亚信科技

As an important research area of AI, facial recognition has been widely used in work and daily life: facial recognition to unlock smartphone, paying with your face, facial recognition-based intelligent access control, etc. However, facial recognition requires high accuracy, so there will be mistakes if facial recognition goes wrong. AsiaInfo’s facial recognition system for intelligent access control developed by AsiaInfo’s AISWare AI² platform is tested with 100% self-optimization method and achieves LFW verification accuracy of over 99.1%, which leads the industry in both convenience and security.

Facial recognition works to identify Mr. Wang only when he stands still

Mr. Wang works at a tech company that uses advanced facial recognition technology for access control system. One day, Mr. Wang found it was almost time for work, but he thought he might not get late if he moved more quickly to do facial recognition for access control, which saved him time for taking out and swiping card. Then Mr. Wang quickened his steps towards the door and smiled at the facial recognition system. But unfortunately, the system failed to identify him and the door remained closed, because he was in such a hurry. Mr. Wang had to step back and stand still for the system to identify correctly. Mr. Wang was late finally.  

Does facial recognition system capture only static image? Do you have to stand still to be correctly identified?

Meanwhile, Mr. Wang also finds access control system doesn’t work in the case of insufficient light. Worse still, some of his colleagues successfully pass access control even with their photos.

AsiaInfo’s facial recognition for intelligent access control: effective solution to all problems

Recognized as an industry leader in telecom software and relating services, AsiaInfo has been dedicated to developing new technologies of big data, AI and IoT, particularly AI is a top priority for AsiaInfo’s R&D departments. In 2018, AsiaInfo set up an AI Communications Laboratory consisting of data scientists, algorithm experts and business experts and independently developed AISWare AI² platform for AGI (Artificial General Intelligence). AsiaInfo’s facial recognition system for intelligent access control is developed by the object & pattern recognition team of AI Communications Laboratory via AISWare AI². The system has solved major problems of failure to capture dynamic image, failure to identify in the case of insufficient light and passing with photos or videos, etc. It has brought a completely new intelligent experience: higher recognition rate, faster and safer.

1.3 frames per second to capture dynamic image: video capture balances real-time factor and computing power, and the webcam takes 3 frames per second to ensure that a person’s facial image can be captured while walking;

2.Multi-target tracking to get best image: multi-target detection and tracking is conducted; facial position, size and expression are judged along with facial profile detection in the tracking process, and highest-resolution and best facial image is captured according to light and imaging conditions, etc.

3.Accept-reject method used to increase recognition speed: the system opts not to recognize facial profile when it surpasses threshold value, which minimizes computing power and time; it takes less than 300 milliseconds from video capture to accurate identification, and it may complete facial recognition when a person goes by;

4.Multiple methods used to improve recognition accuracy: facial adjustment & alignment technology is used to greatly improve recognition accuracy; when the images are free from distortion, dimensionality reduction is done to high-resolution images; image enhancement method is adopted for those exposed to weak light and shadow to increase recognition rate in the case of weak light and excessive exposure;

5.Optical flow detection for accurate identification: LBP, image classification and optical flow model are used to easily identify photo tricks. AsiaInfo has special optical flow algorithm that can distinguish a real person from a photo based on the effect of micro-expressions on optical flow, which effectively prevents the attacks of high-resolution images, images and videos produced by HD display devices (e.g. cell phone, PAD), etc.

With advanced facial recognition technology and strong computing resources based on AISWare AI², AsiaInfo’s facial recognition model for intelligent access control uses data set containing 9,000 people and 2.83 million Asian facial photos and adopts self-optimization DNN training process; pre-processes facial alignment and front face limits; processes captured images in forms of over exposure, under exposure and AR, making LFW verification accuracy beyond 99.1%; the accuracy can also be over 97% in the case of false recognition rate of 1‰. They system has been applied in Jiangsu Telecom, Zhejiang Mobile and AsiaInfo Beijing Summary & Exhibition Dept.

Facial recognition is a small step for AsiaInfo’s AI exploration. AsiaInfo’s innovative breakthrough in facial recognition technology marks its further progress in 5G AI applications. In the future, AsiaInfo will focus more on computer vision and deep learning, continue to pursue cutting-edge technology and framework and take advantage of its own capabilities in technology, product, integration and service to boost AI applications in 5G and provide more smart experience for enterprises.