The process of machine recognition with the biometric feature face

22 September 2020

The face is used as the most widespread biometric feature of the human being. This ranges from interpersonal communication with each other, formal authentication with standardized tokens in identity cards, passports, or driving licenses to dynamic video surveillance in public spaces.

The attractiveness of the face as a biometric feature is based on the various possibilities of unambiguous identification of persons and the disclosure of characteristics such as gender, age, ethnic origin and emotional states.

The human face is limited by the frontal contour from the hairline to the chin. Within this contour, it includes the forehead as the largest part of the face, the mouth, the nose, the cheeks and the eyes. As the most important component of interaction with the outside world, the face includes the most important sensory organs for perceiving the world: sight, smell, taste and hearing.

Representation: Anthropological representation of facial features

The individuality of this biometric feature results from the different compositions of the individual parts of the face. Depending on interaction, mood and external influence, dynamic facial expressions are produced by the activities of individual muscle groups in relation to each other. This is particularly determined by the eyes, eyebrows, forehead wrinkles and mouth as the most mobile parts. In addition, certain similarities and thus affiliations to a group often emerge, for example in family members.
Adults can reliably recognize, remember and recognize faces. This superior perceptual ability is the result of an evolutionary process. Of course, this ability is also finding its way into technological development – the face as a biometric identifier with a high probability has revealed many advantages in recent developments in terms of integration into technological applications. Firstly, the face can be captured at a distance without direct sensor contact; for example, as surveillance cameras in security systems. Secondly, in addition to identity, the face also conveys the current emotional state (happiness, joy, fear, etc.) and therefore enjoys growing importance in the design of applications for interaction between human-machine interfaces.

Please accept statistics, marketing cookies to watch this video.
Video Source: Drowsiness detection by A.Kanasani

Furthermore, it can also be observed that people are more willing to share their face with the public than is the case with other biometric features such as iris or fingerprints. The explosion of digital images in social media and the consistent expansion and use of additional functionalities such as face tagging in social networks are increasingly promoting the commercial significance and the machine recognition possibilities of the biometric feature face.

Process of automatic face recognition

Face recognition deals with the recognition of an already localized known face, for example for access control to physical or virtual rooms. This in turn means that a face recognition process consists of three main core steps with technological modules to handle authentication:

  1. Image acquisition with a sensor module – a sensor records the presented face in the form of biometric data for storage as a biometric template in a database module. The database module then contains a number of similar data sets.
  2. Face detection (also known as face localization or face segmentation) locates a face within a static image or dynamic video and determines the spatial extent of the face within the medium.
  3. Face Recognition – here the features are extracted from the detected face and a matching algorithm compares these sample data with the templates stored in the database. It returns a match value or decides, depending on the threshold value set, whether a comparable data set is available.

Feature extraction and matching

In the overall process of digital face recognition, the procedures for feature extraction and the actual matching play a central role in recognition performance. The formats of the extracted features depend on the recognition methods used.

The extracted feature values are then quantified into discrete values to enable accurate recovery. This basis is then secured with one of the biometric encryption methods, e.g. Salting. Depending on this security procedure, the original data record is inverted, for example, during a comparison for authentication and identification, in this case using Salting.

A short overview of the recognition techniques in three categories of feature extraction is given here:

  • Appearance-based methods– the idea behind appearance-based methods is to represent a given face as a function of various facial images from a training set. For example, a pixel value at a particular location can be determined as a weighted sum of the pixel values of the location from the training images.
  • Model-based procedures should enable a pose invariant representation of an existing face in order to realize face recognition from different poses. For this purpose, prominent points in the face are determined, e.g. corner of the eyes, corner of the mouth, etc. These points are then connected in a representative graph to determine the distance and the respective point is weighted with texture information from the local environment. On this basis the matching is done by comparing the image graph of the sample with the model graph of the stored images.
  • Texture-based methods use the distribution of local pixel values as a basis for the creation of biometric templates. Depending on the procedure, key points with the closer neighborhood are described in matrices, for example, in order to achieve pose invariance.

Effects of the variations on face recognition

At the same time, the many possibilities of using the face as the public biometric feature is one of the greatest challenges. Variations in pose, facial expressions, exposure conditions, aging processes, changes in external appearance, changes in emotional state or family resemblances have a significant impact on the performance of machine face recognition.

Challenge in face recognition: intraclass and intra-user variations

The perfect match between two alphanumeric strings is a necessary prerequisite in password-based systems to verify user identity. Due to the variations described, biometric systems can only determine the identity of a sample based on probability using a match score.

This also leads to errors: false acceptance and false rejection influence the accuracy of the recognition performance. Statistically, these errors can be mathematically described proportionately as the number of failed attempts in the total number of attempts, explained as an example:

  • Failure to detect 5% of known persons out of 100 authentication attempts (genuine users) means a False Rejection Rate (FRR) of 5. This means that persons are incorrectly not detected although the system should have detected them.
  • Detecting 5% of unknown people out of 100 authentication attempts ( impostor users) means a False Acceptance Rate (FAR) of 5, which means that people are detected even though the machine should not have detected them.

The continuous development with methods of 3-D modelling, advanced modelling techniques and simulation methods, for example for aging processes, are increasingly reducing the susceptibility to errors due to intra-class and inter-user variations and are increasing the recognition performance of face recognition, especially in recent years.

Back to top