Technical scheme of face recognition system

1. brief introduction

With the rapid development of artificial intelligence technology , Now many residential areas or office buildings have used face recognition intelligent access control systems . Nowadays, many places do not need to swipe cards to identify when entering and leaving office buildings , You can enter the building directly by brushing your face . New biometrics replace the traditional recognition methods

Face recognition task , It is divided into two parts , Face feature extraction and face verification

Face feature extraction

Face feature extraction needs to judge based on the image features of the face , This characteristic belongs to that person , How to extract facial image features , At present, there are two ways ：

• Aggregate high-dimensional abstract features based on face pixels （ The skin , Facial features are similar to the form of image classification ）
• Features based on face key points , In fact, it is equivalent to the previous method , Added face keys （ Because the position of the key points of the face , It is based on high-dimensional Abstract facial features ）

Features based on face key points

The following figure shows a detailed face recognition algorithm process based on face key points

Key point detection methods can be divided into two types :

• One is to use coordinate regression to solve
• One is to model the key points into a thermodynamic diagram , By pixel classification task , The location of key points can be obtained by regressing the distribution of thermal map .

These two methods , Is a means or a way , The problem to be solved is to find out the position and relationship of this point in the image , In the vernacular, it is to find a pixel location , And the context around it （ Surrounding pixels ） There is a certain combination relationship

Face verification

Face verification is whether the face in the current photo is someone who already exists in the database , There are generally two ways ：

• Direct classification , Distinguish which person is accurate , Output its label （ Image classification ）
• It turns into a binary classification problem , That is to distinguish whether the face in the picture pair composed of two face photos comes from the same person , The output is the confidence of the same person

The first method has many disadvantages , Such as ： When the model training is completed, you can't join new people at any time , And people in each database need to collect more face data , And the accuracy of image classification is not high

The second method is generally realized by twin network （Siamese Network） Realization , The general structure is as follows ：

principle ：

• By the same CNN The network encodes the face image in the same way , Embed a high-dimensional vector space
• Use softmax loss As a loss function, do binary classification training directly on the splicing of two sample embedded vectors , So that the model can directly output the similarity between two samples , When the similarity reaches a certain threshold, it is judged as the face of the same person
• Or use triplet loss、contrastive loss、center loss Equal loss function pairs CNN Network optimization , Make it finally encoded in the high-dimensional vector space , The distance between similar samples is reduced , The distance between different types of samples increases
• Then the similarity between the two samples is measured by the distance between the embedded vectors of the two samples , Similarly, when the similarity reaches a certain threshold, it is judged as the face of the same person

2. Technical solution

The face recognition system uses distributed cluster technology , Based on neural network deep learning algorithm and massive data storage big data computing technology , Realize video surveillance image , Face recognition image

The front end adopts video stream or picture stream for video image transmission , Provide face images of the scene environment , And form a face capture database .

Face recognition service mainly has two modes , There are two comparison modes: verification mode and search mode .

• Verification mode （$1:1$） It is to verify whether the collected image or the specified image is compared with the object registered in the database , To determine whether it is the same person .$1:1$ For authentication mode , By comparing the face features of someone's equipment collection photo and certificate photo , Check whether it is the same person , This mode is mainly applied to scenarios that need to be verified by the real name system .（ It's common to pass the security check by plane ）
• search mode （$1:n$） It means searching all images registered in the database , To find out whether the specified image exists .$1:n$ By collecting someone's portrait , Find the image that is consistent with the current user's face data from the massive portrait database , Identify the identity of the other party through database comparison .（ Community access control ）
  

The second one $1:n$ The face database of is divided into three business databases ：

• Face capture library ： Contains historical snapshot live pictures 、 Small face images and structured face feature data 、 Capture location 、 Capture time and other information , The main business application of this library is image retrieval and comparison , Query the location of the portrait of the target person 、 Time 、 Track tracking, etc .
• Face registry ： Mainly import some large-scale portrait pictures 、 Structured face feature data and identity information , Such as the local social security portrait information database 、 Urban population information database , The main application scenarios after import are image retrieval and comparison and identity information query , Confirm the identity of the person .
• Blacklist database ： Including high-risk personnel 、 Face pictures of special personnel , The main application scenario is the face comparison and early warning of real-time pedestrian flow at each checkpoint .

$N:N$ There are few scenes , In fact, it is equivalent to carrying out multiple 1：N distinguish , be used for “ Prove who is who ”.

3. Face recognition technology expansion

3.1 Face data modeling and retrieval

Face data modeling and retrieval can model the face image data registered in the database to extract the features of the face , And the generated face template can be saved in the database . In face search , Model the specified face image , Then compare it with the template of the owner in the database , Finally, list the list of people with similar degree according to the compared similarity value .

3.2 Dynamic vivisection

Previously, static face recognition was carried out through a specified area or range , That is to say, identify the right distance 、 The location requirements will be relatively high . Static face recognition is characterized by small user capacity , And the safety performance is not high , Sometimes a photo can also be recognized and verified . Now the introduction of dynamic face recognition access control , The system can identify whether the other party is a real person or a photo .

Cooperate with in vivo detection

In the bank app, Common applications , The system determines whether the user is living by prompting the user to complete some actions （ Like blinking , open one 's mouth , Shake your head ）.

3.3 Image quality detection

Image quality directly affects the recognition effect . The image quality detection function can evaluate the image quality of photos , And give the corresponding recommended value to assist in identification .

In the future, more and more cities will become intelligent , Technology oriented products will make citizens' lives more comfortable , Improve the quality of life , Save natural resources .