IRJET- Advance Driver Assistance System using Artificial Intelligence
Facial Recognition and Detection Technical Review
1. Name: Tkeyah Anderson
Advisor: Professor Ayanna Howard
Group Name: PathFinders
Next Generation Facial Recognition and Detection Software
Introduction
This paper reviews current market practices in designing facial recognition systems. The paper will
examine the changes in the implementation of specific algorithms and transition from 2D to 3D facial
mapping. Facial recognition technology is identified as non-intrusive and indirect process that does not
require physical contact unlike fingerprinting. Such capabilities will assist in discretion within
surveillance and security systems. As the need for facial recognition technology increases both in military
and practical commercial usage, improvements in speed and image quality are sought after.
Commercial Applications
Facial recognition systems are designed for use in law enforcement and security systems. As of
September 2014, the Federal Bureau of Investigation, or FBI, in the United States has achieved full
operational capabilities with the use of new facial recognition technology [1], [2]. The United States
Department of Commerce has expressed concerns on the uniqueness of algorithms and templates when
using this technology for image reconstruction [3]. The capabilities aforementioned have extended to
work across databases such that recognition can include images taken in another city or state to address
such concerns. Additionally, the software will provide functionality for status updates on the history of
criminal activity by those stored in the databases that have been labeled as being in “positions of trust
[1]”. The software will be restricted to only tracking those with a criminal history or mug shot.
Operational Algorithms
Facial Recognition technology is designed to first identify a particular object as a face then measure a
variety of facial features through point extraction in order to find a correlation or match. The range of
features includes the distance between an individual’s eyes, nose width, cheekbones, and jaw line. The
industry has since migrated from using a system purely of two dimensional images to being able to accept
and analyze a 3D image or video. Generalized Matching Face Detection Method (GMFD) is incorporated
as a basis for various facial recognition and detection systems [4]. The main logic for facial recognition
within GMFD is a modified Generalized Learning Vector Quantization (GLVQ) algorithm, based on a
neural network, eliminating the issues of facial obscurities. The GLVQ algorithm serves as an integrated
approach to data clustering and detection of various shapes with the use of symmetrical distance analysis.
This algorithm assigns a particular pattern to clusters that are nearby and deemed to be symmetrical. The
2. implementation of the GLVQ permits the detection of linear, spherical, and ellipsoidal detection of
clustered shapes within the contour of the face. As of June 2014, the Face Recognition Vendor Test
(FRVT) serves to implement an enhanced algorithm which reduces the error rate of recognition to 3.1%
with a search speed of one second per 3.2 million populations [5], [6]. The FRVT will have to capability
to pass through demanding national security needs.
Building blocks for facial recognition
Images initially had to be within a database of quality that was in in a controlled environment.
Recognition typically failed if there was a variance in the lighting and shading of the detected image with
that of the stored 2D image forcing a necessary transition to 3D images for accuracy [7]. Under 3D
systems the images are processed through five primary steps; these steps include detection, alignment,
measurement, representation, matching [8]. During representation and matching, 3D images are processed
into a specific numerical code and fed through an algorithm to translate it into a 2D image [8]. The
remaining sixth step of the process is broken into two options based on the needs of the user. The first
option, verification, focuses on a single stored image of the subject for comparison [9]. Identification, the
second option, has a broader range for testing using all available stored images in its database of the
subject to complete the comparison [9]. To extend the analysis, certain systems will take into account the
skin makeup of the subject known as the biometric recognition, the facial skin is analyzed based on
surface texture [10]. Complications arise when the image to be compared has obscurities, such as hair
displaced over the face or lighting that leads to dark shadows on the face.
Current Market Improvement and Practices
The facial recognition technology market is host eight leading companies in the industry [10]. One
particular company, FaceFirst, has shifted its technology to enact a system that is no longer solely based
on the implemented algorithm but includes a human interface [8]. Known as the Operator Center
Software the user will have the option to verify and override a match when the computer cannot
conclusively declare the identity the person of interest [11]. NEC Corporation has developed algorithms
to reduce error in analysis when variations in image quality arise. Such variations include environmental
effects and differences in the angle of the camera when image has been captured [6]. Regions of
complications further extend to the changes in facial expression and aging of the target individual. First
implemented by Google Inc. in Android enabled devices, Apple Inc. holds the patent for 3D object
recognition in facial recognition security for mobile devices [12], [13]. In the company’s patent
application it describes the recurring issue with lighting and facial positioning. The patent includes
methodology behind low power intensive processing and the amount of information displayed by the user
while unlocking the home screen for a cell phone [14].