Machine Vision and Applications (2011) 22:337–348 DOI 10.1007/s00138-009-0235-6 ORIGINAL PAPER Attention control with reinforcement learning for face recognition under partial occlusion Ehsan Norouzi · Majid Nili Ahmadabadi · Babak Nadjar Araabi Received: 20 February 2009 / Revised: 2 September 2009 / Accepted: 13 November 2009 / Published online: 3 January 2010 © Springer-Verlag 2010 Abstract In this paper a new method for handling occlusion in face recognition is presented. In this method the faces are partitioned into blocks and a sequential recognition struc- ture is developed. Then, a spatial attention control strategy over the blocks is learned using reinforcement learning. The outcome of this learning is a sorted list of blocks according to their average importance in the face recognition task. In the recall mode, the sorted blocks are employed sequentially until a confident decision is made. Obtained results of var- ious experiments on the AR face database demonstrate the superior performance of proposed method as compared with that of the holistic approach in the recognition of occluded faces. Keywords Attention control · Face recognition · Reinforcement learning · Occlusion 1 Introduction Face recognition enjoys extensive attention due to its numer- ous applications in comparison to other biometric methods. Nevertheless, high dimensionality of image data in addition E. Norouzi (B ) · M. Nili Ahmadabadi · B. Nadjar Araabi Department of Electrical and Computer Engineering, Control and Intelligent Processing Center of Excellence, University of Tehran, Tehran, Iran e-mail: e.norouzi@ece.ut.ac.ir; enorouzi@yahoo.com M. Nili Ahmadabadi · B. Nadjar Araabi School of Cognitive Science, Institute for Research in Fundamental Sciences (IPM), Tehran, Iran e-mail: mnili@ut.ac.ir B. Nadjar Araabi e-mail: araabi@ut.ac.ir to existence of various distractions in face images—such as facial expressions, lighting variations, and occlusions—is a big challenge of real world applications of face recognition systems. In this paper, an attention-based sequential recog- nition method is proposed to handle the occlusion problem in addition to the high dimensionality of the feature space. Most face recognition systems are appearance-based, i.e., the images are input to the system as pixel arrays. Because of the high dimensionality of the face images, feature extraction and feature selection are integral parts of any face recognition system. In most of the existing systems, feature extraction and selection are done holistically and over the whole image, see [1] for a complete survey of face recognition methods. As a result, even with the best feature extraction and selection methods, the feature vectors are still too long and employ- ing them for recognition is not that easy. Therefore, further reduction of the feature vector size is desired provided that this reduction does not affect the recognition accuracy, and even boosts it [2]. One major way to do so is to implement the recognition process sequentially rather than in a holistic man- ner [3]. In the sequential methods, the images are partitioned into some blocks and feature extraction and classification processes are performed for each block independently. The final decision is made with aggregation of the separate classi- fication results of all or some selected blocks [3–5]. By doing so, the feature vectors that should be handled by the system in each time step are shortened. In addition, sequential rec- ognition makes spatial attention control feasible. Here, we develop a sequential face recognition system. Reviewing the literature of face recognition, it can be seen that the major successes in this field are achieved in controlled conditions, i.e., when the faces to be recognized are not so much different from those in the training data. This is not the case in most real-world tasks, where the images include different types of distractions. In this research, we deal with 123