Why were you mistaken for the police facial recognition system?
According to a German media, someone who participated in an anti-war and peace-calling parade is astonishingly being listed in the extremist database after his face is recognized by the police camera. The existing cases have given us more alarms to us. In London, where the facial recognition technology was adopted for providing assistance to the public security, 35 people were mistakenly arrested in just a day.
The idea that our information might be wrongly used by the state administrative departments really worries and frightens us, for we have no idea what we and Property, reputation, our job, freedom, health or even our lives, everything is possible.
Why is there such a terrible misunderstanding about facial recognition systems for rigorous law enforcement agencies, or how do you judge whether a facial recognition product or solution is good or bad?
Generally speaking, face recognition algorithms have the following core concepts, which are the core indicators to evaluate whether a face recognition algorithm is good or bad.
1) False recognition rate
As the name suggests, it is the probability of false recognition, which is also called False Accept Rate (FAR), referring to the fact that the face comparison is passed but not in person, and the “fake” is recognized as “real”, such as mistakenly taking A It is the same as people sometimes misidentify people.
2) Pass rate
The The pass rate is easy to understand, it is the probability of passing the test with correct recognition.
In contrast to the pass rate, there is another concept called the rejection rate, which = 1 – pass rate, abbreviated as FRR (False Reject Rate). The rejection rate means that although it is you, it just doesn’t recognize you, just like a person who obviously should know you, but he just doesn’t recognize you and says you are not you.
3) Similarity
When When comparing faces, the ratio of how similar the captured face is to the face in the target library can generally be expressed as a percentage, for example, the similarity between the captured face and the face in the target library is 80%, just like when you see someone, he seems to be 80% percent similar to someone in your memory.
4) Threshold Threshold
The concepts of false recognition rate, pass rate, rejection rate, similarity, etc. The above must be combined with the concept of “threshold” in order to have a practical application scenario.
Take the scenario of human eyes seeing someone as an example.
1. you see a person on the street, he seems to be 80% years similar to someone you remember, this 80% percent similarity is the degree of similarity.
2. at this point you follow your usual experience, if there is 70% similarity or more, you decide that this person is the one you know, this 70% years similarity, which is the threshold in your brain.
3. because the person you see has 80% percent similarity, which exceeds the threshold in your brain, so you pass your judgment, which is passed.
4. On the contrary, if the threshold in your brain has to reach 90% percent similarity before you judge that the person is the one you know, then you will think that the person you just saw on the street is not him, there are two cases, 1) he is really not him, then your judgment is right; 2) he is really him, then you are wrong, at this point you reject the truth.
5. at this time, if you find another person on the street is super similar to the one you used to know, reaching 99% people (more than 80% percent of the threshold value in your brain), you will decide that this person is the one you used to know, but go over to say hello, people say you have mistaken the person, at this time it is called misrecognition.
Face recognition is actually simulating the process of human eye recognition, which is almost the same logic as our humans. But the threshold value mentioned here is too critical, the threshold value is high or low, it directly affects the “recognition rate”, “through rate” performance.
First, the false recognition rate, pass rate and threshold value has a correlation, the higher the threshold value, the lower the pass rate and false recognition rate, the lower the threshold value, the higher the pass rate and false recognition rate. This leads to a common misunderstanding, there is a statement called “the higher the pass rate the better”, the initial intention is good, the higher the pass rate the better the customer experiences, but this statement makes two mistakes, on the one hand, the pass rate and false recognition rate should have the threshold value as a premise, simply talking about the pass rate and false recognition rate is not significant; on the other hand, the false recognition rate and pass rate On the other hand, the false recognition rate and the pass rate are mutually constraining, and we cannot pursue the beautiful unilateral data. The other side of the high pass rate is the high false recognition rate, and the risk prevention ability will be reduced, so we should consider the two factors of customer experience and risk prevention ability to determine the threshold value and the corresponding false recognition rate and pass rate. A more accurate statement should be that under the specified threshold, the false recognition rate of company A is lower than other companies and the pass rate is higher than other companies, which means the face recognition algorithm of company A is good.
Secondly, the false recognition rate, pass rate and threshold value are non-linear relationships, and the pass rate and false recognition rate drop sharply after the threshold value is consistently increased. This shows that the threshold value cannot be set too high, but the threshold value will definitely be false recognition as long as the amount is large enough.
So if you see a so-and-so company promote their face recognition algorithm has more than 99.9% percent recognition rate, then you should not easily think that their face recognition algorithm must be good, because whether it is good or not, on this one index, is not enough, he must combine with false recognition rate, pass rate, threshold value, etc., in order to determine more objectively.
The above talked about the terms pertaining to face recognition and how we should determine it, but there is a special phenomenon in the field of face recognition, that is, face recognition does not seem to work well in black and white scenes, why is it so? We will cover this in the next blog.
Post time: Jul-19-2021