IEEE/CAA JOURNAL OF AUTOMATICA SINICA, VOL. 3, NO. 2, APRIL 2016 203 Pose Robust Low-resolution Face Recognition via Coupled Kernel-based Enhanced Discriminant Analysis Xiaoying Wang, Haifeng Hu, and Jianquan Gu Abstract—Most face recognition techniques have been success- ful in dealing with high-resolution (HR) frontal face images. How- ever, real-world face recognition systems are often confronted with the low-resolution (LR) face images with pose and illumina- tion variations. This is a very challenging issue, especially under the constraint of using only a single gallery image per person. To address the problem, we propose a novel approach called coupled kernel-based enhanced discriminant analysis (CKEDA). CKEDA aims to simultaneously project the features from LR non-frontal probe images and HR frontal gallery ones into a common space where discrimination property is maximized. There are four advantages of the proposed approach: 1) by using the appropriate kernel function, the data becomes linearly separable, which is beneficial for recognition; 2) inspired by linear discriminant analysis (LDA), we integrate multiple discriminant factors into our objective function to enhance the discrimination property; 3) we use the gallery extended trick to improve the recognition performance for a single gallery image per person problem; 4) our approach can address the problem of matching LR non-frontal probe images with HR frontal gallery images, which is difficult for most existing face recognition techniques. Experimental evaluation on the multi-PIE dataset signifies highly competitive performance of our algorithm. Index Terms—Face recognition, low-resolution (LR), pose vari- ations, discriminant analysis, gallery extended. I. I NTRODUCTION F ACE recognition techniques are mainly used to match unknown face images with an identified set of images for addressing security problems. As surveillance systems in wide range of public places are increasing, the demand for face recognition techniques for security and surveillance applica- tions is increasing. Though classic face recognition approaches such as [1−4] perform very well, their successes are con- fined to the conditions of strictly controlled scenarios, which are unrealistic in wide range of real applications. Real-life face recognition systems usually suffer from poor resolution, illumination and pose variation conditions, which adversely degrade the performance of most existing face recognition Manuscript received August 19, 2014; accepted February 16, 2015. This work was supported by National Natural Science Foundation of China (60802069, 61273270), the Fundamental Research Funds for the Central Universities of China, Natural Science Foundation of Guangdong Province (2014A030313173), and Science and Technology Program of Guangzhou (2014Y2-00165, 2014J4100114, 2014J4100095). Recommended by Associate Editor Jie Zhou. Citation: Xiaoying Wang, Haifeng Hu, Jianquan Gu. Pose robust low- resolution face recognition via coupled kernel-based enhanced discriminant analysis. IEEE/CAA Journal of Automatica Sinica, 2016, 3(2): 203-212 Xiaoying Wang, Haifeng Hu, and Jianquan Gu are with the School of Information Science and Technology, Sun Yat-sen University, Guangzhou 510006, China (e-mail: wangxiaoy05@163.com; huhai@mail.sysu.edu.cn; gujianquan123@163.com). approaches [5] . Therefore, it is necessary to consider the issue of robust face recognition systems which can jointly handle poor resolution, illumination and pose variations. During the past decades, research in the field of face recognition has been mainly focused on recognizing faces across variations in illumination and pose [6−8] . That is because they are still considered as major challenges encountered by current face recognition techniques. There is a recent trend towards addressing the problem of poor resolution face image recognition [9−14] , as the poor resolution face images are common in real life. For example, when subjects are far from cameras without any restriction, the face regions are very small or of poor quality, which causes low-resolution (LR) face recognition problem. In practical face recognition applications, gallery images are usually considered to be high-resolution (HR). Therefore, low- resolution (LR) will obviously cause the problem of dimen- sional mismatch between HR gallery images and LR probe ones in the special applications such as criminals monitoring. Generally, there are three ways to address the problem. The first way is down-sampling HR gallery images to the resolution of LR probe images, which seems to be a feasible solution for the mismatching problem. Unfortunately, the available discriminating information is drastically lost, especially when the resolution is very low such as 12 × 12 or even 8 × 8. The second and more widely used approach is based on the application of some super-resolution (SR) algorithms [9], [11], and [12], which obtain a high resolution version of the LR face image. Then the recovered HR image can be used for recognition. However, SR algorithms generally are time-consuming and not absolutely suitable for classification. Therefore, SR algorithms cannot meet current demands [15] of face recognition systems. The third way based on coupled mappings (CMs) without any super-resolution techniques is more effective, which is an essential way to solve the mismatch problem. The algorithms based on CMs try to learn two mappings which project the HR and LR face images into a common space where the distance between the LR image and its HR counterpart image is minimized while the distance between images of different subjects is maximized. For example, Li et al. [16] propose to use CMs to project LR and HR face images into a unified feature space. There is a problem with this method, i.e., poor recognition ability. Thus, they further introduce locality weight relationships [17] into CMs and propose coupled locality pre- serving mappings (CLPMs), which considerably improves the performance. However, CLPMs is sensitive to the parameters of the penalty weighting matrix such as scale. Subsequently,