banner

FACE Recognition

In the present day scenario the security systems in place are saturated with cameras which continuously record video data but are not capable of active object identification and verification. As a solution Watch ID goes beyond these limitations and is capable of univocal verification of individuals who come in front of the camera. It capitalizes on the recently found ability of computers to recognize and differentiate between faces.                                        

Technology behind Face Recognition

Every face is unique and has numerous distinguishable points with different peaks and troughs that make up facial features. Every human face has approximately 80 such distinguishable points like width of the nose, distance between eyes, width of the forehead etc. These are measured creating numerical code that represents a face in the database.

There are two basic challenges when it comes to the implementation of face recognition technology – face verification and face identification.

The face verification system is usually limited to personal database, which apart from the typical personal information and the information of the classical face recognition system on the access control, contains characterizations of the face of individuals who have been registered in the system. These face characteristics are called ‘descriptors of the face’ in the nomenclature of face recognition system. They create a compact representation of the face of the concerned person, which makes it possible to effectively store a large database of identified individuals, and secondly allows easy comparison. The verification in security system performs a supplementary function with relation to already existing technologies. The system’s high efficiency allows exact verification of a person on the basis of his/her looks. Therefore, the system makes it impossible to fake identity.

Identification on the other hand is a more complex challenge. It has applications in monitoring systems for object identification (airfields, airports, railway stations, public utility buildings, etc.) wherein many cameras are deployed to simultaneously look over many locations. The personal database in case of the identification challenge is usually a lot larger than in the case of verification and contains descriptors (optionally also photographs) of unwanted or undesirable persons in particular places.

The manner in which the M3S system recognizes faces is based on Principal Component Analysis (PCA) and the Fisher's Linear Discriminant Analysis (LDA). The PCA method on its own does not give good results in recognitions. This is primarily because the system is based on global statistics data, which does not efficiently differentiate faces of individuals that are captured in different lighting conditions and attitudes. It is only in conjunction with LDA that the system makes it possible to obtain minimum levels of incorrect classification. The LDA is the separation of these features that are unique to a particular person. In fact, initial tests have confirmed a high level of efficiency as per the proposed method both in the identification as well as verification of individuals.

Face recognition must be preceded by an arduous process of location of the face in the image, a normalization of found faces and a filtration of the features. These are auxiliary elements in the recognition process. However, their influence on the result of the whole task is very significant. Many described methods of face location exist whose efficiency is relative to the complexities of the analyzed scene and the intensity of the traffic occurrence in the series of images. Since the algorithm of recognition has to integrate itself with monitoring systems, our system accepts solutions which are based on the estimation of traffic. It allows an easy way of locating objects in an analyzed sequence of video images. Objects after segmentation are subjected to the procedure of local feature analysis, which looks at exteriorization of face images out of it.

The normalization process targets at bringing separated objects of the face to the one required by the algorithm for the construction of descriptors of sizes. In this phase, there are practical procedures of scaling and rotation of the image. The transformations are based on determined position of eyes and mouth in the identified image.

The objective of the filtration process is the exclusion of disturbances that are there in the image’s background. It extinguishes pixels surrounding the face to the determined level, so that its values will be equal among different images of the same face.

Face images are subject to the above-tooling and are delivered by the method of construction of descriptors, which brings to light the definite number of features that help determine the identifier of the face.

The above technique, based on descriptors, permits the storage of data on microprocessor cards. It seems that such effective methods of face diagnosing will play a large part in solving problems of identity checks, catching of wanted persons, verification of operations passed with the use of payment cards, etc. in the near future.