- This article is filed under:
- » View All
by Nello Zuech, Contributing Editor - AIA Posted 04/20/2006
For as long as I can remember, facial recognition has been a research topic within the computer vision community - even longer than applications in machine vision were being pursued. Why? Because humans recognize each other by face and faces are already being used for many security applications like passports and drivers licenses.
Significantly, the underlying technology involved in facial recognition is essentially the same as that of machine vision: higher resolution cameras, LED lighting, optics and compute technology. Just as the advances in the underlying technology have made it possible to address more and more machine vision applications, those same advances are now making it possible to perform more rigorous facial recognition algorithms yielding improved recognition performance. There are still challenges, however.
Unlike in most machine vision applications where the scene examined and analyzed is constrained to virtually eliminate challenging scene variables; in most facial recognition applications one has to contend with many scene variables. Challenges include lighting (indoor/outdoor), shadows, shading, face position in terms of skew, orientation and translation, as well as the inherent challenges in appearance changes that might be associated with the person being recognized: facial expressions, facial hair, hair styles, affects of aging, changes in weight, glasses or contacts, sun glasses, etc., not to mention disguises.
There are at least 20 companies suggesting they offer facial recognition products in North America. Some of them have migrated to capturing 3D facial images and adapting mapping techniques analogous to those used in 3D machine vision based on capturing point cloud data and mapping to a CAD file. 3D data capture is based on structured light or stereo/multi-camera correspondence. 3D approaches are able to handle the effects of orientation and illumination, but they are more expensive both in terms of dollars and computational requirements than 2D approaches.
In spite of all the advances in facial recognition, actual successful recognition rates are still poorer than what would be acceptable in comparable pattern recognition machine vision applications. It seems that performance is reasonable when the application involves verification rather than pure recognition. In these cases a person would swipe a card or input a PIN number and recognition would be against his specific template or set of templates representing that person. When it comes to classic facial recognition of a person in a database or true identification, the size of the database becomes a major factor in identification reliability.
There are standards associated with facial recognition. The American National Standards Institute’s (ANSI) International Committee for Information Technology Standards (INCITS) has developed its face recognition format for data interchange. The U.S. Department of Defense and the National institute of Industry and Standards have been evaluating face recognition systems since 1993 and continue to do so. Some of their tests have shown that face recognition performance drops linearly with the logarithm of the database size. The goal is to verify 98% of the faces but keep the false positives verification down to 0.1%.
The specific algorithms associated with face recognition are vendor dependent. In general, images are preprocessed/enhanced and a set of descriptors developed, which are then used as the basis of correlating to the database of stored descriptors. After image capture, a “find” routine is employed to locate the face in the space covered by the field-of-view of the camera. Preprocessing can include correcting for pose (position and orientation) and illumination, aligning the face imagery, equalizing the pixel values, enhancing contrast and brightness, eliminating noise and segmenting features. The processed image data is then compressed into a set of descriptors – geometric, photometric, eigenvectors, etc. Decision-making can be based on any number of approaches: principal component analysis, linear discriminant analysis, nearest neighbor analysis, elastic bunch graph matching, etc.
As with other articles in this series, we canvassed input from the over 20 vendors understood to be offering biometric recognition systems based on face recognition. The following were kind enough to provide responses.
- Frances Zelazny, Director of Corporate Communications, Identix, Inc.
- Myles Kendrick, President, XID Technologies Corporation
What follows are their answers.
1. How would you describe your specific facial recognition implementation? Smart camera-based solution? Board-level solution? Software-based solution? Specific market/applications addressed?
[Frances Zelazny – Identix] Identix provides software based technological modules for face finding, template creation, matching, compression, etc. used by integrators and resellers to develop complete solutions. One of these modules is the Identix ABIS® search engine platform used by government agencies of many countries in such programs as visa, voter registration, drivers’ license and ePassport.
ABIS® eases the delivery of large-scale identification solutions for government, law enforcement and commercial users by providing the capabilities – namely, scalability with respect to database size, number of concurrent searches and response time, as well as built-in redundancy to ensure continuous operation – required by large agencies with millions of records to manage.
Aside from government and homeland security applications, Identix’ facial recognition technology is also being used in consumer applications. For example, Nikon’s COOLPIX 7900, COOLPIX 7600 and COOLPIX 5900 models use Identix’ facial recognition technology to enable the user to capture the face he or she wants, every time. The facial recognition feature offered by Nikon, Face-priority AF (Autofocus) automatically detects human faces in the scene and adjusts camera settings faster than humanly possible. These cameras are in fact, the first major consumer application of facial recognition biometrics.
[Myles Kendrick – XID Technologies] We deploy packaged software/hardware solutions, we OEM, we deploy alone or with system integrators depending on the project size.
Summary of applications:
- Manufacturing Plants, Petrochemical Plants
- Healthcare Facilities
- Dormitories & Construction Sites
- Smart ID Documents
- Immigration - Passports, Boarding Pass
- National ID, Health cards, Driving License
- Identity Theft Prevention
- Computer and Internet
- Mobile Authentication
- Gambling & Media
- Banking & Finance
- Law Enforcement
2. What are the underlying principles associated with your specific facial recognition implementation? Image processing/segmentation principles? And image analysis principles?
[Myles] We developed a unique, award winning technology called Predictive Face Synthesis. This technology predicts how a face may appear in the future – from one 2D facial image, it automatically produces a multitude of images with that same face appearing in different lighting conditions, with glasses, beard, rotated face etc. From each image, we extract the face template and store these for matching. So, the database becomes enriched and the chances of finding a match are much improved. This technology enables us to deploy a system that controls the access of 6,000 blue-collar workers, outdoors, 24/7. It is applied to both 1:1 and 1:N scenarios.
[Frances] Identix’ technology combines traditional facial recognition techniques (Local Feature Analysis or LFA) with new skin biometrics to deliver unprecedented levels of accuracy. Local feature analysis looks at geometry of the face or the relative distances between predefined features (e.g., nose and mouth). Skin biometrics, on the other hand, looks for uniqueness in texture and randomly formed features to form a unique skinprint identifier. The skin algorithm is called Surface Texture Analysis (STA). Because skinprint technology relies on the same image capture devices and the same data, it is easily incorporated into traditional face recognition systems to yield exceptional levels of performance. In effect, two different traits of the same image are analyzed at the same time with the results fused together, yielding a performance boost greater than their individual results.
3. Are the image processing and analysis tools accessible for modification or addition by the user? If so, describe, how.
[Frances] We provide analysis tools for users to see how the technology can work with various databases and with different types of cameras.
[Myles] Yes, all of our tools can be modified. All of our software is offered as a software development kit (SDK). In the case of Predictive Face Synthesis for example, it can be plugged to any other face recognition engine as a means to improve its performance or interfaced to the existing engine.
4. Can you provide a one or two paragraph general description of your graphic user interface/man-machine interface? What interface steps need to be taken to set up the application?
[Myles] Our off-the-shelf solutions provide all necessary tools and functions via the interface; from enrollment to verification score checks, event logs etc. Interfaces are also customized for specific customer’s applications.
[Frances] It's not an off-the-shelf solution so there is no interface per se. The integrator or reseller customizes the interface.
5. How do you characterize the performance of your facial recognition system? False recognitions? Accuracy? Repeatability?
[Frances] Identix’ fusion of traditional Local Feature Analysis with the more recently integrated Surface Texture Analysis, has resulted in a facial recognition technology of unprecedented accuracy on par with fingerprint in one to one applications, highly scaleable, and repeatable across multiple applications, and very stable in that underlying skin characteristics “read” by STA do not change. The system however does require high-resolution pictures.
Identix' internal tests on the combined LFA and STA facial recognition technology have demonstrated that for most data image populations, including the databases utilized in the latest US government Facial Recognition Vendor Test (FRVT 2002), the typical Correct Match Rate (CMR) increased substantially. Tests showed performance significantly exceeding the highest levels of performance from the FRVT 2002 results (of which Identix was in the top tier of performance). In addition, Identix tests demonstrated that when photo capture and quality guidelines associated with mug shots and biometric passports from NIST and ICAO are followed, CMR performance was even higher.
Identix’ facial recognition comprises an advanced, mature product offering for scalable facial search technologies that is akin to exiting fingerprint systems currently in use for positive identification. While fully compatible and extremely accurate with existing databases, it offers consistent breakthrough performance for face images meeting the format and quality requirements for biometric travel documents (visa and passports), which are based upon the recent face recognition standards set by the International Civil Aviation Organization (ICAO), the International Standardization Organization (ISO), and the American National Standards Institute (ANSI).
[Myles] Our offering comprises a face recognition engine that incorporates a face synthesis engine that leverages on synthesized images as reference database.
The principle of the Predictive Face Synthesis technology is to generate additional reference images from a single original photo in order to accommodate to multiple scenarios (different poses, lights, facial hair, glasses). This synthesis technology improves the performance of any face recognition engine: it will indeed provide reference images closer to the actual conditions of use. This effect will dramatically reduce the false rejection rate of a face recognition engine while maintaining the same level of security (False Acceptance Rate).
Moreover our face recognition engine leverages on the availability of thousands of synthetic reference images and scenarios to provide a statistics based matching: this feature combined with current state of the art matching techniques increases the reliability of the system. For example, it enables our engine to function in any environment; such as the outdoor, blue-collar scenarios mentioned above.
Finally, the scenario models of the face synthesis can be expanded and applied at anytime after a deployment. The existing facial database can therefore be upgraded to fit new applications or usage scenarios. This will not require capturing new photos of the end-users as can be the case for traditional face recognition systems.
6. What are the specific application issues that one must be attentive to when applying a facial recognition system in a security application? For example, optics issues, lighting issues, resolution issues? Day/night issues? Scene variables? Etc.?
[Myles] The key success factors of a deployment will be on the quality of the reference data:
- The enrollment photos captured during the initial registration must be of good quality (neutral light conditions, frontal face), typically following ICAO recommendations for travel documents.
- The synthesis engine must be configured to reflect the usage scenario of face recognition: Factors such as the expected user behavior (e.g. rotating face, wearing glasses or medical masks) and the expected environment of use (indoor/outdoor, day and night usage) will be considered to select the correct and relevant synthetic models to use.
- Reviewing preliminary performance results and fine-tuning the synthesizer configuration (by adding models fitting any non-forecasted scenario) will maximize the deployment results.
[Frances] The application issues one should be attentive to when applying a facial recognition system in a security application are camera angle, lighting, quality of image feed, subject participation (i.e. the subject looking at the camera).
7. What is it that you require from a prospective buyer to assure that what you provide will satisfy their application?
[Frances] Our requirements from prospective buyers are different for each implementation. Typically this has to do with database sizes or number of templates that will be issued, environment in which the technology will be deployed, number of users, etc.
[Myles] The most critical information required by us is the scenario of usage and the buyer key priorities (level of security, speed, ease of use…). Based on this information, we will provide a default system configuration and can support the buyer to fine tune if necessary.
8. What are the skills required to integrate a facial recognition system in a security application?
[Myles] The skills required are part of the standard skill-set of a system integrator: Database management to support traditional large volume of data and software integration programming to integrate the SDK C++, C and java interfaces available as part of a larger solution infrastructure. For this purpose we provide demonstrators to illustrate how the SDK is integrated in a standard project. System integrators can reuse these examples as part of their own solution. Additionally we provide domain knowledge and training to address the most common issues of a face recognition project: for example, data quality assessment, system performance analysis and fine tuning (e.g. synthesizer configuration).
[Frances] The skills required are application development; our Software Development Kits can be programmed in C++. The ABIS® system is customized using Java and XML.
9. How do you support your products? E.g. help with set-up? Post service support? Training? Warranty? Documentation?
[Frances] Our team provides a wide range of support services that help to ensure success at the lowest cost possible.
Our menu of offerings include:
- Exceptional RFI and RFP support
- System design consultation
- Project implementation support
- Proof-of-concept support
- Application development services
- 24/7 help desk
[Myles] We provide end-to-end support for our partners:
- Documentation on best practices for system installation, usage and configuration
- Support to refine and address specific customer requirements at commercial proposal stage
- Support to design and implement SDK custom functionality at proof of concept stage
- Support to plan and implement a first project
- Support to assess and maximize the performance of the solution provided after project kick-off
- 24x7 technical support
Further, we recommend that each partner implement their first project jointly with XID so as to share our knowledge and overall know-how within the frame of a real case scenario. This principle will help them to gain confidence of the product during future implementations and to quickly anticipate all potential project issues proactively.
10. What are the most important differentiators between competing facial recognition products in the security market today?
[Myles] Most facial recognition vendors are competing on the accuracy of their face matching algorithms. They have currently reached a performance plateau with the major players having only marginal differences in accuracy and their systems are affected by a lack of stability in changing environments. XIDs Predictive Face Synthesis represents a leap forward.
It addresses exactly these remaining issues of these products such as the sensitivity to changing lighting conditions, rotation of faces, growth of a mustache, glasses, etc.
[Frances] The most important differentiators between competing facial recognition products in the security market today are accuracy of facial recognition algorithms, back-end database search and scalability and standards-based interoperability.
11. In your opinion, what are the main product features of the facial recognition products you offer that customers are looking for today?
[Frances] First, customers are looking for scalability and redundancy, so that multiple searches can be conducted at one time. Secondly, they are looking for portability, and therefore require platforms that can be built into devices such as cell phones and cameras. Identix is well represented in both areas.
[Myles] End users want robustness with the ability of the system to function outdoors and for blue-collar worker scenarios. They want quality after-sale service and maintenance. Other vendors are showing interest in Predictive Face Synthesis as a means improve their own engine performance.
12. Are there some emerging changes associated with the underlying technology used in facial recognition in security applications that will impact the performance of facial recognition in security applications? What will those impacts be on specific applications?
[Myles] XID’s Predictive Face Synthesis will make systems more robust and allow greater scope of applications such as time and attendance in the workers market.
[Frances] One emerging change Identix is pioneering encompasses 3D modeling with 2D matching. This entails a modified enrollment procedure whereby multiple views of the subject are captured at the enrollment image and saved to database, so that when the system captures the subject in different poses at different angles and in a variety of lighting conditions, there is statistical increase in accurate matching.
13. What are specific camera and camera connectivity trends associated with data capture for facial recognition applications in the security market?
[Frances] For high-end screening applications, we require standard, off-the-shelf high-resolution cameras.
[Myles] Different engines have different requirements. For XID, we do not require any proprietary hardware.
14. What are other major trends you see associated with future facial recognition in security applications?
[Frances] We see two major trends for the future. First, the use of facial recognition in security applications will entail subjects passing through a chokepoint at which the technology is deployed, rather than reliance on wide area surveillance. Also, we are seeing a trend toward tapping various databases simultaneously, and sharing data across agencies to get the most accurate information in real time.
15. Any advice or tips you can give to someone investigating the purchase of a facial recognition system for security applications?
[Myles] Before you buy, see the same system in operation at a live customer site. Also, if possible, try it as a Proof-Of-Concept (POC) at a similar site/scenario to which you wish to deploy the system.
- Make sure your vendor is going to be there for the long term.
- Check to see what other large scale or similar deployments your vendor has successfully completed.
- Buying the cheapest system is often more expensive in the long run.
There are currently no comments for this article.
Leave a Comment:
All fields are required, but only your name and comment will be visible (email addresses are kept confidential). Comments are moderated and will not appear immediately. Please no link dropping, no keywords or domains as names; do not spam, and please do not advertise.