Face Recognition Using Optimized 3D Information from Stereo Images

Human biometric characteristics are unique, so it can not be easily duplicated [1]. Such information includes; facial, hands, torso, fingerprints, etc. Potential applications, economical efficiency, and user convenience make the face detection and recognition technique an important commodity compared to other biometric features [2], [3]. It can also use a low-cost personal computer (PC) camera instead of expensive equipments, and require minimal user interface. Recently, extensive research using 3D face data has been carried out in order to overcome the limits of 2D face detection and feature extraction [2], which includes PCA [3], neural networks (NN) [4], support vector machines (SVM) [5], hidden markov models (HMM) [6], and linear discriminant analysis (LDA) [7]. Among them, PCA and LDA methods with self-learning method are most widely used [3]. The frontal face image database provides fairly high recognition rate. However, if the view data of facial rotation, illumination and pose change is not acquired, the correct recognition rate remarkably drops because of the entire face modeling. Such performance degradation problem can be solved by using a new recognition method based on the optimized 3D information in the stereo face images. This chapter presents a new face detection and recognition method using optimized 3D information from stereo images. The proposed method can significantly improve the recognition rate and is robust against object’s size, distance, motion, and depth using the PCA algorithm. By using the optimized 3D information, we estimate the position of the eyes in the stereo face images. As a result, we can accurately detect the facial size, depth, and rotation in the stereo face images. For efficient detection of face area, we adopt YCbCr color format. The biggest object can be chosen as a face candidate among the candidate areas which are extracted by the morphological opening for the Cb and Cr components [8]. In order to detect the face characteristics such as eyes, nose, and mouth, a pre-processing is performed, which utilizes brightness information in the estimated face area. For fast processing, we train the partial face region segmented by estimating the position of eyes, instead of the entire face region. Figure 1. shows the block diagram of proposed algorithm. This chapter is organized as follows: Section 2 and 3 describe proposed stereo vision system and pos estimation for face recognition, respectively. Section 4 presents experimental, and section 5 concludes the chapter.