The “facial composite” is one of the major fields in the forensic science that helps the criminal investigators to carry out their investigation process. The survey conducted by United States Law Enforcement Agencies confirms that 80% of the law enforcement agencies use computer automated composite systems whereas Sri Lanka is still far behind in the process of facial composite with lot of inefficiencies in the current manual process. Hence this research introduces a novel approach for the manual facial composite process, while eliminating the inefficiencies of the manual procedure in Sri Lanka. In order to overcome this situation, this study introduces an automated image processing based software solution with 2D facial feature templates targeting the Sri Lankan population. Thus, this was the first ever approach that creates the 2D facial feature templates by incorporating both medically defined indexes and relevant aesthetic aspects. Hence, this research study is comprised of two separate analyses on anthropometric indices and facial feature shapes which were carried out targeting the local population. Subsequently, several evaluation techniques were utilized to evaluate this methodology where we obtained an overall success rate as 70.19%. The ultimate goal of this research study is to provide a system to the law enforcement agencies in order to carry out an efficient and effective facial composite process which can lead to increase the success rate of suspect identification.
In the occurrence of crime incidents, the law enforcement agencies and police officers often depend on eyewitness testimony in order to establish the complete view of the crime situation. Eyewitness testimony is an account given by a victim unfolding the occurrence of a criminal activity [
In the event of criminal identification, the way Sri Lanka is analyzing the physical evidences is more similar to a certain extent with international context but still suspect identification through facial composite is far behind than we could ever imagine.
A facial composite is a graphical representation of a human face according to eyewitnesses’ description. In order to overcome the current situation in Sri Lanka, an automated image processing based software solution has been introduced which includes 2D facial feature templates created using the anthropometric indices and shapes.
The failure rate of suspect identification through facial composite sketch has been declared as 92.86% in the year of 2014 [
However there is several recognized composite sketch software available in the international market such as FACES 4.0, E-FIT which is currently used by CIA, FBI and the US Military [
Since Sri Lanka is at its infancy level in facial composition, our attempt is to pinpoint the gap in order to find out the optimum way to automate this traditional procedure by incorporating the information technologies (image processing) and statistical analysis techniques and its related aspects. Since this research proposes a template based composite system, as a first step the need is to construct 2D facial feature templates by analyzing the Sri Lankan dataset. In early studies it was proven that the ability of assessing facial feature appearance has been achieved by anthropometric proportion indices and this has been used for a number of forensic and clinical practices [
Therefore, to create these templates, in the first place a preliminary study was done with 140 undergraduate students (both male and female) aged 20 - 25 to find out the anthropometric proportion indices measurements of eyes, nose, upper vermilion, lower vermilion and these facial feature shapes of Sri Lankans.
Furthermore, this research also consolidates a detailed rule set based decision support system to filter the most possible facial feature templates according to eyewitness description. Thus the incorporation of Information Technology into the Sri Lankan facial composite process will be a guide to a more accurate and efficient suspect identification procedure with the help of more advanced facial composite process. Hence we have taken the very first step to provide Information Technology based solution for facial composite process in Sri Lanka.
The “Facial composite” process has a very long history where it evolved from manual composition to computer based facial composition solutions. The entire facial composite evolution can be categorized into four [
The “Photo fit” was developed in three versions in order to cover the population of male Caucasian, female Caucasian and male Afro Asian [
The third generation evolved with the software solutions where the individual facial feature templates are composited with the help of the computer technology [
Finally, the fourth generation of facial composite system has been evolved with the concept of genetic algorithm [
From 1990, feature based selection method was used for most of the composite software solutions. The products such as FACES, Mac-a-Mug Pro are commonly used software by the United States [
The entire aforementioned computer based software systems such as EFit, EvoFit and Faces increase the efficiency of traditional manual methods. Since almost all these international composite systems were developed targeting the facial features of foreign people, these are not feasible to utilize these software systems in Sri Lanka [
Anthropometric proportion indices have been used in a considerable number of facial analysis studies conducted for various facial feature types and different races in several fields such as medicine, forensic, plastic surgery. Previous researches related to numerical measurements of facial feature analysis have done similar calculations with the use of medically accepted facial indexes using anthropometric soft tissue landmarks [
According to the literature, different formulas have been stated for Facial Index, Nasal Index, and Eye Index. A similar anthropometric calculations was applied in the upper and lower vermilion index calculations as there were no such formulas were defined for these two indexes.
As stated in previous studies facial index [
Research conducted in [
Mainly, the proposed solution consists of two steps which need to be carried out to create an ultimate composite image. They are Pre requirement stage and Building composite image.
As the pre requirement stage of our proposed process, the 2D facial feature templates were created based on the analysis of two relevant parameters known as facial feature indices and facial feature shapes. Henceforth the index measurements of facial features were taken to obtain the size whereas the available shapes of main facial features of Sri Lankan people were considered to obtain the shapes in the process of template creation.
The dataset which consists of 140 frontal standard facial photographs of undergraduate students comprising both male and female between 20 to 25 age categories, were captured in order to create the facial feature templates. The same standard environment setup was used to provide a similar condition (frontal view, neutral expression, daylight condition, same distance from camera to personal object) for the entire process of capturing photographs.
This research study has focused only on main facial components and subsequently the 2D facial feature templates have been created for eyes, nose, mouth and face. These templates were created using the results obtained from the two sub studies carried out for index analysis and facial shape analysis of Sri Lankans. Since the weight of a person can affect the width and height of the facial features, the collected data set was categorized based on the weight categories using BMI (Body Mass Index) values, as BMI can be used to screen for different weight categories.
In the creation of facial feature templates, it was required to carry out an in-depth analysis of facial features. Even though there have been initial studies conducted with regard to facial indices in Sri Lanka [
The facial feature detection library, Luxand Face SDK was used for the facial feature detection from facial photographs as it is comparatively more reliable and accurate over manual measuring techniques and other related tools. Using such libraries would overcome inefficiencies related to the manual measuring techniques such as subjectivity and high time consumption. Subsequently the distance between the feature points and the indices were calculated via the Java based tool stated above. Initially, the facial feature index formulas which are mentioned in the literature review were used to measure the index values of each facial feature. Subsequently they were classified further into different categories as shown in Tables 1-5 in order to obtain a more accurate categorization. Once after classifying each facial feature, the width and height of these facial features were measured in millimeters and the average value of these measurements were calculated to decide the most possible width and height to represent the whole data set for each category.
Eye classification | Eye index range | Eye index mean | Eye width mean | Eye height mean | ||||
---|---|---|---|---|---|---|---|---|
Male | Female | Male | Female | Male | Female | Male | Female | |
Narrow | <274.99 | <249.937 | 255.95 | 228.94 | 7.46 ±0.44 | 7.26 ± 0.66 | 2.88 ± 0.22 | 3.10 ± 0.22 |
Medium | 275 - 325.1 | 250 - 293.999 | 296.91 | 272.59 | 7.44 ± 0.51 | 7.25 ± 0.49 | 2.51 ± 0.28 | 2.69 ± 0.22 |
Wide | >325.29 | >294 | 341.37 | 326.32 | 7.44 ± 0.68 | 7.56 ± 0.33 | 2.13 ± 0.21 | 2.31 ± 0.14 |
Nasal classification | Nasal index range | Nasal index mean | Nasal width mean | Nasal height mean | ||||
---|---|---|---|---|---|---|---|---|
Male | Female | Male | Female | Male | Female | Male | Female | |
Narrow | <86 | <82.999 | 83.54 | 78.04 | 9.38 ± 0.90 | 8.44 ± 0.58 | 11.40 ± 0.88 | 10.80 ± 0.90 |
Medium | 86.1 - 99.2 | 83 - 91.999 | 94.84 | 88.24 | 9.54 ± 0.71 | 8.65 ± 0.59 | 10.39 ± 0.84 | 9.90 ± 0.67 |
Wide | >99.3 | >92 | 102.69 | 94.49 | 9.80 ± 0.68 | 9.23 ± 0.64 | 9.62 ± 0.89 | 9.64 ± 0.58 |
U.V classification | U.V index range | U.V Index mean | U.V width mean | U.V height mean | ||||
---|---|---|---|---|---|---|---|---|
Male | Female | Male | Female | Male | Female | Male | Female | |
Narrow | <594.1 | <808.999 | 514.13 | 1487.95 | 12.06 ± 0.95 | 11.75 ± 1.05 | 2.33 ± 0.31 | 1.87 ± 0.32 |
Medium | 594.2 - 789.8 | 809 - 1191.999 | 686.94 | 897.92 | 12.67 ± 1.04 | 11.59 ± 0.65 | 2.07 ± 1.53 | 1.30 ± 0.1 |
Wide | >789.9 | >1192 | 879.49 | 630.97 | 12.63 ± 1.08 | 11.81 ± 0.98 | 1.49 ± 0.34 | 0.79 ± 0 |
L.V classification | L.V index range | L.V index mean | L.V width mean | L.V height mean | ||||
---|---|---|---|---|---|---|---|---|
Male | Female | Male | Female | Male | Female | Male | Female | |
Narrow | <443 | <444.999 | 430.37 | 396.71 | 12.35 ± 1 | 3.05 ± 0.29 | 2.87 ± 0.34 | 2.94 ± 0.31 |
Medium | 443.1 - 553.1 | 445 - 535.9 | 541.11 | 484.06 | 47.54 ± 1.03 | 11.25 ± 2.44 | 2.36 ± 0.26 | 2.98 ± 2.04 |
Wide | >553.2 | >536 | 648.46 | 586.96 | 13.37 ± 0.84 | 12.07 ± 2.83 | 1.95 ± 0.29 | 2.00 ± 0.20 |
Face classification | Face index range | Face index mean | Face width mean | Face height mean | ||||
---|---|---|---|---|---|---|---|---|
Male | Female | Male | Female | Male | Female | Male | Female | |
Narrow | <79.9 | <77.9 | 76.49 ± 2.82 | 75.39 ± 1.64 | 178.04 ± 12.80 | 168.59 ± 10.21 | 227.34 ± 15.35 | 216.34 ± 14.60 |
Medium | 80 - 88.9 | 78 - 82.9 | 84.29 ± 2.49 | 79.76 ± 1.44 | 174.64 ± 8.85 | 166.18 ± 8.89 | 242.27 ± 17.32 | 220.36 ± 13.75 |
Wide | > 89 | >83 | 91.99 ± 2.42 | 84.93 ± 1.97 | 174.69 ± 9.59 | 168.42 ± 11.22 | 243.40 ± 28.24 | 233.86 ± 12.59 |
The results of the eye, nasal, upper vermilion, lower vermilion and face index analysis of each facial feature are presented in Tables 1-5 respectively. As the main outcome of this analysis, average values of index, height and width were calculated. The below mentioned mean values of width and height were used to construct the 2D facial feature templates.
Since there are no other studies conducted on facial feature shapes for Sri Lankan context, shape categorization has been conducted based on aesthetic aspects with the help of the domain experts’ knowledge. Therefore, two domain experts in the aesthetic field were consulted to classify the facial feature shapes.
The same 140 frontal photographs were used to detect the different shapes of Sri Lankan people’s facial features. All the 140 photographs were inspected by the domain experts in aesthetic field Mr. Winnie Hettigoda, the Senior lecturer, of multimedia art Faculty of Visual Art in University of the Visual and Performing Arts and the forensic artists at Criminal Records Division in Sri Lanka categorized the different shapes of facial features.
During the classification process the two domain experts were able to identify 6, 10, 8 and 7 types of shapes for face, eye, nasal and mouth respectively for males. Similarly 6, 8, 8 and 7 types of shapes for face, eye, nasal and mouth were identified respectively for females.
A facial feature template based composite system was implemented to computerize the manual hand drawn facial composite process to enhance the efficiency of the current manual process. The created 2D facial feature templates were then added to the system afterwards the facial composite process will be take place at this stage with the help of the eyewitness description. Subsequently an interface has been implemented to derive the facial feature templates and each type of templates will be shown in different tabs in composite software system from which the eyewitness can select the appropriate facial feature templates to produce the ultimate suspect’s face.
Face pool comparison and resemblance techniques were carried out as the evaluation methods for this study. The evaluation was conducted only for male category as the facial feature templates were constructed only targeting male indices and shape categories. As shown in
As the second evaluation method the resemblance evaluation technique was conducted via online survey with 51 respondents. As shown in
Image No | Test Case 1 | Test Case 2 | Test Case 3 | Test Case 4 | Test Case 5 | Test Case 6 |
---|---|---|---|---|---|---|
Image 1 | 13% | 0% | 72.7% | 6.1% | 0% | 0% |
Image 2 | 13.3% | 12.1% | 9.1% | 6.1% | 18.2% | 81.8% |
Image 3 | 4% | 69.75% | 12.1% | 12.1% | 18.2% | 0% |
Image 4 | 78.7% | 12.1% | 6.1% | 72.7% | 18.2% | 6.1% |
Image 5 | 2.7% | 6.1% | 0% | 3% | 45.5% | 12.1% |
Features | Case 1 | Case 2 | Case 3 | Case 4 | Case 5 | Case 6 |
---|---|---|---|---|---|---|
Face | Similar 47.4% | Similar 31.6% | Similar 42.1% | Average 42.1% | Similar 36.8% | Similar 37.3% |
Eye | Similar 42.1% | Very Similar 36.8% | Similar 47.4% | Similar 47.4% | Average 36.2% | Similar 47.4% |
Nose | Similar 36.8% | Similar 21.1% | Similar 42.1% | Similar 47.4% | Slightly similar 31.6% | Similar 35.3% |
Mouth | Average 47.4% | Similar 31.6% | Average 47.4% | Average 47.4% | Slightly similar 42.1% | Average 33.3% |
The overall success rate which obtained from this research study was 70.19% for the 6 test cases whereas the success rate of the prevailing international software gives 60% [
According to the evaluation phase, test cases 1, 2, 3, 4 and 6 have taken the higher accuracy rate than the success rate gain from international software solutions. In the proposed solution, the facial feature templates were constructed using the medically defined indexes which were ideal for Sri Lankan people and the aesthetic aspects were also considered when placing the facial templates for the ultimate composite image. These well accepted measurements and aesthetic aspects were put together in constructing facial composite image via this process. Therefore this has become one major reason behind the highest accuracy rate achieved from the proposed process.
However case study 5 has obtained the lowest result as 45.5% and it clearly depicts that this attained success rate is lower than the average success rate which was derived from the proposed process. Moreover the results obtained from the resemblance evaluation shows face, eye features of composite image of case 5 were presented high equivalence rate whereas the features that presented a low equivalence were nose and mouth. Hence it is obvious that the nose and mouth features were the major causes which mislead the audience to select the wrong person for case study 5. And the root cause for this mismatching nose and mouth templates were selected to construct this composite image because only limited nose and mouth shapes were identified from the process as our data set was 140 and it was not enough to represent the whole population. But still this drawback can be eliminated, if the same procedure was applied for large dataset and as a result more nose and mouth shapes can be identified through the shape classification method as mentioned in section “Commonly available shapes of facial features”.
The present study has been able to establish an efficient facial composite process for Sri Lankans which can achieve a considerable amount of high success rate. Therefore the findings and the results of this study will definitely helpful in the forensic science and will eliminate the drawbacks in suspect identification via facial composite.
In the field of forensic science, the facial composite plays a major role in the process of tracking down the criminal when there is no physical evidence found from the crime incident. Henceforth the facial composite image becomes one of the major clues in the investigation procedure.
The study introduces an automated image processing based composite software solution which includes 2D facial feature templates created for the first time targeting the local population to eliminate the inefficiencies in the current manual facial composite procedure in Sri Lanka.
Since these 2D Facial Features templates have been created targeting Sri Lankan context, the analysis of facial feature variations of Sri Lankans were required before creating these facial feature templates. Therefore two major sub researches have been carried and the results obtained from these sub studies were incorporated to construct the 2D facial features templates needed to carry out the proposed facial composite process.
The study conducted was evaluated with face pool and the resemblance techniques and obtained 70.19% accuracy indicating that this proposed process has a significant impact on dramatically increasing the success rate of the facial composite comparative to the prevailing success rate. Hence this computerized system will significantly improve the existing manual process and benefit professionals involved in the criminal justice process mainly police and crime investigators and ensure justice for the victimized person.
A special acknowledgement goes to the advisor Mr. Winnie Hettigoda, senior lecturer of University of the Visual & Performing Arts and Senior Superintendent of Police Mr. Priyantha Jayakody, Director of Criminal Record Division in Sri Lanka who helped us in gaining the relevant domain knowledge. Furthermore, authors are thankful to Mr. Nadaun Liyanage for the immense assistance given for the success of the automated composite system.
P. B.Jayasekara,L.Sivaneasharajah,M. A. S.Perera,J.Perera,D. D.Karunaratne,K. D.Sandaruwan,R. N.Rajapakse, (2016) Facial Index Based 2D Facial Composite Process for Forensic Investigation in Sri Lanka. Forensic Medicine and Anatomy Research,04,7-16. doi: 10.4236/fmar.2016.41002