During the epidemic, wearing a mask has become an indispensable "equipment" for people to go out, but wearing a mask has also brought a lot of inconvenience to daily life. Especially after the resumption of work and resumption of production, problems such as the inability to pass a gate with a mask, and the low efficiency of manually registering access information are even more disturbing.
Recently, a set of "black technology"-"precision epidemic control" system jointly developed by a Shenzhen company was equipped with "new three-dimensional face recognition technology", making it possible to "face brush" with a mask. How does this technology guarantee the accuracy of recognition? What other application space does the existing results have?
New 3D face recognition technology is developed based on 2D face recognition and low-precision 3D measurement recognition technology
At the moment, face recognition is no longer new. At the high-speed rail station, passengers can confirm their identity through face recognition and enter the station. Manual sign-in to work has also gradually been replaced by "swipe your face".
It is understood that the current face recognition technology is divided into two-dimensional face recognition technology and three-dimensional face recognition technology. "Most of the current markets use two-dimensional face recognition technology with relatively low accuracy." Mike Wang, CEO of Chongqing Huifan Technology Co., Ltd. said that two-dimensional face recognition technology is greatly affected by the external environment and has weak anti-counterfeiting capabilities. Therefore, in areas with high security requirements, two-dimensional face recognition technology cannot carry the banner of security responsibility.
"Make-up, cover the face, wear a headgear, are all tricks in the face of high-precision 3D face recognition technology." Mike Wang said that with the market's increasing requirements for face recognition accuracy, 3D face recognition technology came into being.
"Compared with our common two-dimensional face recognition technology, the three-dimensional face recognition technology has a big difference in measuring depth accuracy." Mike Wang told reporters that the two-dimensional face recognition technology is based on the planar image of the face. It ’s like a photo of a face. But in fact, our faces are bumpy. ”Mike Wang, for example, said that the nose of a westerner is generally higher than that of an Easterner, and Zhang ’s eyes are more sunken than Li ’s. "Three-dimensional face recognition technology incorporates these facial features into the measurement."
Three-dimensional face recognition technology also has different levels of accuracy in measuring depth. According to reports, the new three-dimensional face recognition was developed on the basis of two-dimensional face recognition and low-precision three-dimensional measurement recognition technology. It was developed to solve the bottlenecks and major defects of the two.
Number of point clouds: To improve the accuracy of the measurement depth and ensure that enough point clouds are collected in the area not covered by the mask
When wearing a mask, only the roots of the nose, eyes, and eyebrows are left on the face, which is difficult to recognize by human eyes alone. So, why can the new 3D face recognition technology recognize faces with masks? "First of all, this has to do with the depth measurement accuracy of the technology." According to reports, the general 3D face recognition technology has a measurement depth accuracy of 1 to 3 mm, the new technology can reach 0.2 mm or less.
In the new three-dimensional face recognition technology "eye", the human face is more complicated than expected. For example, from the perspective of the new three-dimensional face recognition system, moles on human faces are of different sizes. "How to collect a sufficient number of point clouds in the part that is not covered by the mask is a difficult point for face recognition while wearing a mask." Mike Wang said.
The number of point clouds can also be called feature information. "For example, the shape of eyebrows, the size of moles, the thickness of lips, etc." Mike Wang said that the new 3D face recognition technology has high measurement depth accuracy, so it can capture more point clouds. "The greater the number of point clouds, the more facial feature information captured, the more detailed information, and the higher the accuracy, the better the face recognition."
"The number of point clouds that iPhoneX captures on a person's full face is 30,000, and the new three-dimensional face recognition technology is 50,000 to 80,000." Mike Wang said, taking iPhoneX as an example, when the mask covers When the lower part of the face, a large part of the information features will be lost, and it is difficult for the system to piece together a face based on the remaining facial detail information. However, because the new three-dimensional face recognition technology can capture a large base number of point clouds, even if the mask covers most of the information features, the technology can still successfully recognize faces with the remaining number of point clouds.
3D face modeling: Through algorithm optimization, a 3D face model that approximates a real face is established.
"The temperature is normal and you can pass." The reporter saw in the company that the staff stood in front of the access control system without removing the mask. In just 2 seconds, the machine completed face recognition and completed the entry and exit registration in the background.
"In terms of preliminary test results, the recognition accuracy of wearing masks per 1,000 people is 98%." Mike Wang said that achieving such a high accuracy rate is inseparable from the accurate 3D face modeling of the technology.
"Face recognition is generally divided into two steps, one is face detection, and the other is face matching." Mike Wang told reporters that the recognition process is that the face machine first finds the face from the video image and analyzes it. The "key points" are calibrated to obtain the shape and position information of the facial features on the face, such as the distance between the two eyes or the angle of the nose-to-mouth corner line in the horizontal direction, etc. The face reconstruction model is finally compared with the existing data in the personnel database to identify the corresponding person.
"Only when the three-dimensional face model approaches the real face to the maximum, the recognition accuracy can be higher." Mike Wang said that the key point is the algorithm. It turns out that the machine is not good at identifying portraits. For example, a middle-aged man in the eyes of the machine is just a string of numbers composed of 0 and 1. It can't understand the meaning of this image. If a machine wants to learn to recognize an image, it needs a programmer to write a programming algorithm for it.
At the algorithm level, the existing face recognition algorithm model can be optimized for a combination of facial features and local features. "We have trained the data specifically on the face from the nose to the forehead through the texture information and depth information of the face to optimize the algorithm," said Mike Wang.
I can "recognize" even a side face (small)
Prior to the outbreak, the "Precise Epidemic Prevention and Control" system equipped with the "New 3D Face Recognition Technology" system has been used in hospitals and construction sites. Relying on this system, doctors at the West China Second Hospital of Sichuan University can enter and leave the operating room without removing their masks or their badges and access cards. "This saves time and is more hygienic," said Mike Wang.
With the improvement of the domestic epidemic situation, various places are returning to work and school in an orderly manner. In places where there is a large number of people, such as airports and schools, manual registration of entry and exit information and one-to-one detection of body temperature are time-consuming and error-prone. Zhang Zili said that the system can realize the intelligent management of enterprises, units, and schools, and complete the work of quickly registering access information, tracing the trajectory of personnel movements, and dining at different times.
The system can also play a big role in public safety. For example, in the field of security, many criminals often avoid surveillance systems such as "sky eye" by wearing hats and covering their faces. New three-dimensional face recognition technology can help police better solve cases. "Even if it is a side face, the new 3D face recognition technology can make criminals nowhere to be."