Synonyms

Three-dimensional face identification; Three-dimensional face authentication; Three-dimensional face verification

Definition

3D face recognition is referred to as person identification using the three-dimensional geometry of the face.

Introduction

The authentication or recognition of an individual’s identity is an integral part of numerous systems developed in the context of a variety of applications ranging from law enforcement and surveillance of public places to digital rights management and engaging human-computer communication. Biometrics, the physical and behavioral traits that characterize uniquely a human, such as fingerprints, signature, voice and face, have attracted growing research interest as an alternative means for reliable person identification over passwords, PINs, smart cards and similar mechanisms. In particular, person identification from face images has gained a prominent position among other biometrics. This is mainly due to the social acceptability of face as a biometric and the user-friendliness of face recognition systems, but is also motivated by the inherent ability of humans to recognize faces visually in their natural inter-personal communication with apparently minimum effort.

For many decades, face images and image sequences have been used as a basis for establishing identity and a plethora of algorithms have been proposed; an extensive review may be found in [1]. However, despite the great advances, face recognition from images in vivo is still far from being a reliable and robust way for establishing identity. Recorded images of the same face may have large variations in appearance, caused for example by changes in the illumination of the scene or orientation of the head. This contributes considerably to the performance of existing systems when applied to non-controlled conditions. To cope with these difficulties, a promising new technology, 3D face recognition, was recently investigated. This relies on discriminative features extracted from the 3D geometry of the face, which is captured by special 3D scanners or 3D cameras. Since 3D geometry is independent of appearance variations, 3D face recognition has demonstrated very good accuracy even in “difficult” conditions.

A typical face recognition system based on 3D data consists of four parts in general, which involve data acquisition, pre-processing, feature extraction and feature classification. 3D data (also called depth) may be captured directly, using a laser scanner, or indirectly, applying stereo photogrammetry techniques to color images captured by multiple cameras or using a single camera and a structured-light projector. The result of the acquisition is usually a “range image” with each pixel recording the distance of a point on the object from the camera. The accuracy and resolution of the data depends on the acquisition technology and environment. Laser scanners for example lead to more accurate 3D face models, but need several seconds to grab a single image, while being also relatively expensive. On the other hand, stereo-based and structured-light based techniques are faster and cheaper but produce noisy data.

Once 3D facial data is acquired, it undergoes standard pre-processing, such as noise removal, hole filling and smoothening, or more elaborate pre-processing, such as head pose compensation usually using variants of the Iterative Closest Point algorithm [2]. Then, several surface features that hopefully characterize unambiguously each facial surface are computed. These features, also called descriptors, may be local or global and they are usually defined in terms of surface curvature, distance between surface points, moments and other geometrical attributes of the surface. Point signatures [3], spin images [4] and snapshots [5] are examples of local surface descriptors, while contour-lines [6] and canonical images [7] are more global descriptors. Particular emphasis is placed on these features being pose-invariant to dispense with pose compensation and eliminate the associated errors. Exhaustive literature surveys for 3D and multimodal 2D and 3D face recognition algorithms may be found in [8, 9].

Feature classification usually follows vector-oriented pattern classification techniques such as Support Vector Machines, Neural Networks, Principal Components Analysis and Linear Discriminant Analysis, or statistic-based techniques, such as Bayes Classification and Hidden Markov Models [10]. Most of the classifiers proposed in literature require training with a representative number of training faces to adjust a set of parameters. Then, given this set of parameters and a gallery of already processed faces whose identity is known, recognition of new unseen faces, called probe faces, may be performed. The features of the probe face are extracted first and then they are matched against the corresponding features of the gallery faces to establish a measure of similarity. The probe face is finally classified to the most similar gallery face.

Quantitative evaluation and comparison between face recognition algorithms is usually performed under two experimental scenarios, face identification, which responds to the question “Who is he?”, and face authentication, which responds to the question “Is he really who he claims to be?” Authentication is a 1:1 match process where the face in question is compared with the face whose identity is claimed. On the other hand, identification is a 1:N match process, since the face in question is compared with all faces stored in a database. Performance is assessed by means of the Cumulative Match Characteristic (CMC) and the Receiver Operating Characteristic (ROC) curves for the identification and the authentication scenario respectively. The CMC is the graph of recognition rate versus the ranking of similarity measures, while the ROC curve is the graph of False Rejection Ratio versus False Acceptance Ratio. Recently, Face Recognition Grand Challenge (FRGC) [11], has been established as a framework for evaluation of 3D face recognition algorithms. FRGC is a project trying to promote and advance 3D face recognition technology and to this end it offers a common data set and a common infrastructure for experiments and quantitative assessment. Its goal is to exceed by an order of magnitude the performance of state-of-the-art 2D face recognition algorithms promoted by a corresponding project, the Face Recognition Vendor Test (FRVT), and by this time many algorithms capable of achieving recognition rates greater than 90% have been proposed.

The availability of cheap and accurate 3D sensors has led to a booming of 3D face recognition technology over the past 10 years and research has already reached a sufficiently mature level. Although pitfalls inherent in 2D systems have been successfully encountered or alleviated, a new challenge has come up regarding 3D face recognition, facial deformations that alter the geometry of the face such as those caused by facial expressions. Most algorithms proposed so far are sensitive to the deformation of the surface caused by facial expressions and their performance decreases considerably when dealing with non-neutral faces. Of course, there are other problems that need to be encountered too, such as the age variation and occlusion, but current research is primarily focusing on expression-invariant 3D face recognition.

Cross-References

Face and Facial Expression Recognition Using 3D Data: An overview

Three Dimensional Face Identification

Three Dimensional Face Verification