The truth: The current recognition accuracy of AI technology exceeds 90%
To welcome the arrival of the "National Day" holiday, many scenic spots have installed facial recognition gate machine system at the entrances to check tickets. This also made the topic of " face recognition technology unable to accurately identify twin " aroused the curiosity of many people, and there are many people who agree with this statement.
"In fact, using face recognition technology to accurately identify twins is not a new problem in the artificial intelligence (AI) world. Even identical twins has long had AI to achieve more than 90% recognition accuracy, although this is only achieved under laboratory shooting conditions, and the scale of experimental samples is relatively small." Qiu Bo, director and professor of the Department of Electronic Information Engineering, Hebei University of Technology, introduced.
At present, the common face recognition system mainly includes 4 components: face image acquisition and detection, face image preprocessing, face image feature extraction, and matching and recognition. When performing facial recognition, the facial recognition system will first use computer image processing technology to extract portrait feature points from the video, and then use the principle of biostatistics to analyze and establish a mathematical model, that is, face feature templates; then compare and analyze the built face feature templates with the face of the subject, and give a similar value based on the analysis results. Finally, this value is used to give the face recognition result of the subject.
"The overall appearance of the facial features of identical twins is very similar, so it is indeed difficult for AI systems to recognize them." Qiu Bo said that even for twins, facial features are still subtle, and when face recognition is performed on-site, different expression changes can be captured using video information. Such face recognition technology has the ability to distinguish the subtle differences in the face of twins, and can even build the unique dynamic facial 3D model of the twins.
However, this requires collecting a large number of face data of identical twins as training samples, allowing the face recognition system to conduct special training and recognition. If deep learning technology is used, the number of neural networks and layers and nodes are often involved, resulting in the cost of each training being very high. Therefore, although facial recognition can achieve relatively high accuracy in distinguishing identical twins under laboratory conditions, in real life, other biometric features such as fingerprints, iris, and voiceprints are generally used to assist in recognition.
Source: China Science Network
