Face recognition is a method of identity recognition, the purpose of which is to determine what the identity of a face is in pictures and videos.
This article will introduce in detail the four characteristics, four steps, five difficult points and the development trajectory of face recognition.
What is face recognition?
Face recognition is computer vision-computer vision Typical applications in the field.
The purpose of face recognition is to determine the identity of the face in pictures and videos (videos are composed of pictures).
Face recognition is a type of identity recognition, which is similar to ID card identification, fingerprint recognition, and iris recognition. You can compare face recognition with the familiar ID cards:
- Input information (open ID card-input face information)
- Store the information in the database (ID information-face information)
- When you need to verify your identity, compare the newly collected information with database information (scan ID card-swipe your face)
Face recognition has 4 characteristics
Compared with other identity recognition, face recognition has four characteristics:
- Convenience. Faces are biometric and do n’t need to carry something like an ID card
- Not mandatory. The recognition process does not even require the cooperation of the object, as long as the face is captured, it can be recognized, for example, in the field of security.
- Non-contact. There is no need to contact the device, which is more secure than a fingerprint.
- Parallel processing. When there are multiple faces in a photo, they can be processed together. Unlike fingerprints and irises, one by one is required.
Based on the above characteristics, face recognition is being widely used in various fields. Everyone can see the application of face recognition everywhere in life.
4 steps for face recognition
There are 4 key steps in the face recognition process:
- Face Detection
- Face alignment
- Face encoding
- Face matching
The four steps are explained in detail below.
The purpose of face detection is to find the position of the face in the picture. When a face is found in a picture, no matter who the face is, the coordinate information of the face will be marked or the face will be cut out.
Directional Gradient Histogram (HOG) can be used to detect the face position. First gray the picture, then calculate the gradient of the pixels in the image. By converting the image into a HOG form, the face position can be obtained.
Face alignment is to align face images of different angles into the same standard shape.
First locate the feature points on the face, and then use geometric transformation (affine, rotation, scaling) to align the feature points (displace the eyes, mouth, etc. to the same position).
The pixel values of the face image are converted into compact and discriminable feature vectors, which is also called a template. Ideally, all faces of the same subject should be mapped to similar feature vectors.
In the face matching building module, the two templates are compared to get a similarity score, which gives the possibility that both belong to the same subject.
5 Difficulties in Face Recognition
The presentation of face images in the real world is highly variable. So face recognition is also one of the most challenging biometrics methods. Where face images are variable include:
- Head pose
- Lighting conditions
- Facial expressions
Development track of face recognition algorithm
The face recognition field has also transitioned from traditional machine learning algorithms to deep learning algorithms.
Traditional machine learning algorithms
During the machine learning phase, face recognition also went through three important stages:
- Geometric feature stage
- Representational Stage
- Texture feature stage
Deep learning algorithms
In the deep learning phase, the development of the algorithm also went through three stages:
- From the initial VGG network to the Inception network to the Resnet network, the network model generally shows a deeper and wider trend.
- These vendors, such as Despise and Shang Tang, who have achieved good results in academic open competitions, began to develop actual business as a starting point. By continuously expanding their actual data collection, the algorithm performance is gradually improving.
- In addition to further increasing the amount of data to improve the performance of the algorithm, contrary to the first stage, everyone began to study the lightweight of the network without reducing the recognition performance. The main purpose of lightweighting is two, one is to improve the speed of the algorithm, and even can be deployed to the mobile end; the other is to facilitate hardware implementation, so that the face recognition algorithm can be directly made into a hardware module.
To learn more about the technical details of the different stages, you can read this article "Understand the development of face recognition technology in one article"
Typical applications of face recognition
The application of face recognition is becoming more and more widespread. As long as it is related to identity recognition, it is possible to use face recognition. A few typical application scenarios are listed below.
- Access control system
- security system
- Unmanned supermarket
- EPassport and ID
- Autonomous service system (such as ATM)
- Information security systems, such as face payments
- Entertainment applications, such as some props in Douyin