How do facial recognition biometrics work?

Facial recognition biometrics is an advanced form of image recognition based on artificial intelligence (AI) and machine learning (ML). It can be used on still or video images.

This article begins by discussing the basics of facial recognition, including facial detection and feature extraction using convolutional neural networks (CNNs). It then moves on to consider two-dimensional (2D) and three-dimensional (3D) facial recognition and concludes by examining how facial recognition can work in the dark.

Facial recognition uses an algorithm that begins by finding the eyes to identify features. Additional facial characteristics like eyebrows, mouth, nose, nostrils, and irises are identified from there. Basic facial recognition algorithms typically look for the following (Figure 1):

• Distance between the eyes
• Depth of the eye sockets
• Distances from the forehead to the chin and from the noise to the mouth
• Contours and shapes of the cheekbones, lips, ears, chin, and other features

Face detection

Before facial features can be extracted, a face must be identified. That begins with training an AI with a large dataset. There are several face identification techniques, including:

• Feature-based
• Knowledge- or rule-based
• Template-matching
• Appearance-based

Once a face has been identified, face normalization can speed up the feature extraction process. Normalization adjusts a detected face to remove variations like different poses, lighting, and backgrounds to make it consistent with other faces in the database.

Feature extraction

There are also several ways to extract facial features. Most common methods use a form of CNN to process pixel-level data. In facial recognition applications, a region-based CNN (R-CNN) is often used as the classifier. There are four steps in R-CNN operation (Figure 2):

1. Identify potential object regions of interest (ROIs) in an input image. These regions can range into the thousands for complex images and are rectangles that may represent the boundary of an object.
2. Use a CNN to extract features from each ROI.
3. Support-vector machine (SVM) classifiers determine what type of object is in each ROI.
4. Non-maximum suppression (NMS) eliminates duplicate or overlapping boundary boxes for each ROI, leaving the most probable result.

2D vs 3D and facial recognition

2D facial recognition systems use standard digital images. This was the original method and is still the most common form of facial recognition. Newer 3D facial recognition systems use stereoscopic images to create a 3D model of a face. 3D recognition can overcome lighting, expression, and pose variations and produce a more reliable result. But it’s not as convenient and is more costly.

There are more cost-effective approaches to 3D face recognition, such as using multiple 2D cameras and constructing a 3D representation based on several 2D images.

Alternatively, deep-learning construction of a 3D image uses a grid of light projected onto a face. The reflected light pattern can create an accurate 3D facial image. Some of these systems combine a visible light image with an infrared (IR) camera to accurately map facial features.

Facial recognition in the dark

Thermal imaging using IR cameras instead of visible light imaging is used for facial recognition. Instead of using reflected light, the IR camera produces an image, a thermogram, of the heat radiation emitted from the face.

The thermogram provides data about the face’s physiological characteristics, which can be used similarly to a visible light image for face recognition. This approach has the major advantage of producing highly reliable results in low-light conditions.

Summary

Facial recognition is a well-established and reliable biometric identification technology. A range of implementations are optimized for specific application needs and varying resource availability. In addition, IR radiation emitted by a face can be used for facial recognition under low light conditions without relying on a visible light image.

References

3D face recognition: A comprehensive survey, SpringerLink
Computer Vision, Dive into Deep Learning
Facial recognition and fundamental rights 101, European Digital Rights
Reading Your Face: How Does Facial Recognition Work?, Aratek
The future of biometrics technology: from face recognition to related applications, Cambridge University Press
Thermal Image Generation for Robust Face Recognition, MDPI applied sciences
What is face detection?, TechTarget

Related WTWH content

How do sensors play into AI’s facial recognition capabilities? Part 1
New Model For Large-Scale 3-D Facial Recognition
If you are working with video signal processing here are some tools to consider: Part 2
How do sensors impact ATMs?
Semi-custom SoC design excels at high-performance data processing