MEET
THE ALGORITHM

CERTIFIED
TOP PERFORMER
1ST PLACE
VERIFICATION ACCURACY
IDENTIFICATION SPEED
THE III BEST PERFORMANCE
ON THE DETECTION
OF PEDESTRIANS AND CYCLIST
THE 2ND BEST ACCURACY
PERFORMANCE ON DETECTION
ACTIVITY IN EXTENDED VIDEO

HOW FACIAL
IDENTIFICATION WORKS

How neural network works

The neural network is trained to identify unique facial features so that it can then find similar faces in the database. NtechLab’s face detection algorithm works with global facial databases, allowing for a split-second search.

Recognition accuracy FNMR=0.008 @FMR<10-6
250 millionimages under 0.2 seconds
500 millionimages under 0.3 seconds
1 billionimages under 0.5 seconds
Recognition speed<1 second
icon

When working with facial features it is impossible to restore the original image of the face, which ensures compliance with the rules of personal data protection.

icon

Working with facial features requires up to 100 times less computing power compared to face image processing.

RECOGNITION MECHANISM

01Detecting the face and silhouette in the image
02Visual distortion correction
03Retrieval of facial features
04Verification or identification of a face

The algorithm analyzes
an individual video frame

A video consists of frames. Each frame is nothing more than a collection of pixels.
Ресурс 5

Pixel color
determination

Each pixel has a unique color code. The color code in the RGB palette is represented as three numerical values.
icon

The neural network receives a matrix of RGB values of pixels as input.

33333333

The algorithm detects faces

icon

The algorithm determines where faces are on the image.

NtechLab’s face detection algorithm is capable of detecting an unlimited number of faces in a frame, making it an ideal solution for security in crowded areas.

The speed of the detector does not depend on the number of faces in the frame.

faces-border faces-border-mb-fullEn
icon

The algorithm of facial recognition technology provides bounding box coordinates: the top left and right bottom borders of a face for further work with each one of them.

Visual distortion
correction

A dedicated algorithm is able to determine the position of the head and correct visual distortions: for example, «turn» the face to a full-face position.
NtechLab’s face recognition technology works under difficult conditions and effectively displays faces in an image or video even in a significant lack of light, disregarding altered position or head turns and tilts.
faces-normal1
icon

The image goes through the following stages: points on the eyes, corners of the mouth, nose — 5 points total.

icon

Algorithm retrieves
facial characteristics

The network finds and assigns to each person a vector of features or, in other words, a biometric face pattern.
icon

A biometric face pattern is a certain sequence of numbers formed by a neural network as a result of the initial image transformation and applied for comparison with other patterns.

NtechLab’s algorithm took the 1st place in the IARPA face recognition accuracy competition

Face detection and identification
against the image database

Identification is the comparison of facial features with those in the database. Proven by independent tests, NtechLab’s algorithm is number one automatic face recognition software in solving the identification problem. It finds faces, even if there are significant age-related changes, a beard or mustache, glasses, or any other means of partial face concealment. This ensures maximum effectiveness of facial recognition for law enforcement organizations
94
98%MATCH PROBABILITY

FACE ATTRIBUTES

icon

NtechLab’s algorithm works successfully with the following attributes:

GENDER
AGE
BEARD
GLASSES
EMOTIONS
MASK
The features of the face recognized by the system are called attributes. Such attributes can be used for a quick search in monitoring databases, for example — to find all faces with glasses.
Each of the attributes is defined by a separate neural network, with all involved networks working simultaneously.
87
87%BEARD
92%GLASSES
main-face-attr-2En1
87%BEARD
92%GLASSES
The features of the face recognized by the system are called attributes. Such attributes can be used for a quick search in monitoring databases, for example — to find all faces with glasses.
Each of the attributes is defined by a separate neural network, with all involved networks working simultaneously.

LIVENESS

Allows distinguishing a real person in front of the camera from a photo or video.
NtechLab uses a proprietary Liveness detection technology in its products. The technology is based on a passive detection method, which is ideal for access control and authorization tasks in mobile applications.
72
icon

Does not require additional actions from users (like smiling, blinking or interacting with the identification system otherwise)

Works with existing equipment (no need to install specialized hardware)

WORKS WITH VIDEO AND PHOTO
SPOOFING DETECTION ACCURACY 99,9%
VERIFICATION TAKES LESS THAN 1 SECOND
AVAILABLE THROUGH API
live-mb-linesEn
Allows distinguishing a real person in front of the camera from a photo or video.
NtechLab uses a proprietary Liveness detection technology in its products. The technology is based on a passive detection method, which is ideal for access control and authorization tasks in mobile applications.

SILHOUETTE
DETECTION
AND RECOGNITION

NtechLab’s algorithm detects the silhouette and path of the person passing by the camera.
The main problems effectively solved by the silhouette recognition software are an instant and accurate calculation of the massive number of people in a video stream, as well as inter-camera silhouette tracking.
02
icon

Can be used as an add-on to a facial recognition system, or as an independent module that performs silhouette recognition only

faces-silhouette-mdEn
NtechLab’s algorithm detects the silhouette and path of the person passing by the camera.
The main problems effectively solved by the silhouette recognition software are an instant and accurate calculation of the massive number of people in a video stream, as well as inter-camera silhouette tracking.

HOW NTECHLAB ALGORITHM WORKS

about-ii-bannerEn
NtechLab’s algorithm uses multiple neural networks for face image search and identification. One network detects the face in a photo or video stream, the other extracts a biometric pattern, while others work with attributes (gender, age, glasses, beard and others).

Global acknowledgment

NtechLab’s neural networks have a unique architecture developed by our own specialists.

They have been trained on large data sets of photos that are very close to real-life images in terms of shooting conditions, lighting and position in the frame, which ensures highly advanced computer vision.

The success of the neural network is determined by:

NEURAL NETWORK ARCHITECTURE

DATASETS THAT THE NETWORK HAS BEEN TRAINED ON

NEURAL NETWORK TRAINING ALGORITHM

PAD (Presentation Attack Detection) testing at the reputable iBeta Quality Assurance laboratory verifies the ability of face recognition algorithms to establish whether the face on video belongs to a live person or whether it’s an image or a mask.

«The NtechLab algorithm passed the PAD Level 2 testing and proved full compliance with the ISO/IEC 30107−1 and ISO/IEC 30107−3 standards»

Testing conducted by the National Institute of Standards and Technology of the USA

The NtechLab algorithm was recognized the world’s best based on the results of 7 independent tests

Within three tests the algorithm has beaten a record in the entire history of testing

International competition in detecting deepfake techniques applied to the video

«3rd place in deepfake detection»

International video activity recognition competition

«2nd place in video activity recognition.»

A competition in pedestrian and cyclist detection on the road

«3rd place in recognition of pedestrians and cyclists through city surveillance cameras.»

Testing conducted by the National Institute of Standards and Technology of the USA

«The best recognition result from the databases of photos taken under uncontrolled conditions»

A competition among algorithm developers for emotion recognition, organized by the University of Ohio, USA

«Best emotion recognition results.»

International competition organized by the United States Intelligence Advanced Research Projects Activity

«Best verification accuracy»

«Best verification speed»

International Face Recognition Algorithm Evaluation Competition, hosted by the University of Washington, USA

«Best face recognition accuracy»

Top position among facial recognition algorithm developers, including Google

DOWNLOAD THE QUALITY
ASSESSMENT MANUAL