Colour vs Infrared Face-Rec: Ready for Full-Scale Deployment
Aurora-AI is best known for development of Artificial Intelligence solutions. One example being our world-class Infrared (IR) Face Recognition systems, with thousands of installations globally including major implementations at Heathrow and Manchester, which have been helping to secure the UK Border for many years. The use of IR provides stability against changes in lighting conditions, meaning our performance is consistent year-round with no impact from sunlight, dull lighting or shadow casting.
Our first generation of deployed AI-based face recognition engine (Stingray-v3) was released over three years ago. The improvement in verification accuracy in comparison to our previous pre-AI face-rec engines was astounding. Not least because those improvements were achieved in just a matter of weeks thanks to the revolutionary AI technology and our unique approach to training these artificial brains on our proprietary AI training cluster. Since then, our team has been working hard on improving the capabilities of our AI cluster, making it bigger, faster, more reliable and capable of self-analysis to automate AI training.
For years, the Stingray-v3 face recognition engine has successfully processed millions of people. It is being used to verify all domestic passengers at Heathrow and Manchester airports and has also been used globally on construction sites to prevent wage fraud. The increase in the accuracy over previous engines allowed for faster enrolments and verifications but also extended use to self-boarding capabilities with the automated identification of passengers. We believe that even now, this is the only deployment of its kind in the world: processing 100% of passengers.
At Aurora-AI we are constantly working on improving the face recognition AIs. Our approach is to train the AI for very specific needs, using exactly the image type, quality and comparison pairs as expected in the real-world use case. This will always perform significantly better than a general face recognition engine trying to suit all purposes. Therefore, our current portfolio contains several face-rec AIs trained on different modalities of images. The modality of choice, always with the highest performance is infrared. It provides stability in changing lighting conditions through our unique flash adjustment technology as well as huge accuracy improvements even when lighting is stable. Besides the infrared face recognition engine, we also offer colour face recognition engines (working on full-colour RGB photographs) and a colour-vs-infrared engine (CvIR) which matches colour passport-style photographs with IR images captured using our specialist IR cameras. Interestingly, the accuracy of the later has just surpassed the accuracy of the original infrared Stingray-v3 engine, which has proven itself on many installations worldwide. In other words, this now means that we can match against passport images to the same level of accuracy and performance, known and proven in real operational environments on a massive scale. The new AI which achieved this impressive milestone is named Orca!
New generation: Orca engines
At Aurora-AI we use a sea creature naming convention to differentiate between new generations of Artificial Neural Network. Our latest, 3rd generation Orca AI engines provides fantastic reductions in error rates:
- IR Orca reduced error rates by 90% over Stingray-v3.
- Colour Orca reduces error rates by 85% compared to our previous colour engines and 17.5% lower than the Stingray-v3 engine.
- Colour-vs-IR (CvIR) Orca provides a 90% reduction in error in comparison to our previous CvIR technology, which is 42% lower error rate than the IR Stingray-v3 engine.
Colour vs Infrared
The fact that our latest colour-vs-IR engine has surpassed the accuracy of the original IR Stingray-v3 engine is a big milestone for the team here at Aurora-AI. It might be difficult to understand how a colour photograph can be compared with an IR image as they each come from different modalities, the colour photograph contains RGB data whereas the IR images are single channel grayscale of a completely different wavelength of light. Our AI research team has found a clever way to train an AI that operates between the two modalities, learning only the common features and ignoring colour information, after all the facial feature shape and geometry are the most important factors in the biometric comparison.
The new CvIR engine, with 42% lower error than the IR Stingray-v3 engine (used for millions of transactions worldwide) proves that the technology has matured enough to be used on a very large scale.
To see this exciting new technology in action, please visit us at PTE 2019 in London ExCel (26th – 28th March), we would love to speak to you. You’ll find us on stand #1090. Please use our contact page to make an appointment.