What Is Face Recognition? Understanding the Basics – Mechanisms, Types, and Accuracy ~ “Back to Basics” Series Vol. 1 ~
- erikaitagaki
- 6 days ago
- 6 min read
Updated: 2 hours ago
In recent years, face recognition technology has been increasingly used in various everyday situations, such as unlocking smartphones or managing access control in offices.
In this article, we explain the basic mechanism, different types, and accuracy of face recognition in the simplest terms possible.
1.How Face Recognition Works
Face recognition is a technology that determines the statistical likelihood that a person is who they claim to be by extracting facial features (quantified numerical representations of facial characteristics) from an image of a face, then comparing these features with others to calculate their similarity.
General Flow of Face Recognition
・Face Detection
Identify and locate the “face” area within an image.
・Image Alignment
Rotate and adjust the face image to make it easier for the recognition process.
・Feature Extraction
From the face image, the system identifies numerous landmarks such as the eyes, nose, mouth, and facial contour.
The positional relationships between these points are calculated vectorially and converted into an array of numerical values—this is called the feature vector.
For security and data handling purposes, the feature vector is usually encrypted or transformed into a string of alphanumeric characters before being stored.
The process of generating these landmarks and feature vectors is performed by a face recognition model, which is the core of the technology.
Each model has its own characteristics and strengths, and even with the same image, the extracted features can differ depending on the model used.
・Matching
The extracted feature vectors are compared vectorially to determine the distance (similarity) between them.
If the calculated similarity exceeds a pre-defined threshold* and is the highest among compared features, the system determines that the two images belong to the same person.
*Threshold: A preset value below which the system will determine that the images do not belong to the same person.
2.Types of Face Recognition
Face recognition can be implemented in various ways depending on its purpose and system configuration.
Here, we explain two common distinctions: 1:1 vs. 1:N recognition and edge-based vs. server-based recognition.
1:1 Recognition and 1:N Recognition
・1:1 Recognition
This is a method of authentication where one registered individual is compared with one verification subject.It is used for purposes such as smartphone login, airport face recognition gates, and identity verification (eKYC).For example, the system compares the feature vector extracted from the face photo on an ID with that extracted from the person presenting it, determining whether they are the same person.Since the comparison is limited to a single target, even relatively compact face recognition models can achieve high-speed, high-accuracy authentication.
・1:N Recognition
This method searches for one individual from among many registered images.
It is used in scenarios such as office entrance/exit control systems, membership verification or reception at stores and facilities, and blacklist-based suspicious person detection.
For users, it offers the convenience of complete hands-free authentication without the need for cards or IDs.However, the larger the number of registered individuals, the more time-consuming the matching process becomes, and the more critical accuracy tuning becomes to meet the required precision level.
Edge-Based Recognition and Server-Based Recognition
In face recognition system architectures, where the authentication is performed generally falls into two categories: edge-based recognition (device-side) and server-based recognition.
・Edge-Based Recognition
In this method, the authentication process is completed locally on the device (edge side), such as a tablet with an integrated camera.
Because the authentication does not rely on a network, it can be performed quickly and stably on-site.
However, edge devices are often preferred to be compact and lightweight, which limits their computational power and storage capacity.
This makes it difficult to run large-scale face recognition models, and the number of people that can be authenticated by a single device depends on its hardware specifications.
・Server-Based Recognition
In this method, the authentication process is performed on the server side. Facial data captured by a camera-equipped device is transmitted via a network and matched by software running on the server.
The server can be deployed either in the cloud or in an on-premises environment.
Server-based recognition allows the use of larger-scale face recognition models, enabling even higher accuracy.
By upgrading the server specifications, it can handle a greater number of registered individuals, offering high scalability and making it suitable for medium- to large-scale systems.
However, it is more susceptible to delays and communication issues depending on the network environment.
3.What Is a Threshold?
In a face recognition system, the threshold is a value used to determine whether a person is the same individual.
If the similarity score is lower than the set threshold, the system judges that the person is not the same.
The threshold setting directly affects both security and convenience.
Therefore, in face recognition systems where the threshold can be adjusted, it is extremely important to set it appropriately based on the intended use case.
Both security and convenience are important factors.
If you wish to prioritize security, you can set the threshold higher.
If you wish to prioritize convenience, you may slightly lower the threshold setting.
4.What is the Accuracy of Facial Recognition
The accuracy of facial recognition is improving rapidly, and there are now many systems with extremely high accuracy, such as 99.9% recognition rates.
However, to give an easily understandable extreme example, if a single authentication test succeeds once, it could be considered 100% accurate.
In reality, there are multiple important metrics for measuring facial recognition accuracy. A model is not considered good simply because one metric is outstanding; all metrics must be balanced at a high level.
To know the true accuracy, it is crucial to understand what each metric represents and how the other metrics perform. Here are the key indicators used to evaluate facial recognition accuracy.
Key Indicators of Facial Recognition
・False Acceptance Rate (FAR)
This metric indicates the probability that the system incorrectly recognizes an unauthorized user as the correct person and grants access.
The lower the FAR, the better, as this means higher security.
Raising the threshold reduces the FAR.
・False Rejection Rate (FRR)
This metric indicates the probability that the system incorrectly rejects the legitimate user as an unauthorized person.
The lower the FRR, the better, as this means higher convenience.
Lowering the threshold reduces the FRR.
By nature, facial recognition systems exhibit a trade-off between FAR and FRR: raising the threshold lowers FAR but increases FRR, while lowering the threshold decreases FRR but increases FAR.
However, a high FAR renders the system ineffective as a security measure, and a high FRR makes it impractical in real-world use. Since it is important to balance security and convenience, a good facial recognition model is one that maintains a very low FAR at a high threshold setting without increasing FRR.
5.Features of Nabla Works’ Facial Recognition Technology
Nabla Works’ facial recognition technology is characterized by its ability to achieve high accuracy even under challenging conditions.
While it is naturally capable of accurately recognizing faces in ideal situations—captured from the front, with clearly visible eyes, nose, and mouth, and high image quality, like a passport photo—real-world conditions often present various challenges for facial recognition.
Examples of challenging conditions for facial recognition:
In Nabla Works’ facial recognition system, when registering a face, the system estimates unseen parts of the face from a single photo taken at a certain angle and converts them into feature data.
Of course, it is not necessary to register multiple photos from every angle.
Registering just a single frontal photo is enough.
This allows the facial recognition system to be used conveniently and reliably even under challenging conditions.
6.Nabla Works’ Facial Recognition Solutions
At Nabla Works, we offer a wide range of facial recognition solutions, from facial recognition-based entry/exit management and reception systems to various software and cloud services that allow you to integrate facial recognition into your own systems.
Experience firsthand the convenience and high accuracy of our facial recognition solutions!
