Face Recognition using Parallel Associative Memory Karm Veer Arya AVB-Indian Inst. of Info. Tech. & Management Gwalior, India kvarya@gmail.com Viraat Singh Siemens System India Ltd. Bangalore,India viraat.singh@gmail.com Pabitra Mitra Department of C. S. E. I.I.T. Kharagpur Kharagpur, India pabitra@cse.iitkgp.ernet.in Phalguni Gupta Department of C. S. E. I.I.T. Kanpur Kanpur, India pg@cse.iitk.ac.in Abstract—This paper proposes a Parallel Associative Mem- ory (PAM) based face recognition system. An image is split into blocks which are processed in parallel by conventional associative memory. The combined output of these are used to discriminate between faces. Further, a parameter termed as run length count has been used to discriminate between similar faces. This approach helps to scale up associative memory to higher resolution images with more detailed features thus improving recognition performance. To analyze the efficiency and goodness of the proposed system the experiments have been performed on ORL face database. The proposed method out performed most of the existing methods. Index Terms—Associative memory, Face recognition, Parallel associative memory, Run length count I. I NTRODUCTION Face verification is the process of one-to-one matching of images where the query face image (i.e., the image whose identity is being claimed) is compared with the template face image. In recent past face recognition has attracted a lot of attention of researchers from pattern recognition, computer vision and biometric communities [1], [2], [3], [4]. Face recognition has been well studied in literature [5], [6], [7], [8], [9], [10], [11]. However, most of these algorithms perform poorly in the presence of uncertainties in the face image. Face images inherently contain uncertainties including spectacles, beard and poor lightening condition etc. To handle these uncertainties face recognition system based on associative memories have been proposed in literature [5], [9], [13], [14]. Associative memory is a neural architecture with human like capacity to recognize pattern in an uncertain environment [13], [12]. Associative memory usually enables a parallel search in a stored data file [12]. The above property may be used to build robust face biometric systems. However, the main drawbacks of associative memories are that they do not scale to large image sizes having high resolution. As an example the associative memory based ARENA Face Recognition Algorithm [5], [14] reduces the image size to 16 × 16 pixels and as a result, it looses many features of the face which can be used in differentiating the images. In the RCE-based Associative Memory approach to human face recognition [8] the data set is divided in blank, smile, angry and surprised groups. This approach works well for blank faces but performance degrades drastically for other expressions. Moreover, a lot of preprocessing is required on the given data set. In Kernel Autoassociator Approach (KAA) to Pattern classification [11], kernel autoassociator takes a kernel feature space as the nonlinear manifold, and places emphasis on the reconstruction of input patterns from the kernel feature space. The computational complexity of the kernel autoassociator approach is very high even for moderate resolution images. The concept of sub-images and parallel processing were used in [23] to identify frontal view of human face. In [23] sub- images are referred as decomposition of large image and Fast Neural network is used for parallel processing. Decomposition and parallel processing is used for human face detection in a large image. This paper proposes parallelization of the auto-associative memory in order to apply it for recognition of high resolution face images. The goal is to retain discriminative feature of the face and thus to improve the recognition performance. Fusion of the output of each component is performed using a novel scheme which considers run lengths of the binary output of the component. The system has been found to be efficient and performed satisfactorily under varying illumination conditions, change in facial expressions, using spectacles etc. The Olivetti Research Laboratory (ORL) face database [15] has been used to evaluate the performance of the proposed system. Six face images of the same person are used for training the proposed neural network and ten face images of the same person and 390 images of different persons are used for testing. The False Rejection Rate (FRR) and False Acceptance Rate (FAR) are used as performance metrics. Rest of the paper is organized as follows. Auto-associative memory for face recognition is proposed in Section 2. Next section deals with parallel associative memory for face recog- nition. Section 4 describes the face similarity measure for recognition. Experimental results are discussed in Section 5. Conclusions are given in the last section. II. AUTO- ASSOCIATIVE MEMORY FOR FACE RECOGNITION Associative memories mimic the capacity of human brain to recall information in a robust and associative access mode [12]. The concept of auto-associative memory to store and recall the data has been proposed by Kohonen [16], [17]. Associative memory usually enables a parallel search in a stored data file. The purpose of search is to output either one or all stored data items that match the given search argument, and to retrieve it fully or partially [12]. An associative memory is a distributed memory with human brain like capacity which 1332 1-4244-2384-2/08/$20.00 c  2008 IEEE Authorized licensed use limited to: National Kaohsiung University of Applied Sciences. Downloaded on January 13, 2010 at 07:37 from IEEE Xplore. Restrictions apply.