Home → Startups → Face Recognition
The task was reliable protection of office premises of a mid-sized IT company

Face Recognition (AI-PASS)

Key Points





Type of Project

AI & Machine Learning


6 Months of Realization

Expert Team

Business Analyst, QA Engineer, 2 Developers

Non-disclosure agreement

The Main Goal is The Safety of Access to The Premises

Our task was reliable protection of office premises of a mid-sized IT company. Facial recognition systems have become a kind of natural choice for us. Our first internal project aimed at developing a facial recognition system was implemented by providing access to our company premises. We used software technologies and our office became the main platform for experiments.

Non-disclosure agreement

Introduction to Face Recognition Technology

Our ultimate goal was to develop a face recognition system that not only accurately identified the face and gave an alarm if the face was not recognized, but was also integrated with the card system, blocking access to the company’s premises under certain circumstances.

Long-term security (a well-designed face-recognition application turns out to be at least 10% more “observant” than a human security guard, who can easily get distracted or become less alert from fatigue).

As a result, we concluded that the following combination of tools and methods will be more optimal for achieving the goals:

• HOG (Histogram of Oriented Gradients) – preferably for face detection.

• Dlib NN – optimal for getting a face descriptor (unique face identifier, consisting of 128 facial features or dots).

• SVM – Preferred for comparing the received visual information with the face descriptors stored in the application database.

Important Notices

1.  A face recognition application should recognize not only faces but also objects. This would be especially true for large corporations with a very large number of employees or for those enterprises that have increased safety requirements due to the nature and specifics of their production process. For example, someone carrying a rifle will not be welcomed at the airport. A face recognition application can be trained to identify such objects. Perhaps it would be convenient to use it to prevent theft at production facilities.

2.  Most often, a person’s head does not remain motionless while walking. The quality with which a face recognition application can identify a face depends on the distance of the camera from that face, the position of that face, and its movements. Several tricks can help solve the face distance problem:

Positioning the cameras of the face recognition solution so that their resolution is optimal for the distance to the future objects.

Installing enough light sources in the right places.

Using a face recognition application to enhance the resulting image to match a specific pattern suitable for recognition purposes.

Using a face recognition application in such a way that it takes into account how significantly you needed to adjust the incoming image. For example, if, as required by the recognition algorithm, a 30×30 face image needs to be enlarged to 150×150, the application should give that image a lower rating.

3.  The angle of the face is also very important and is often quite difficult to deal with. Ideally, 128 different facial features or points are required to immediately identify a human face. The further from the front view the angle of rotation of that face, the more difficult it is for your face recognition application to correctly identify the person.

We have done a lot of research, experimenting with different angles and images of different parts of the face, and we have found an effective approach to the face recognition process:

We take a full set of images from different angles taken by the camera, determine how close they are to the angle of the front face (we calculate the angle of rotation for each of the images), and measure the distance to the object for those images that can be considered facial.

Images that have a better ratio of angles to front-to-front spacing take precedence over other images. However, the rest of the images are also used in the recognition process.

4. Every time a new person is added, a new cluster containing at least 50 images (including both front and left and right rotated images) must be added to the face recognition application database. The SVM must then be retrained to adapt to this information and be able to use it.

Write to Us

Describe your project, briefly the goals that you need to achieve and the tasks that need to be solved.

Enjoy our insightful articles and step-by-step implementation guides to ensure your market success

Fideware Related Blog Posts