Deep Learning Face Recognition Application

Advances in deep learning technology have led to explosive growth in face recognition applications. At present, the application scenarios of face recognition can be divided into face authentication and face recognition.

1. Face Authentication: That is to prove that you are you

A typical application is a 1:1 human-certification-one system, which compares the user's face collected on the spot with the face photo corresponding to the certificate to verify whether it is a legal holder. Take the most commonly used second-generation ID card photos (photos only 102 * 126 pixels, compressed into 1K bytes of storage, many people have a large age span) vs users with the scene as an example, Dahua adopts deep learning The face recognition algorithm has a pass rate of more than 98% when the false alarm rate is less than one in 100,000. The accuracy of deep learning algorithm in face authentication has surpassed the average human recognition level. Therefore, it has been widely used in scenarios such as remote opening of accounts by financial institutions, high-speed rail airport security, and entrance to the examination room.

2. Face identification: It can be divided into static comparison and dynamic comparison

Static comparison, usually closed set, means that the test person exists in the registration database, and the previous 1 or previous N recognition accuracy rate is used as the evaluation index. The typical application is (identity of the suspect/lost population, etc.) identity confirmation - the high-resolution photographs collected under the conditions of the use of the test person are searched in large-scale registration libraries (escaped fugitives, permanent populations, etc.). At present, on the scale of 10 million registered databases, Dahua's deep learning algorithm has achieved a recognition rate of over 90%, and many successful cases have been achieved. In addition to the improvement of the accuracy rate, another major advantage of deep learning is that the features learned are very concise (usually a few hundred dimensions, and can be further quantified and compressed on the premise of maintaining the accuracy rate), and large-scale face retrieval The speed has also been greatly improved, and the tens of millions of bottom libraries can generally return search results in seconds or even less.

Dynamic comparison, usually open set, that does not confirm whether the tester exists in the target library, needs to ensure the accuracy rate in the case of ensuring that the false alarm rate is not higher than a certain value. The performance of the algorithm is related to the application scenario (whether the tester cooperates with the size of the base library). In the case of a base library of 10,000 people and a false alarm rate of no more than 1%, the accuracy of the user's general deep learning algorithm in the scene is more than 98%, which can basically meet the needs of applications such as access control and attendance. For non-matching scenes, the recognition rate will be greatly affected, generally below 90%. For example, the size of the public security domain registration database can reach 100,000. There are often unfavorable factors such as the testers not cooperating with or even deliberately avoiding and monitoring the poor video quality. In the densely populated places, ensuring that false positives are sufficiently low, there are often more leaks. Newspaper.

In general, deep learning has greatly enhanced the face recognition effect and directly promoted the application of face recognition. Although the accuracy of deep learning algorithm has surpassed humans to some extent, there is still much room for improvement to meet more application requirements. The author believes that the current application should not only be based on the “recognition rate” theory, but should be based on the specific scenario to design reasonable products and solutions based on the thinking of system engineering, so as to maximize the existing technology. For example, even if the accuracy rate of human comparisons exceeds 99.99%, we still need to cooperate with other verification methods to ensure sufficient security in the payment and other scenarios. On the other hand, although the recognition rate of deep learning algorithms under some conditions is very low, There is a lot of use value, such as the identification of missing children, even if the accuracy rate is only 1%, it is worth trying. Similar scenarios such as suspects' control and pedestrians' red lights only require a small number of successful cases (the recognition rate may be very low), which can play a good role in shock and warning.