1. Introduction

In recent years there have been great expectations of biometric authentication in view of increasing vicious crimes and terrorist threats. Biometric authentication is automated identification or identify-verification of an individual based on physiological or behavioral characteristics such as fingerprint, iris, face, vein, voice, and so on. These kinds of authentications are the most commonly used to safeguard international borders, control access to facilities, and enhance computer network security.

Face authentication has interesting characteristics that other biometrics do not have; facial images can be captured from a distance, any special actions are not always required for authentication, and a crime-deterrent effect can be expected because the captured images can be recorded and we can see who the person is at a glance. Due to such characteristics, face recognition technique is expected to be applied widely not only to security applications such as video surveillance but also to image indexing, image retrievals and natural user interfaces.

Media and Information Research Laboratories of NEC has a couple of research activities concerning face recognition, which will be introduced here.

2. Three-dimensional face recognition using geodesic illumination bases [1,2,3,4]

Appearances of a face changes drastically when imaging conditions such as pose and illumination are varied, creating serious difficulties in identifying the resulting image. Even when there are only the illumination changes, its effects override the unique characteristics of individual features and thus greatly degrades the performance of state-of-the-art systems. Therefore, robust recognition systems need to model and compensate for variations in the image caused by pose and illumination changes.

Although pose can be concisely modeled by using six physical parameters, describing illumination conditions with a small number of physical parameters is quite difficult because types or quantity of light sources have an infinite degree of freedom. Previous studies, however, have shown that illumination variations of an image can be described concisely, especially in the case of faces.

The authors’ group proposes a 3-D face recognition method that is robust against any lighting variations and a large extent of pose variations ranging from frontal to profile views. The proposed method constructs an appearance description model of a face that consists of a 3-D shape and geodesic illumination bases (GIBs). The appearance model can describe any image variations due to pose and illumination changes, including shadows. An appearance model on the surface of a 3-D object, not in an image space is constructed (fig.1). The GIBs describe irradiances of the object surface under any illumination and generate the illumination subspace of an image in an arbitrary pose to describe illumination variations. To cope with background separation and self occlusion, the model uses individual 3-D shapes. The appearance model is automatically aligned with the target image by estimating the pose parameters from roughly given values, so it does not require any precise prior pose knowledge that cannot be obtained in real situations.

Experimental results showed that the performance of the proposed method achieved a first success ratio of 94.2% when used with an extensive database that had 14,000 test images consisted of 200 individuals with drastic illumination changes in a large extent of posed ranging up to 60 degrees sideward and 45 degrees upward from a frontal pose.

3. Two-dimensional face recognition using perturbation space method[5,6]

The authors’ group also develops a universalized face model to generate various facial appearances, which does not require the individual’s own facial shape and reflectance on the surface. This face model consists of a generalized 3-D facial surface and a set of illumination bases also generalized, for diversified range of persons (fig.2). For pose changes, the enrolled image is mapped onto the 3-D facial surface, and is rotated in 3-D space to simulate out-of-plane pose changes. For illumination changes, various shaded images are generated from the enrolled image using the illumination bases which were obtained by executing principal component analysis (PCA) on various facial images.

If all of these generated images are enrolled to the database, it may require huge computational time to compare between a query and each image among the database in the recognition process. To solve this problem, these images are compressed using PCA, and the comparison is executed between the query and compressed linear subspace including almost all enrolled images. This method, which is called the perturbation space method, takes account of pose and illumination changes even within short calculation time, avoiding any pose and illumination estimation process.

4. Conclusion

For reducing the adverse effect caused by pose and illumination changes, a model based approach is very effective, because such variations are induced by physical properties. This paper described two types of face recognition methods using both the individual and generalized 3-D based model. In future works, the authors’ group would like to reduce specularity effect with strongly view-dependency and large amount of calculation cost of the pose optimization process in 3-D face recognition. In ordinary 2-D face recognition, estimating 3-D facial shape from an image should be developed to improve the recognition performance even further.

Acknowledgements
The author gratefully thanks the biometrics group of Media and Information Research Laboratories, especially Dr. Atsushi Sato, Dr. Hitoshi Imaoka, and Mr. Rui Ishiyama, for their great efforts to solve the problems.

Reference