So-called Biometric identification Technology is the close integration of computer technology with optics, acoustics, biosensors, and biostatistics principles to identify individuals using their inherent physiological characteristics (such as fingerprints, facial features, irises, etc.) and behavioral characteristics (such as handwriting, voice, gait, etc.). Fingerprint recognition has the highest market share. It is estimated that the biometric identification technology market will reach $25 billion by 2020, with an average annual growth rate of about 14% over 5 years.

　　 Facial and iris recognition gaining popularity

　　Li Ziqing, director of the Biometrics and Security Technology Research Center at the Institute of Automation, Chinese Academy of Sciences, told the author that non-contact biometric identification technology refers to technology that can collect identity characteristic information without direct contact between the terminal device and the user. Iris recognition, facial recognition, voiceprint recognition, and retinal recognition all belong to this type. Unlike voiceprint recognition, which has high requirements for the sound environment, and retinal recognition, which may cause inconvenience to people with corrected vision, the industry generally believes that facial recognition and iris recognition have more extensive and convenient application scenarios, and their commercialization prospects are more promising.

　　Iris recognition has entered our daily lives. Currently, more than 50 airports worldwide have systems that allow passengers to pass through directly using only iris recognition. Singapore also began using "iris scanning" this year as another form of identity verification for travelers and Singaporean citizens in addition to photos and fingerprints. Iris recognition is also playing a major role in the financial field: In order to avoid the theft of traditional bank "username + password" information databases by illegal personnel, some banks have begun to add iris recognition verification, that is, only when the iris scan information and account information completely match can customers obtain authorization for fund withdrawal, upgrading the security level comprehensively.

　　Compared with fingerprints and digital passwords, iris recognition has higher security in internet payment applications, which has also made iris scanning systems favored by Samsung and other brands of mobile phones. In the view of Qiu Xianchao, deputy general manager of Beijing Zhongke Hongba Technology Co., Ltd., "iris phones" may become an indispensable "smart organ" for people in the future.

　　However, the public's expectations for facial recognition are even higher than for iris recognition: Can "facial recognition for meals" and "facial recognition for passage" really be realized? "Now in the Baidu Science and Technology Park, employees can pass through the "facial recognition access control gate" in 1 second, without the need for employee cards or waiting." Lin Yuanqing, director of Baidu Deep Learning Laboratory, described the application of Baidu's facial recognition technology to the author: "The next application scenario may be the Baidu cafeteria—Baidu employees can say goodbye to employee cards when they go to the cafeteria to eat, and tens of thousands of people will be the first to enjoy the convenience of "facial recognition".

　　 Pursuing more natural AI interaction

　　Facial recognition has two different scenarios: one-to-one facial comparison and one-to-N facial recognition. The former is commonly used in banks and public security systems, generally requiring the submission of personal ID information and facial features, and then the system compares the ID photo with the submitted information one-to-one; the latter has also been used in criminal investigation and other fields, but due to the difficulty of one-to-N recognition accuracy, the public security system traditionally only uses this technology to assist in investigations, and is "not sure" about the results and "cannot lock down" the suspect.

　　Baidu's "facial recognition" technology chooses accurate one-to-N recognition. Lin Yuanqing revealed that Baidu has been focusing on one-to-N recognition applications since 2016 and has made great technological breakthroughs, "and can now make one-to-N facial recognition very accurate." Baidu previously announced that the accuracy of Baidu Brain's facial recognition monitoring has reached 99.7%.

　　If the accuracy is not enough, the application of facial recognition access control gates may cause Baidu employees to be blocked outside, which is unacceptable." Lin Yuanqing told the author that Baidu's facial recognition accuracy has two indicators: false acceptance rate and acceptance rate. "The acceptance rate is easy to understand, which refers to the accuracy of matching. The false acceptance rate refers to "letting non-Baidu people pass easily"; this indicator is close to 0; the acceptance rate is above 99%.

　　In fact, achieving high accuracy in one-to-N accurate recognition is not easy. This also makes scholars including Li Ziqing doubt: "Is it really as high as advertised?" "Achieving high accuracy in one-to-N is the most difficult." Lin Yuanqing also admitted that, especially as the database grows—currently the database for Baidu's access control gates and future Baidu cafeterias is 20,000 to 50,000 people—this poses a more serious challenge to the accuracy and security of recognition.

　　But this is exactly what Baidu excels at." Lin Yuanqing said that only by achieving accurate one-to-N recognition can we truly bring revolutionary changes to the original facial recognition methods, making the interaction in the artificial intelligence (AI) era more intelligent and convenient. "We must achieve this, because we hope that future AI interfaces will be very natural and seamlessly integrated with our lives.

　　 Quietly entering the 2.0 era

　　In fact, when facial recognition was first applied, "accuracy" has always been a bottleneck for in-depth application. In particular, the initial two-dimensional facial recognition could not achieve a high degree of matching between "face" and "person." "Facial tracking and recognition technology has entered the 2.0 era." Ye Zhou, founder of the artificial intelligence technology company ULSee, told the author: "Traditional facial recognition technology is mainly based on facial recognition of visible light images, but this method has insurmountable defects, especially when the ambient light changes, the recognition effect will decrease sharply, and there are other problems such as large differences between side faces and front faces, and fraud using photos.

　　A product manager from a biometric identification technology company told the author, "Now, liveness detection has replaced static image recognition, and infrared light as a means of collecting facial features has made up for the shortcomings of visible light recognition; real masks, photos, and mobile phone videos cannot "get away with it." In addition, this product manager said that today, people no longer need to "shake their heads" and "interact" with the machine to ensure the authenticity of the information when facing a facial scanner.

　　Currently, facial key point detection technology can accurately locate key areas of the face, and can even support a certain degree of occlusion and multi-angle faces, which greatly improves accuracy. Even so, the recognition of twins and before and after cosmetic surgery remains a technical bottleneck for facial recognition. In this regard, Li Ziqing told the author that iris recognition may be one of the breakthroughs to solve this problem.

　　Unlike the uncertainty of facial morphology in different life stages, the iris has high certainty and uniqueness. From fetal development to death, it maintains its morphological certainty and difference from other individuals throughout the process. This attribute of the iris has long been valued by experts in biometric identification technology. As an important part of non-contact biometric identification, iris recognition also does not require direct contact between the eyes and the device, making it more convenient and faster, and dual authentication of iris and face provides "double insurance" for facial recognition.

　　However, the average human iris diameter is 12mm, while the average iris diameter of Chinese people is smaller. The smaller exposed area makes it difficult to collect iris feature information. Li Ziqing said, "In addition, the richness of iris color in Chinese people is far lower than that of Westerners, which is also one of the problems faced by iris recognition."

　　With further technological development, iris recognition has dedicated infrared cameras and dedicated infrared lights, making recognition more feasible. In addition to the upgrade of recognition scanning rays, the iteration of machine learning and algorithms themselves is also an important aspect of solving the credibility problem. Lin Yuanqing said that data, algorithms, and products are a positive cycle, the iteration of data and the evolution of algorithms promote the production of better products, and better products in turn promote the further improvement of data and algorithms. "In fact, the breakthrough of Baidu's facial recognition technology is inseparable from the large investment in algorithms in recent years." Lin Yuanqing said, "Baidu's entry into the facial recognition industry follows a technological route."