Abstract—Face recognition algorithms work properly only under strict conditions. There is no doubt that the system accuracy drops significantly when some of those conditions are not applied. In this paper, we take care of the most detrimental factor, variant head poses, to recognize the face. Novel method is proposed to construct the virtual view of face from a single frontal face. Instead of using the whole face, some patches are selected from only half of the face. Features are extracted via biorthogonal wavelet transform. Block of feed forward neural networks is applied to infer the mapping ratio between poses and gallery. CMU database is utilized to demonstrate the robustness of the proposed method. Index Terms—Pose, biorthogonal wavelet transform, face recognito, human machine interaction. I. INTRODUCTION Although more than three decades of efforts have been dedicated to face recognition systems, today’s systems are only reliable under strict conditions. The three most important restrictions to recognize the face in real world are illumination, expression and pose. However it has been proven that the pose is more difficult to deal with and it is well-known as a bottleneck in recognizing the face [1]-[4]. It could be shown that the extracted data from different identities show more similarity when compared to data extracted from the same identity but in different poses. If the system could recognize the face without preconditions, it will be useful for variety of applications such as security and surveillance systems and human-computer interactions. Still there is yet another problem which has been neglected in most of previous works [5]-[12]. Specifically it is the number of face images per person used as gallery. Since in real-world it is a tedious task to provide many images from different poses of one identity in the database, it is noteworthy for the system to remain accurate by relying only on the single image per person. The frontal face is more appropriate to be used as a gallery because of the important and necessary information it provides in applications such as in a access control system as well as for biometric database (passports or ID cards). There are different categories in the recognition of face with varying poses. In the view point of types of image, it is divided to 2D and 3D. There are many works which used three dimensional images [11], [13]-[17]. These systems could obtain higher accuracy compared to 2D images. Manuscript received April 9, 2012; revised May 10, 2012. The authors are with the Computer Vision, Video and Image Processing (CvviP) Research Lab, Faculty of Electrical Engineering, Universiti Teknologi Malaysia, 81310 UTM Skudai, Johor, Malaysia (e-mail: shahdi@fkegraduate.utm.my, syed@fke.utm.my) However, due to their high dimensionality of data, they are time consuming and thus not suitable for real-time applications. Another drawback of using 3D images is the huge database needed for every image at different angles. On the contrary, 2D image could be easily gathered via ordinary camera. It also does not need too much time to process the data because of its low dimensionality. Another categorization is according to their usage of either the entire face (global methods) [5], [7], [18]-[22] or only using some patches in the face (local methods) [23]- [25]. Theoretically using the entire face should give the system better performance compared to using local methods. Nevertheless, since some regions of the face do not provide extra information, their inclusions do not improve the accuracy and in some situation those regions may cause misclassification. Thus, in this case local methods provide more promising results. In the method proposed by Li et al. [1] similarity between faces is evaluated via correlations in a media subspace among different poses. Their work is also based on the patch level and in their method media subspace is constructed by Canonical Correlation Analysis in order to maximize the intra-individual correlations. However, the pose is limited to angles with which two eyes are visible. Du et al. [2] partition the whole image into 7 fixed size patches. The virtual frontal for each patch is then estimated separately. The virtual image is constructed by integrating these virtual frontal patches. Similar to the previous methods [1], [3], the pose is confined to the orientation whereby two eyes are visible. The other problem with their method is that the fixation of patch size and its only dependency on the two centers of the eyes may cause a considerable reduction in accuracy. Fig. 1. Proposed method in a glimpse, where P and G represent pose and gallery (frontal) images respectively. Index i shows the i`th person used as a training image and index j shows j`th person used as a probe image. Index k shows the corresponding pose. Face Recognition under Variant Head Poses Using Single Frontal Gallery Image and Facial Symmetry Seyed Omid Shahdi and S. A. R. Abu-Bakar 288 International Journal of Computer and Electrical Engineering, Vol. 4, No. 3, June 2012