Home > News > Industry News > How accurate is the facial rec.....
Browse Categories
Face Recognition Temperature(7)
New Arrival(26)
Hot Products(12)
Fingerprint Scanner(23)
Handheld Terminal Series(14)
POS Terminal(5)
Time Attendance Series(14)
Access Control Series(6)
Personal Security Protection(4)
Sales Promotion(35)
Door Lock Series(11)
Camera Series(11)
Explosion Models
Contact us
Tel:+86-23-67305242 :
Facebook:Huifan Technology
Contact Now


How accurate is the facial recognition system? Why is it important?

  • Author:HFSecurity
  • Release on:2020-12-03

In just a few years, face recognition capabilities have been significantly improved. According to tests conducted by the National Institute of Standards and Technology (NIST), as of April 2020, the error rate of the best face recognition algorithm is only 0.08%, while the leading algorithm in 2014 was 4.1%. As of 2018, NIST found that the accuracy of more than 30 algorithms has exceeded the best performance achieved in 2014. These improvements must be considered when considering the best technical specification method. Government actions should be taken to deal with the risks of technological development, not the current risks. Further improvement of accuracy will continue to reduce the risks associated with incorrect identification and expand the possible benefits of correct use. However, as performance improvements provide the impetus for wider deployment, the need to ensure proper governance of technology will become more urgent.

What is facial recognition?

Facial recognition systems are a sub-field of AI technology that can identify individuals from images and videos based on the analysis of their facial features. Today, facial recognition systems are powered by deep learning, which is a form of AI that operates by passing input through multiple stacked layers of simulated neurons to process information. These neural networks are trained on thousands or even millions of examples of problem types that the system may encounter, so that the model can "learn" how to correctly recognize patterns from the data. The facial recognition system uses this method to isolate certain features of the face that have been detected in the image, such as the distance between certain features, the texture of the individual's skin, and even the heat sensation of the face, and compare the resulting facial contours to recognize others Human faces.

Facial Recognition Device
RA08T Face Recognition Temperature Device with 4G Function 

Broadly, facial recognition systems can be used to accomplish one of two different tasks: verification or identification. Verification (also known as 1:1 match) is used to confirm that a person is who they say they are. An example of verification is when a person uses their face to unlock a smartphone, log in to a banking application, or verify their identity while boarding an airplane. Take sample images of human faces during the login process and compare them with known images of the person they claim. Facial recognition algorithms tend to have high accuracy on verification tasks, because subjects usually know that they are being scanned and can position themselves so that their cameras can clearly see their faces.

Recognition (also known as 1:N or 1:Multiple matches) means that the software compares an unknown face with a large database of known faces to determine the identity of the unknown person. It is possible to identify "cooperative" objects that are aware of being scanned or "non-cooperative" objects that are not being scanned. The latter has caused the greatest concern because of concerns that law enforcement agencies or private companies will use the technology to collect data about individuals remotely without their knowledge. However, remote identification can also be used to identify suspects from surveillance video, track missing persons or kidnapping victims, and improve private sector services. Compared with verification systems, the accuracy of remote identification systems is often lower, because fixed cameras are difficult to capture consistent, high-quality images of individuals moving freely in public places.

How accurate is facial recognition?

Under ideal conditions, a facial recognition system can have nearly perfect accuracy. The verification algorithm used to match objects to clear reference images (such as passport photos or facial photos) can achieve up to 99.97% accuracy on standard evaluations (such as NIST's Facial Recognition Vendor Test (FRVT)). This is comparable to the best results of an iris scanner. This facial verification has become so reliable that even banks can rely on it to log users into their accounts.

NIST's FRVT found that when trying to match photos taken 18 years ago, many middle layer algorithms showed an error rate almost 10 times higher.

When considering how the facial recognition algorithm processes the matching faces recorded in surveillance videos, the sensitivity to external factors can be seen most clearly. NIST's 2017 Video Face Evaluation (FIVE) tested the algorithm's performance when applied to videos captured in environments such as airport gates and sports fields. The test found that the accuracy of the best algorithm was 94.4% when using the lens of passengers entering from the boarding gate (relatively controlled setting).  In contrast, the accuracy of the leading algorithm used to identify individuals walking in sports fields (more challenging environments) is between 36% and 87% depending on the location of the camera.

HFSecurity Face Recognition Device

RA08T Facial Recognition Temperature Access Control Device with HDMI 

The five results also proved another major problem in the accuracy of face recognition-the large differences between vendors. Although a top-level algorithm has an accuracy of 87% on the sports field, when the median algorithm only processes images from the same camera, its accuracy is only 40%. [8] NIST's test of image verification algorithms found that the error rate of many facial recognition service providers on the market may be orders of magnitude higher than the leader. [9] Although some vendors have built highly accurate facial recognition algorithms, it is still difficult for ordinary providers on the market to achieve similar reliability, and even the best algorithms are highly sensitive to external factors. According to NIST, the wide range of accuracy among vendors "shows that facial recognition software is far from being commercialized."

What is a confidence score and why is it important?

When adjusting the algorithm to avoid false alarms, it is also important to consider the impact on accuracy. Because facial recognition may be used in situations where users want to minimize the risk of misidentifying the wrong person, for example, when law enforcement agencies use the technology to identify suspects, the algorithm is usually set to only when the matching item is a certain evaluation of them Confidence. The use of these confidence thresholds can significantly reduce the matching rate of the algorithm by forcing the system to cancel correct but low-confidence matches. For example, when matching without any confidence threshold, a set of indicative algorithms tested under FRVT had an average miss rate of 4.7% on photos taken "from the wild". Once the threshold requires the algorithm to return a result only when the result is determined to be 99%, the miss rate jumps to 35%. This means that the algorithm identified the correct individual in about 30% of the cases, but did the identification with a confidence of less than 99%, and therefore reported that no match was found.

The introduction of such a confidence threshold is very important for the following situations: in this case, people are not checking the matches made by the algorithm, and any errors may have a serious impact on those that are misidentified. In these cases, a higher miss rate may be better than allowing false positives, and strict confidence thresholds should be applied to prevent adverse effects. However, when using facial recognition to conduct a so-called survey, you only need to return a list of possible candidates for the operator to check. Since people are checking the results and making a final decision on how to use the information, the confidence threshold is usually lowered return. In these cases, facial recognition is just a tool to speed up human recognition, not for recognition itself. Theoretically, the incorrect match in the lineup produced in this way should be no higher than in the case of not using the technology, because in both cases, the match is ultimately made by humans. However, there is some concern that if some matches are returned with a higher confidence score than others, the operator may be biased towards accepting the conclusion drawn by the algorithm.

HFSecurity Face Recognition Device

When considering how facial recognition is deployed, it is important to understand the appropriate role of confidence intervals. In 2018, the American Civil Liberties Union (ACLU) made headlines when they discovered that Amazon’s facial recognition technology incorrectly matched 28 members of Congress with the arrested. In their test, the American Civil Liberties Association entered photos of members of Congress and searched a database of 25,000 portraits of arrested individuals to see if the system would return any matching documents. In 28 instances (about 5% of all test members), Amazon returned a match. The American Civil Liberties Union believes that these results show that facial recognition is not accurate enough to be deployed, and there is no serious risk of abuse due to incorrect matches.

However, this accuracy can only be achieved under ideal conditions where the lighting and positioning are consistent and the subject's facial features are clear and unobstructed. In real-world deployments, the accuracy rate is often much lower. For example, FRVT found that the error rate of a leading algorithm climbed from 0.1% when matching high-quality facial photos to 9.3% when matching personal photos taken "in the wild", and when shooting with the subject in the "field" The error rate is different. It may be blocked by objects or shadows on the camera.  Aging is another factor that may seriously affect the error rate, because over time, changes in the subject's face will make it difficult to match photos that are many years apart.

Related News:

Facial Recognition Is Everywhere. Here’s What We Can Do About It

Face recognition visitor machine makes front desk visitors smarter

The future development of face recognition will break through the scope of security applications

website: www.hfteco.com

website: www.china-attendance.com

Email: info@hfcctv.com

HFSecurity Face Recognition DeviceFacial Recognition DeviceFacial Recognition DeviceFacial Recognition Device