Article Preview
Top1. Introduction
Face Recognition (FR) is a reliable biometric technique for identification and verification of persons using their face. FR has become more and more popular and intended to be used in many areas including access control applications (airports, authorized areas, mobiles, personal computers, etc.), surveillance in public spaces (stadiums, supermarkets, etc.), law enforcement applications, human-computer interaction applications, etc. Human face provides not only its identity, but also important soft biometric traits for human recognition such as gender, age and ethnicity. 2D FR systems that rely only on intensity information are more sensitive to pose, occlusions, environmental changes such as illumination and background and facial texture variations in makeup and facial hair. Despite significant improvements for more than half a century, 2D-based FR systems have not accomplished certain accuracy.
A face is not only a 2D image but also a three-dimensional form representing depth information whose major advantages are robustness to changes in lighting, pose, background and their resistant to spoofing security attacks. With technological advances reached during the last two decades, the development of 3D sensors has opened new frontiers for FR systems. With the introduction of 3D information, the problem of segmenting the background of the face has become much easier. Furthermore, more representations and descriptors are derived and used to handle the variations of face pose and expression. Nerveless, 3D face images are still difficult to acquire due to several issues such as cost and accessibility. With the rapid development and decreasing cost of 3D data acquisition devices, 3D facial data can be easily captured for real time application. A good compromise between RGB-D data acquisition and the cost of the sensor is provided. With the emergence of this type of scanners and the advancements of deep learning based methods, one promising solution is to use 3D information of the face to train deep learning based methods to deal with face pose variations, illumination changes and partial occlusions.
Surface normal has not been sufficiently explored in terms of 3D face representation and description. In this paper, the authors propose a method for learning surface normal derived from 3D information for 3D face recognition. To fully exploit surface normal, the three components of normal vectors of each face are first estimated, and then rendered as an RGB image used as input of deep convolutional neural network. The main contributions of this work can be summarized as follows:
- •
The use of normal maps instead of depth maps for 3D face recognition using low quality depth sensor
- •
The signal-level fusion of the normal maps components to construct an RGB image
- •
The investigation of the order of components that yields to the best accuracy
- •
The evaluation of verification mode using data captured by low quality depth sensor
The remaining of this article is organized as follows. First, Section 2 gives an overview of the related work. Then, the proposed depth-based face recognition approach is presented in detail in Section 3. Section 4 summarizes the performed experiments and the obtained results. Finally, Section 5 concludes the article with some observations and perspectives for future work.