI-101	en_r-il_r $|$ en_l-il_l	Endocanthion - Lateral iris (Same side)
PAI-102	en_r-il_l $|$ en_l-il_r	Endocanthion - Lateral iris (Different side)
PAI-103	en_r-im_r $|$ en_l-im_l	Endocanthion - Medial iris (Same side)
PAI-104	en_r-im_l $|$ en_l-im_r	Endocanthion - Medial iris (Different side)
PAI-105	en_r-li $|$ en_l-li	Endocanthion - Lower lip
PAI-106	en_r-ls $|$ en_l-ls	Endocanthion - Upper lip
PAI-107	en_r-mid $|$ en_l-mid	Endocanthion - Midnasal
PAI-108	en_r-n $|$ en_l-n	Endocanthion - Nasion
PAI-109	en_r-pu_r $|$ en_l-pu_l	Endocanthion - Pupil (Same side)
PAI-110	en_r-pu_l $|$ en_l-pu_r	Endocanthion - Pupil (Different side)
PAI-111	en_r-sn $|$ en_l-sn	Endocanthion - Base of the nose
PAI-112	en_r-sto $|$ en_l-sto	Endocanthion - Stomion
PAI-113	en_r-zy_r $|$ en_l-zy_l	Endocanthion - Zygion (Same side)
PAI-114	en_r-zy_l $|$ en_l-zy_r	Endocanthion - Zygion (Different side)
PAI-115	g-gn	Height of the face
PAI-116	g-go_r $|$ g-go_l	Glabella - Gonion (Same side)
PAI-117	g-il_r $|$ g-il_l	Glabella - Gonion (Different side)
PAI-118	g-im_r $|$ g-im_l	Glabella - Medial iris
PAI-119	g-li	Glabella - Upper lip
PAI-120	g-ls	Glabella - Lower lip
PAI-121	g-mid	Glabella Midnasal
PAI-122	g-n	Glabella - Nasion
PAI-123	g-pu_r $|$ g-pu_l	Glabella - Pupil
PAI-124	g-sn	Glabella - Base of the nose
PAI-125	g-sto	Glabella - Stomion
PAI-126	g-zy_r $|$ g-zy_l	Glabella - Zygion
PAI-127	gn-go_r $|$ gn-go_l	Chin - Gonion
PAI-128	gn-il_r $|$ gn-il_l	Chin - Lateral iris
PAI-129	gn-im_r $|$ gn-im_l	Chin - Medial iris
PAI-130	gn-li	Chin - Lower lip
PAI-131	gn-ls	Chin - Upper lip
PAI-132	gn-mid	Chin - Midnasal
PAI-133	gn-n	Chin - Nasion
PAI-134	gn-pu_r $|$ gn-pu_l	Chin - Pupil
PAI-135	gn-sn	Chin - Base of the nose
PAI-136	gn-sto	Chin - Stomion
PAI-137	gn-zy_r $|$ gn-zy_l	Chin - Zygion
PAI-138	go_r-go_l $|$ go_l-go_r	Inter-gonion width
PAI-139	go_r-il_r $|$ go_l-il_l	Gonion - Lateral iris (Same side)
PAI-140	go_r-il_l $|$ go_l-il_r	Gonion - Lateral iris (Different side)
PAI-141	go_r-im_r $|$ go_l-im_l	Gonion - Medial iris (Same side)
PAI-142	go_r-im_l $|$ go_l-im_r	Gonion - Medial iris (Different side)
PAI-143	go_r-li $|$ go_l-li	Gonion - Lower lip
PAI-144	go_r-ls $|$ go_l-ls	Gonion - Upper lip
PAI-145	go_r-mid $|$ go_l-mid	Gonion - Midnasal
PAI-146	go_r-n $|$ go_l-n	Gonion - Nasion
PAI-147	go_r-pu_r $|$ go_l-pu_l	Gonion - Pupil (Same side)
PAI-148	go_r-pu_l $|$ go_l-pu_r	Gonion - Pupil (Different side)
PAI-149	go_r-sn $|$ go_l-sn	Gonion - Base of the nose
PAI-150	go_r-sto $|$ go_l-sto	Gonion - Stomion
PAI-151	go_r-zy_r $|$ go_l-zy_l	Gonion - Zygion (Same side)
PAI-152	go_r-zy_l $|$ go_l-zy_r	Gonion - Zygion (Different side)
PAI-153	il_r-il_l $|$ il_l-il_r	Maximum iris width
PAI-154	il_r-im_r $|$ il_l-im_l	Diameter of the iris
PAI-155	il_r-im_l $|$ il_l-im_r	Lateral iris - Medial iris (Different side)
PAI-156	il_r-li $|$ il_l-li	Lateral iris - Upper lip
PAI-157	il_r-ls $|$ il_l-ls	Lateral iris - Upper lip
PAI-158	il_r-mid $|$ il_l-mid	Lateral iris - Midnasal
PAI-159	il_r-n $|$ il_l-n	Lateral iris - Nasion
PAI-160	il_r-pu_r $|$ il_l-pu_l	Lateral iris - Pupil (Same side)
PAI-161	il_r-pu_l $|$ il_l-pu_r	Lateral iris - Pupil (Different side)
PAI-162	il_r-sn $|$ il_l-sn	Lateral iris - Base of the nose
PAI-163	il_r-sto $|$ il_l-sto	Lateral iris - Stomion
PAI-164	il_r-zy_r $|$ il_l-zy_l	Lateral iris - Zygion (Same side)
PAI-165	il_r-zy_l $|$ il_l-zy_r	Lateral iris - Zygion (Different side)
PAI-166	im_r-im_l $|$ im_l-im_r	Minimum iris width
PAI-167	im_r-li $|$ im_l-li	Medial iris - Lower lip
PAI-168	im_r-ls $|$ im_l-ls	Medial iris - Upper lip
PAI-169	im_r-mid $|$ im_l-mid	Medial iris - Midnasal
PAI-170	im_r-n $|$ im_l-n	Medial iris - Nasion
PAI-171	im_r-pu_r $|$ im_l-pu_l	Medial iris - Pupil (Same side)
PAI-172	im_r-pu_l $|$ im_l-pu_r	Medial iris - Pupil (Different side)
PAI-173	im_r-sn $|$ im_l-sn	Medial iris - Base of the nose
PAI-174	im_r-sto $|$ im_l-sto	Medial iris - Stomion
PAI-175	im_r-zy_r $|$ im_l-zy_l	Medial iris - Zygion (Same side)
PAI-176	im_r-zy_l $|$ im_l-zy_r	Medial iris - Zygion (Different side)
PAI-177	li-ls	Lip height
PAI-178	li-mid	Lower lip - Midnasal
PAI-179	li-n	Lower lip - Nasion
PAI-180	li-pu_r $|$ li-pu_l	Lower lip - Pupil
PAI-181	li-sn	Lower lip - Base of the nose
PAI-182	li-sto	Lower lip - Stomion
PAI-183	li-zy_r $|$ li-zy_l	Lower lip - Zygion
PAI-184	ls-mid	Upper lip - Midnasal
PAI-185	ls-n	Upper lip - Nasion
PAI-186	ls-pu_r $|$ ls-pu_l	Upper lip - Pupil
PAI-187	ls-sn	Upper lip - Base of the nose
PAI-188	ls-sto	Upper lip - Stomion
PAI-189	ls-zy_r $|$ ls-zy_l	Upper lip - Zygion
PAI-190	mid-n	Midnasal - Nasion
PAI-191	mid-pu_r $|$ mid-pu_l	Midnasal - Pupil
PAI-192	mid-sn	Midnasal - Base of the nose
PAI-193	mid-sto	Midnasal - Stomion
PAI-194	mid-zy_r $|$ mid-zy_l	Midnasal - Zygion
PAI-195	n-pu_r $|$ n-pu_l	Nasion - Pupil
PAI-196	n-sn	Nose height
PAI-197	n-sto	Nasion - Stomion
PAI-198	n-zy_r $|$ n-zy_l	Nasion - Zygion
PAI-199	pu_r-pu_l $|$ pu_l-pu_r	Inter-pupil width
PAI-200	pu_r-sn $|$ pu_l-sn	Pupil - Base of the nose
PAI-201	pu_r-sto $|$ pu_l-sto	Pupil - Stomion
PAI-202	pu_r-zy_r $|$ pu_l-zy_l	Pupil - Zygion (Same side)
PAI-203	pu_r-zy_l $|$ pu_l-zy_r	Pupil - Zygion (Different side)
PAI-204	sn-sto	Base of the nose - Stomion
PAI-205	sn-zy_r $|$ sn-zy_l	Base of the nose - Zygion
PAI-206	sto-zy_r $|$ sto-zy_l	Stomion - Zygion
PAI-207	zy_r-zy_l $|$ zy_l-zy_r	Face width

Equations

$$\text{iris\_ratio}=\frac{d(i{m}_L,i{l}_L)+d(i{m}_R,i{l}_R)}{2}$$

$${\text{F}}_1=2\ast \left(\frac{precision\ast recall}{precision-recall}\right)$$

Full text

I Introduction

Anthropological knowledge can be used to support forensic investigations of the deceased and the living [1]. In the first case, postmortem (PM) profiles of victims are reconstructed to narrow the number of comparisons between missing persons and unknown bodies, as described in [2]. The profiling process is carried out by retrieving information regarding sex, age, stature and ancestry from the deceased, especially from skeletal remains, and comparing them with antemortem (AM) data from the alleged victim [3]. In order to increase reliability, the collected information is combined with AM and PM evidence obtained by primary means of human identification, namely fingerprint, dental and DNA analyses [4].

On the other hand, forensic anthropology applied to the living usually relies on morphological and biometric information of victims and suspects of crimes registered in footage of closed-circuit television and photographs [5, 6]. The identification of children that suffered sexual exploitation, as well as their perpetrators, figures among the procedures requested by Law in the routines of medico-legal institutes [7]. Over the last decade, requests for anthropological examination of the living became more common following an increasing trend in cybernetic crimes [8]. This new outlook justified the need for developing advanced tools to support forensic casework [9].

The photo-anthropometric analysis of the human face emerges in this context as an alternative tool for searching, collecting, and quantifying morphological features and using them for forensic purposes [10]. Working at the interface of forensic anthropology and computer science, this non-invasive and low-cost approach is founded on the registration of landmarks on photographs and the calculation of ratios between facial distances [7, 9, 10]. The morphometric information retrieved from the human face can be used within a comparative basis, between reference and target persons, or in a reconstructive basis, where sexual dimorphism and age estimation of the living are performed [11].

This study was designed with four aims for the use of photo-anthropometric data of the human face: to propose an automatic solution based on an artificial neural network to estimate anthropological information using photo-anthropometric indices (I); to test the diagnostic accuracy of the solution with cut-off points between male and females and threshold limits for the ages of 14 - related to sexual consent [12, 13], and 18 - related to legal majority [14, 15] (II); to analyze the correlation between sex and age using photo-anthropometric indices of the human face (III) and; provision, for academic purposes only, of a complete data set of facial photo-anthropometric points marked and checked by forensic experts, measured on over 18.000 faces of individuals from Brazil over the last 4 years (IV).

The manuscript is organized as follows: Section II presents the proposed method and all the processing details for classification using a model based on artificial intelligence techniques. It also includes a description of the database and the inputs used in this manuscript. Section III presents the experimental results. Finally, Section IV presents the discussion, main conclusions and future works.

II Materials and Methods

The main goal of this manuscript is to evaluate the use of photo-anthropometric data from human faces to create an automatic classifier based on an artificial neural network to estimate age and sex. In this Section we will describe the approach used to create an automatic solution using an artificial intelligence model and the details of the inputs and the proposed tests.

II-A Photos Database and Cephalometric Landmarks sets

The principal proposed and used database is composed of photo-anthropometric index data from 18,000 frontal face photos of 18,000 different people from Brazil. All photos were acquired in accordance with ICAO 9303 normative [16] used for Machine Readable Travel Documents. All photos have been captured over a white background, stored at a resolution of 480x640 pixels, 24bits of color, with no glasses and with a natural expression. The photos are divided into male (9,000) and female (9,000) and 18 age groups (from 5 up to 22 years old), totaling 500 photos for each individual group (sex and age).

The cephalometric landmarks adopted in this manuscript, all of which were used to create the photo-anthropometric indexes, were described in [9] and [7]. In this methodology, one expert manually located 28 cephalometric landmarks in each of a 1000 pre-training images using a SAFF-2D software, applying an identification methodology proposed in [10, 17, 18]. All 18,000 faces were processed and all facial landmarks were marked and checked following the methodology developed in [19]. The following is a complete list of cephalometric landmarks which were identified, according to Caple and Stephan standard nomenclature [20]: Entocanthion (en’), Exocanthion (ex’), Iridion laterale (il), Iridion mediale (im), Pupil (pu’), Zygion (zy’), Alare (al’), Gonion (go’) and Cheilion (ch’), Crista philtri (cph’) bilateral landmarks. The remaining can be found on face’s midline as: Glabella (g’), Nasion (n’), Subnasale (sn’), Labiale superius (ls’), Stomion (sto’), Labiale inferius (li’), Gnathion (gn’), Midnasal (m’) [21]. Figure 1 shows all the landmarks used and their locations on the face.

II-B Photo-anthropometric indexes

Using the cephalometric landmarks detailed in Subsection II-A, the authors of [9] and [7] present details about 10 and 40 facial photo-anthropometric indexes (PAIs) respectively. They are facial measures which permit the explanation of people’s anthropometric information, specifically age and sex, by only using facial growth information.

The main feature of the PAIs is the use of the mean iris diameter as a proportional factor defined as the iris ratio to solve scale and calibration issues (e.g. different metrics relations to the pixel size and to the intrinsic parameters of the capture device which was used). The iris ratio is essentially the Euclidean distance defined by a function $d(p,q)$ from Iridion laterale (il) and Iridion mediale (im) of both eyes. The iris ratio is detailed in Equation 1. When a facial measurement generated by the Euclidean distance between two cephalometric landmarks is divided by the iris ratio the result is a stable growing factor which solves the image scale problem and removes effects of varying landmark positions over the face.

[TABLE]

Using a specialized software to automatically extract the cephalometric landmarks, we applied the methodology (iris ratio) proposed in [9], combining all of the 28 cephalometric landmarks creating 208 PAIs. The result is stored in a comma-separated values file (CSV) composed of the photo-anthropometry data, including 208 PAIs per image, labeled by sex and age for the 18.000 images used to build the data set. More details about the generated data set will be described in Section II-D. The complete description of all 208 PAIs can be found in the supplementary material in Section B.

II-C Experimental set-up and evaluation metrics

All the proposed tests were executed on an Intel AI DevCloud [23], a cloud framework for applications in artificial intelligence powered by Intel, running the Keras API version 2.2.0 [24] and Tensorflow version 1.8 [25]. During the training process we adopted an artificial neural network (ANN) model based on a multilayer perceptron (MLP) [26], using one dense (fully connected) layer with 128 neurons with the Adamax optimizer from Adam [27]. Table I presents the detailed MLP structure used to execute the tests.

In a classification process, all results require a test accuracy analysis and a traditional measure is the F1 score (also known as F-score or F-measure) [28]. This accuracy metric is composed of four parameters: True Positives ($n_{tp}$), True Negatives ($n_{tn}$), False Positives ($n_{fp}$) and False Negatives ($n_{fn}$). At the end of the estimation process, we evaluated the estimate as a true positive ($n_{tp}$) when the classifier agreed with the anthropological data (age/sex) for the validation data and as $n_{fp}$ (false positive) when the estimate (age/sex) disagreed. We have used the F1 score method shown in Equation 2,

[TABLE]

where the variables $precision$ and $recall$ are defined as $precision=n_{tp}/(n_{tp}+n_{fp})$ and $recall=n_{tp}/(n_{tp}+n_{fn})$, respectively. The F1 score metric varies between 1 (best) and 0 (worst). To evaluate each test in detail, we adopted the confusion matrix method [29]. The confusion matrix allows the comparison of each “predicted label” with the “true label” resulting in a table which presents the classifier’s behavior, providing a more complete visualization than the F1 score.

As described above, we propose the use of 208 PAIs per photo and provide them as input to an automatic classifier to estimate anthropological information. In order to evaluate the expected performance of the classifier model over the proposed data set, a k-fold cross-validation was used in the testing process [30, 31]. In our case, a ten-fold cross-validation procedure was used, randomly separating our data set five times with $90\%$ of photos for training the classifiers and $10\%$ of photos for testing.

To evaluate the PAI methodology for anthropological estimation, we defined 3 groups of tests. The first one, Group A relates to the estimation of sex, the second one, Group B, to the estimation of age and Group C to the estimation of the age group. These tests are described as follows:

• •

Group A - Sex estimation: The group is composed of two test sets where one classifier was trained with age as an input and the other classifier avoiding111When the term “avoiding” is used throughout the manuscript, it means information from a specific class was not used in the proposed tests. it, working similarly to an adversarial neural network [32].

• •

Group B - Age estimation: This group is composed of 4 tests to evaluate age estimation using only the PAI data. For all these tests, we focus on identifying if the sex data can provide improvements over (or interfere with) the age estimation using only the facial measurements.

• •

Group C - Group age estimation: In this group two classifiers are proposed in order to evaluate if the facial measurements are able to identify if an individual belongs to a specific age group. The classifiers analyze the thresholds of 14 and 18 years old. The first one identifies if the person is 14 years old or more and the second one if the person is 18 years old or older.

As stated above, in Group A the estimation of sex is analyzed. Table II presents the structure of the tests for Group A. In Test 1, we trained the classifier separately for each age group, totaling 17 individual tests from 5 up to 22 years old. This test was developed to try to evaluate the classification process applied on each age range using only the PAI data. In Test 2 we used all the data observations (avoiding the age information) to evaluate the sex classification over all age groups.

Meanwhile in Group B the age estimation process is analyzed. Table III presents the test structure of the Group B tests. We defined 4 tests to evaluate the classification process for age estimation using the PAI data. For each test we analyzed the impact of sex information on the estimation process.

For each proposed test, we trained three different classifiers. Two by splitting the PAI data by sex (female and male) and the other classifier using all the data observations in the same classification process. The 4 tests in Group B are described as follow:

• •

Test 1: We selected the PAI data from 6 to 22 years old at age intervals of 1 year. In this test we are focused on evaluating the classifier accuracy for age estimation with a total of 17 age classes as output.

• •

Test 2: We selected the PAI data from 6 to 22 years old at age intervals of 2 years. In this test, as in Test 1, we evaluate the classifier accuracy for age estimation with 9 age classes as output: 6, 8, 10, 12, 14, 16, 18, 20 and 22.

• •

Test 3: We selected the PAI data from 6 to 22 years old at age intervals of 4 years. As above, we are evaluating the classifier accuracy for age estimation with a total of 5 age class outputs: 6, 10, 18 and 22.

• •

Test 4: We selected the PAI data from 5 to 20 years old at age intervals of 5 years. As in all the previous tests of this group, we are interested in evaluating the classifier accuracy for age estimation with a total of 4 output classes: 5, 10, 15 and 20.

Finally, for Group C, Table IV presents the test structure for age group classification, with thresholds of 14 and 18 years. In these tests we evaluated if the PAIs can be used to estimate if the person is older/under 14 years old or older/under 18 years old. We adopted the same test procedures as presented in Group B. In this case we evaluated if the sex data affects the age group estimation process. For both tests we trained classifiers for females, for males and for all observations without the sex information as input.

II-D Descriptive Statistical Analyses of the Data Set

In this section we undertake a descriptive statistical exploration of the new data set which serves as input to the proposed machine learning model222The data set and the developed machine learning model will be available to download soon, after the blind review process is completed, in attendance of the journal’s submission criteria.. As described above, the full data set contains 500 observations on 208 variables for each category of Age (18 levels) and Sex (2 levels), yielding $18,000$ observations in total. Given the large number of variables and categories, it is not a viable option to describe all the variable distributions in all the categories. Therefore, we performed a visual and detailed analysis of boxplots of the distributions of all the variables in all the 36 categories. An example for the variable PAI-10 (wing of the nose - chin) can be found in Figure 2. Males present larger values than females and the differences increase with age. Similar figures for all 208 variables can be found in the supplementary material333All boxplots are available in the specific section of the Supplementary Material files (File PAIs_sex_age_boxplots_supplemental.pdf)..

We performed Shapiro-Wilk Normality Tests [33] in each age-sex group for each variable. The null hypothesis for this test is that the data are normally distributed. $P$ values lower than $0.01$ were considered indications of significant deviations from normality. Three variables were non-normal in all age-sex groups: PAI-154 (diameter of the iris), PAI-160 (lateral iris - pupil), and PAI-171 (medial iris - pupil, same side). For the other variables, approximately $12\%$ of the tests rejected normality (about two age-sex groups per variable). However, given the large number of observations in each category and the presence of many outliers444See Figure 2, for example, where outliers can be identified as the points well above or well below the boxes in the boxplot., both of which increase the probability of rejecting normality, this suggests that much of the data is normally distributed and that analysis techniques that depend on normality can be used, albeit with caution.

Two-way analyses of variance (ANOVA) [34] were performed on each PAI to assess statistically significant changes in facial parameters in response to the factors sex, age, and the interaction between sex and age. $P$ values lower than $0.01$ were considered significant. For each of the 208 PAIs both sex and age were significant factors. The interaction between sex and age was not significant for only 4 of the 208 PAIs, specifically PAI-50 (labial commissure - zygion, same side), PAI-154 (diameter of the iris), PAI-160 (lateral iris - pupil), and PAI-171 (medial iris - pupil, same side). Three of these four variables were identified as non-normal in all categories by the Shapiro-Wilk tests presented above, only PAI-50 was not. As an example, the distributions of PAI-160 are presented in Figure 3. It is interesting to observe that the inclusion of these variables in the proposed machine learning model does not introduce any difficulties, as would occur in a standard regression model where large correlations and/or lack of variation would cause instability in the model estimates. The model simply will not use variables which don’t provide additional predictive power.

III Results

This section presents all the results obtained using the proposed artificial neural network architecture detailed in Subsection II-C, including all the F1 scores and confusion matrices for each proposed test.

III-A F1 results

For Group A, the evaluated tests describe how the PAI indexes can classify the sex information. Figure 4 presents the F1 score for sex estimation for each age. The last column (all) was obtained using all the ages in the classification process in order to evaluate the sex classification baseline test or “global” test. We can compare the mean of the results for each age to the “global” test, obtaining similar F1 scores: $0.83$ and $0.81$ respectively.

The tests from Group B use the PAI indexes for age classification. Figure 5 presents the F1 scores for age estimation divided in four groups: age intervals of 1 year (green), age intervals of 2 years (blue), age intervals of 4 years (red) and age intervals of 5 years (purple). Each group test was divided into three subtests: the first one is the age estimation on male individuals. The second one just female individuals and the last one using both and avoiding the sex information.

The Group C tests are focused on evaluating if the PAI indexes can classify the age group information correctly. Figure 6 presents the F1 scores for age group estimation divided into two tests: older/under 14 years old and older/under 18 years old. Each group test was divided into three subtests, as were the Group B tests: male individuals; female individuals; and all individuals avoiding the sex information.

III-B Confusion Matrices

For each group test, we used confusion matrices to provide a detailed report, including all achieved results for age and sex classification. The matrices demonstrate the classifiers’ accuracy by comparing the “predicted labels” to the “true labels”.

Figure 7 presents the confusion matrices for the Group A tests. The figures from Figure 7 LABEL:sub@fig:mf_age_1 to Figure 7 LABEL:sub@fig:mf_age_22 present the confusion matrices for each age group separately, from 5 to 22 years old respectively. Figure 7 LABEL:sub@fig:mf_age_avg presents the average results of all age groups, meanwhile the Figure 7 LABEL:sub@fig:mf_age_all presents the results using all data avoiding the age information.

Figures 8, 9 and 10 present the confusion matrices for tests in Group B using 1 year for age intervals. Figure 8 presents the results for age estimation avoiding the sex information. Figure 9 presents the results for age estimation for females. Figure 10 presents the results for age estimation for males.

Figures 13, 14 and 15 present the confusion matrices with the detailed results for Group B: age intervals of 2 years. Figure 13 presents the detailed results for age estimation avoiding the sex information. Figure 14 and Figure 15 present the detailed results for age estimation for female and male individuals, respectively.

Figure 16 presents the confusion matrices with the detailed results for Group B: age intervals of 4 years. Figure 16 LABEL:sub@fig:age_4_4 presents the detailed results for age estimation avoiding the sex information. Figure 16 LABEL:sub@fig:age_4_4_F presents the detailed results for age estimation on female individuals. Figure 16 LABEL:sub@fig:age_4_4_M presents the detailed results for age estimation on male individuals.

Figure 17 presents the confusion matrices with the detailed results for Group B: age intervals of 5 years. Figure 17 LABEL:sub@fig:age_5_5 presents the detailed results for age estimation avoiding the sex information. Figure 17 LABEL:sub@fig:age_5_5_F presents the detailed results for age estimation on female individuals. Figure 17 LABEL:sub@fig:age_5_5_M presents the detailed results for age estimation on male individuals.

Figure 11 presents the confusion matrices for Group C: older/under 14 years old. Figure 11 LABEL:sub@fig:age_group_14 presents the detailed results for age group estimation avoiding the sex information. Figure 11 LABEL:sub@fig:age_group_14_F presents the detailed results for age group estimation on female individuals. Figure 11 LABEL:sub@fig:age_group_14_M presents the detailed results for age group estimation on male individuals.

Figure 12 presents the confusion matrices with the detailed results for Group C: older/under 18 years old. Figure 12 LABEL:sub@fig:age_group_18 presents the detailed results for age group estimation avoiding the sex information. Figure 12 LABEL:sub@fig:age_group_18_F presents the detailed results for age group estimation on female individuals. Figure 12 LABEL:sub@fig:age_group_18_M presents the detailed results for age group estimation on male individuals.

IV Discussion

Estimating the age of victims and suspects of crimes is an essential procedure in the forensic casuistic of human identification [35, 36, 37, 38]. Age estimation becomes even more important when legal age thresholds are determined by the Court [39, 40]. Currently, the ages of 14 and 18 years represent the legal age threshold of sexual consent and majority in the judicial system of several countries worldwide [41, 42]. Designing scientific tools that allow the investigation of age with accuracy and evidence-based standards must be continuously encouraged to promote optimal forensic practices. Justified by the need for improving facial age estimation through photo-anthropometric analysis and founded on the hypothesis that the photo-anthropometric analyses of the human face can distinguish individuals younger and older than 14 and 18 years, this study aimed to propose and test an automatic solution to distinguish (male and female) individuals younger and older than 14 and 18 using photo-anthropometric analyses of the human face.

Population-specific studies on the use of photo-anthropometry of the face for age estimation are scarce in the scientific literature – especially those with large and standardized samples [9, 10, 11]. In this study, a large sample of Brazilian participants was collected (n=18000) and organized in a detailed, standardized data set available only for academic purposes. Sample standardization was accomplished by selecting male (n=500) and female (n=500) participants equally distributed (n=1000) in age intervals of one year (from 5 to 22 years). Moreover, all the digital photographs of the participants were taken with the same equipment and followed protocols previously reported in the scientific literature [9] and by the International Civil Aviation Organization (ICAO). In the scope of sampling strategy, an additional gain was obtained by choosing Brazilian participants, which are known for their multiracial formation [43] and phenotypic diversity [44]. This characteristic makes the study outcomes novel and eventually more reproducible in other populations.

In relation to the study outcomes, inferences about age were first investigated in association with sex. More specifically, the morphological information retrieved from the human face was tested based on its performance to classify males and females in age categories (5-22). A separate statistical test was evaluated for each of the 18 age groups. The outcomes endorsed the scientific literature by revealing difficulties for sexual dimorphism in young participants ($<$12 years). In particular, the mean F1 scores in young age groups ranged between $0.7$ and $0.78$. On the other hand, for participants aged from 13 to 22 years the mean F1 scores increased from $0.81$ to $0.91$, showing evident improvement for classifying males and females. According to the scientific literature, the difference between young and old participants is explained by the lack of mature secondary sexual features depicted in facial photographs of children [45].

Recently, Kloess in [45] investigated the challenges in classifying child sexual abuse images. In their study, inferences about age (minor or not) were more easily and accurately given in images of babies and toddlers, while difficulties increased when the age of interest approached young adolescence. Interestingly, the larger size of the eyes in comparison with the other facial features and the presence of (milk) teeth and interdental gaps emerged as potential indicators of youth, while the use of make-up was a confounding factor [45]. In the present study, the diameter of the iris (iris ratio) [9] was used as fixed reference to calculate morphological ratios from the human face – which enables a quantification of the qualitative information provided by the previous authors. Additionally, photographs of participants depicting unnatural facial expressions (e.g. smiling) or using make-up were part of the exclusion criteria in the present study. Corroborated by the scientific literature [45], this methodological set up promoted a reduction of age classification bias as function of sex. In this context, sexual dimorphism, regardless of age, was performed as a quality-control procedure to eliminate the influence of age over the classification performance. The mean F1 score reached $0.81$, indicating that the classification tool was able to properly distinguish most of the males and females of the sample if they were combined in a single group. In practice, the use of age-specific classification tools is recommended (whenever applicable) to best-fit the needs of each case – especially if the case involves age interests between 13 and 22 years, such as the age of sexual consent [42] and the age of legal majority [41].

In a second phase, the present study engaged in a deeper investigation based on age. This phase was justified by the uncertainty regarding the chronological age of victims and suspects of crimes that are commonly observed during the routine of forensic services. The methodological set up at this phase clustered together not only participants in age intervals of one year, but also in larger age intervals (e.g. within two, four and five years). Within each group, the performance was better for classifying the age of males. Between groups, the mean F1 scores were progressively higher with the increase in age interval size. Consequently, the best age estimates were found in the group with age interval of five years (mean F1 score: 0.74 combining males and females). Clearly, this outcome shows that the classification process becomes more difficult by refining the sample based on age. A similar approach was recently used by Machado in [9]. The authors performed a photo-anthropometric analysis to investigate the allometric growth of the human face by grouping together individuals within age intervals of four years. Despite the evident contributions towards the analysis of facial alterations over the time, the methodological set up proposed by the authors was limited compared to the present study. The advantages highlighted in the present setup include not only the sample stratification in groups of four different age intervals, but also the collection and quantification of much more morphological information from the human face. While the authors in [9] mapped human facial growth with ten measurements calculated from ratios based on the diameter of the iris, the present study outcomes were founded on 208 measures calculated with the same rationale. In practice, the improved methodological setup proposed in the present study induces more reliable and accurate age estimations.

The third and final phase addressed in this study focused on testing the classification system to distinguish participants that were younger or older than specific legal age thresholds of interest. This set up was justified to specifically meet the needs of justice when it comes to answer legal requests regarding the ages of 14 – related to sexual consent, and 18 – related to legal majority. The tests that were carried out in this phase showed, again, better classification of males, both in relation to the threshold of 14 and 18 years. When males and females were combined, the mean F1 scores reached 0.93 and 0.85 for the ages of 14 and 18 years, respectively. Satisfactory outcomes were also recently observed by Borges in [7] with a similar approach. The authors obtained accuracy (Area Under the Curve) estimates from Receiver Operating Characteristic (ROC) curve analyses that reached 0.96 and 0.90 for the ages of 14 and 18, respectively. Differences between studies include the larger sample size in the current investigation (n=18000 in face of 1000 used by the authors [7]) and, again, the larger number of measurements from the human face (n=208 compared to the 40 measurements used by the authors [7]). In forensic practice, the performance of the classification systems used in this study strengthens and supports its use for distinguishing victims and suspects of crimes aged below or above 14 and 18.

The automatic solution developed to classify individuals based on age and sex using morphological information retrieved from photo-anthropometric analyses of the human face reached optimal outcomes. More accurate age estimates were found in subjects aged between 13 and 22 years; estimates for the age of males were better than for females; and classifications based on legal age thresholds of sexual consent and majority were feasible and promising.

The proposed methodological setup presents multiple advantages compared to the available scientific literature. However, translating it to practice requires careful implementations and follow-up work to include updates in scientific evidence. Future studies in the field should test the reproducibility of this methodological setup and inherent outcomes in different populations. Investigations based on other legal age thresholds of interest are also encouraged to best-fit the judicial systems of different countries. Advances in the methodological setup for further improvements should include longitudinal sampling and three-dimensional imaging. Another approach in the computer vision field is to evaluate a classifier using deep learning techniques for age and sex estimation [46, 47], combining photos with photo-anthropometric indexes creating a cross-domain classifier. This allows the evaluation of possible improvements achieved by the inclusion of PAIs when compared to using only images as input.

Acknowledgment

The authors would like to acknowledge the team of Federal Police of Brazil, specially the forensic experts of National Institute of Criminalistic. This work was conducted with financial support from Coordination for the Improvement of Higher Education Personnel (CAPES) and Federal Police of Brazil (Grant Number: 001 Pro-Forenses 25/2014 CAPES).

Appendix A Confusion Matrices

Appendix B Photo-anthropometric Indexes