«Age Estimation of Adolescents and Adults Using the Dimensions of the Eye and Pupil in “Selfie” Photographs KariAnna Baber, BA, BS Research ...»
Age Estimation of Adolescents and Adults
Using the Dimensions of the Eye and Pupil
in “Selfie” Photographs
KariAnna Baber, BA, BS
Research Advisors: Dr. Terry Fenger (MUFSC), Joshua Brunty (MUIS&T),
Cpl. Robert Boggs (WVSP), Ian Levstein (MUFSC)
Summer 2014 Internship with Marshall University Forensic Science Center
West Virginia State Police Digital Forensics Unit
The goal of this project is to demonstrate the effectiveness of investigating digital
images and correlating an age with the photographed individual. This becomes most relevant for individuals in the teenage age range, who often appear older in age due to the use of makeup, posing, and filter technologies commonly used to take a picture of oneself -- otherwise known as a “selfie.” By analyzing features of the face, particularly the eye and pupil regions, the subject has less ability to hinder age estimation based solely on physical appearance. Institution Review Board (IRB) approval was obtained in order to use human subjects. The target age group of participants was between 11-19 years old, however participants not within the age range were accepted. Because there are many variables that influence the functionality of the pupils, such as mood, eye problems, medications, and lighting, images were taken under controlled conditions which include using the same room and lighting. Each subject was given an ID number for the project, and asked a simple list of questions detailing his/her mood, medications, and eye problem history as well as his/her age, birthday and other demographic information. A series of pictures were taken of the individual with a Nikon® D3100 digital camera and Apple® iPad® iOS Version 7.1.1, along with a short video of around the participant’s face. The images were downloaded onto a computer for analysis using Photoshop®. Each image was calibrated so that the pupillary diameter, area, and interpupillary distance could be determined and compared using formulas given in MacLachlan & Howland, 2002. Using the known age of each participant and the age from the formulas, the effectiveness of age estimation can be determined.
2|Page Literature Research In the forensic analysis of digital images and age, there are two possible avenues that can potentially determine a correlation. First, there is age synthesis, which involves the process of changing the appearance of the face to mimic the natural aging process (Fu, Guo, Huang, 2010). An example of this process is using a child image and synthesizing the face to appear older. This can be seen on missing children posters. Second, there is age estimation, which involves assigning an exact age or age group to the subject featured in an image. This is a soft biometric technique in that it does not aim to identify an individual, but provides supportive, descriptive information about an individual (Fu, Guo, Huang, 2010). While the perceived age and appearance age can be easily altered, it would be beneficial to be able to determine an estimated age that would place a subject in a minor or non-minor grouping. This is the target process that has potential to be applicable to forensic digital examiners in crimes involving minors.
Age estimation has been found to be most difficult with younger subjects for a variety of reasons (Zeng, et al., 2012). First of all, the image factors such as resolution, quality, and lighting increase the difficulty in an age estimation exercise. A possible way to circumvent these issues is converting the image into gray scale. Second, images of younger individuals are more difficult to assign a perceived age, especially when factoring ethnicity and gender features.
Conversely, children were easier to age estimate than adults because of the growth stages of different features in the face.
Certain features of the face have been used for soft biometric age estimations. The ears, nose, mouth width, and facial form are some physical features that continue to change with 3|Page age. In a study done by Guyomarc’h & Stephen in 2012, features of the ears were measured and it was determined that this method and feature was minimally accurate in for several reasons. Most strikingly is the subjectivity involved in finding the significant feature positions for measurement and the lack of reference data. Another study conducted by El Dib & Onsi (2011) used multiple eye wrinkle and forehead feature points for age estimation. The process involved cropping an image in different shapes depending on the number of feature points and active shape models, Gabor functions, support vector machines (SVM), and support vector regression (SVR). The mean absolute error (MAE) is around 3 years using these algorithms.
However, this study is not representative of minors because of the databases used for analysis (El Dib & Onsi, 2011).
Other computer generated algorithms, along with SVM and SVR, include partial least squares methods (PLM) (Guo & Mu, 2011). This method produced a MAE around 4.5 years, while a linear SVM produced an MAE of 5 years when used on a particular database.
Furthermore, using another database, the MAE produced from different algorithms where within 0.45 years of each other. More databases and algorithms along with the use of biologically inspired features have been examined which reduce the MAE to as low as 2.61 years for females and 2.58 for males in different combinations (Guo et al., 2009).
In the field of ophthalmology and optometry, studies and work have been done focusing on using the pupil of an eye. Two important studies in particular, have found a relationship between the pupil measurements and age. In 2009, Lavezzo, Schellini, Padovani, and Hirai conducted a study using preschool aged children and focused on the differences in the pupil with gaze types. An attentive gaze is considered initial and focused, while a spontaneous gaze is 4|Page considered comfortable and exploring the environment. After taking a digital image of the child and controlling the illumination, the image was inverted to take measurement. It was found that the diameter of the pupil differs in spontaneous and attentive gazes, but that the right pupil and the left pupil could be considered equal and within the error range of the mean.
Another study included the tracking of children from 1 month old to 19 years of age (MacLachlam & Howland, 2002). Each year, the individual’s eyes were photographed using a fiber optic light guide in a camera lens and flash gun at two different illumination settings. One image at ambient light, 300 lux, and another in dimmed lightening, 15.9 lux. The images were analyzed by measuring pupillary diameter, area, and interpupillary distance. It was found that all three of these measurements have the potential to correspond with age, however the results of children in the 12-19 year old range has a decreased amount of data due to subjects leaving the study. From this study, equations were that correlate each measurement with an age as well as consider gender and the change of illumination in different photographs.
Furthermore, an additional measurement could be potentially useful in the field of forensics.
Considering the ratio of the cornea to the pupil diameter, may allow for a decreased MAE in age estimation in forensics, although it is used in ophthalmology for surgical planning (Cakmak, et al., 2012).
Considering the pupil for soft biometrics may be new, but using another feature of the eye is known. Hard biometric techniques, which aim to identify an individual, have aimed to use the iris as a means of identification and security (Fu, Guo, Huang, 2010). The iris is the colored part of the eye, which is converted from a round, donut image to a rectangle (Poonguzhali & Ezhilarasan, 2012). The features within the iris are transformed into a pattern using 5|Page normalization, sharpening, and Gabor functions. This final information image is then read in a way that matches the pattern to a known database image with corresponding information, similar to a bar code, in order to identify.
Using the iris biometerics is not without its problems, particularly when considering how the pupil works (Hollingsworth et al., 2009). Pupil dilation is affected by many factors such as drug usage, mood, light exposure, and health problems such as cataracts (Lavezzo et al., 2009).
When the pupil increases, less of the iris is exposed and when the pupil in constricted, more of the iris is exposed. One of the easiest variables to control is lighting. It then becomes important to normalize the iris pattern when the pupils are at the most extreme degrees of dilation.
Complicating the procedure more, the resolution of the image or scanner needs to be able to detect pattern well enough to procure a match from the database (Hollingsworth et al., 2009).
Introduction Today, in these modern times, the dependence people have on digital devices has increased and is still increasing. From digital cameras to smart phones to tablets, there is a constant opportunity for nearly anyone and everyone to be connected to the internet in order to find information, store information, post about daily routines to social media, and purchase merchandise, not to mention the ever growing downloadable apps that are featured to assist a user in a particular task. Keeping in contact across long distances is easier, deals are easier to find, and massive amounts of information can be obtained and stored in reachable locations and are available anytime, anywhere.
While this makes many daily activities and communication simpler, a new wave of criminal behavior has emerged with unique digital evidence to be analyzed. In particular, digital 6|Page images have the potential to be found anywhere from the data in suspect cell phones to personal computer hard drives. These images can be stored and shared with others on the cloud, through social media sites, e-mail, applications, and multimedia messages (MMS).
Increasing in popularity is the trend of taking selfies. A selfie is defined by the Oxford dictionary as “a photograph that one has taken of oneself, typically one taken with a smartphone or webcam and shared via social media.” Taking a selfie can occur anywhere, anytime, and shared with others, keeping a constant, open line of communication. People take selfies when they are mad, sad, happy, doing something crazy, wearing something stylish, or even to show boredom.
This trend has even grown to include taking large group selfies known “usies.” One of the more recent and well known examples of this was the picture taken by Ellen DeGeneres during the 2014 Academy Awards. In fact, this usie image was so popular that it crashed the social media site Twitter.com. A feature common selfies which cannot be consciously controlled is the pupil.
Even with the most magnificent make-up job or silliest facial expression, the pupil is going to respond in an unconscious way. Pupil size can be altered by numerous factors such as drug usage, mood, light exposure, and health problems such as cataracts (Cakmak, et al., 2012).
However, these images may not be as ordinary as a picture of one’s dog or from the night of a best friend’s wedding. Digital images have become prevalent in criminal behaviors such as the distribution and possession of child pornography, sexting, stalking, harassment, and prostitution and solicitation. Many of the crimes listed are of concern because of the involvement of minors as well as the frequency at which these types of cases are being seen by forensic digital analysts. While it may be common to utilize the terms “child” and “minor” interchangeably in lay conversation, legally they can have different implications depending on 7|Page state statues. The terms generally differ in the age associated with maturity. According to one legal dictionary, a child is anyone under the age of 14 while a minor is under the age of 18 (http://legal-dictionary.thefreedictionary.com/child). In the state of West Virginia, however, a child is anyone under the age of 18 years old, which is the same as the definition of a minor (http://www.legis.state.wv.us/wvcode). This becomes an important fact to consider when determining what kind of crime was committed, if one was at all.
Typically, when thinking of the crimes listed previously and thinking of children, the image of a kindergartener, someone around 5 years old, comes to mind. But, there is an age group of children that can frequently appear much older than they actually are adolescents.
With the use of make-up, posing, lighting, outfits, etc. adolescents can alter their perceived and appearance age while masking their actual age in images taken on digital devices. Then, these pictures may end up on social media sites, be sent as an MMS to a friend, or used in matters of child exploitation. In evidentiary images, making a determination which describes the subject as a minor or minor is difficult and problematic based on physical appearance. So, the question becomes whether age can be determined from a digital image using pupil and eye measurements that cannot be consciously controlled and is there an appropriate methodology to determine age from a digital image?
Institutional Review Board (IRB) Approval Although previous studies have used existing databases to retrieve and analyze for age estimation purposes, it was believed that having the ability to control as much of the environment in which the pictures would be taken would be most beneficial. This entails using actual adolescents as participants to photograph. Because children are a protected population 8|Page and the personal, face-to-face interaction became necessary, IRB approval needed to be granted.
In a meeting with Bruce Day at the Office of Research Integrity at Marshall University, it was required to apply for an expedited review of a social research project. Social research projects differ from medical research projects in that there is a lack of medicine trials, and typically the risk-benefit components of participation are minimal. An expedited review was chosen because of the absence of risk and benefit to participants. In other words, by participating in this project, a participant would not be at any risk or gain any prize or reward.