Face Recognition by Extending Elastic Bunch Graph Matching with Particle Swarm Optimization Rajinda Senaratne 1 , Saman Halgamuge 1 , Arthur Hsu 2 1 Department of Mechanical Engineering, Melbourne School of Engineering, The University of Melbourne, Australia; 2 Bioinformatics Division, The Walter and Eliza Hall Institute of Medical Research, Australia Email: rajinda s@hotmail.com, saman@unimelb.edu.au, hsu@wehi.edu.au Abstract— Elastic Bunch Graph Matching is one of the well known methods proposed for face recognition. In this work, we propose several extensions to Elastic Bunch Graph Matching and its recent variant Landmark Model Matching. We used data from the FERET database for experimentations and to compare the proposed methods. We apply Particle Swarm Optimization to improve the face graph matching procedure in Elastic Bunch Graph Matching method and demonstrate its usefulness. Landmark Model Matching depends solely on Gabor wavelets for feature extraction to locate the landmarks (facial feature points). We show that improvements can be made by combining gray-level profiles with Gabor wavelet features for feature extraction. Furthermore, we achieve improved recognition rates by hybridizing Gabor wavelet with eigen- face features found by Principal Component Analysis, which would provide information contained in the overall appear- ance of a face. We use Particle Swarm Optimization to fine tune the hybridization weights. Results of both fully automatic and partially automatic versions of all methods are presented. The best-performing method improves the recognition rate up to 22.6% and speeds up the processing time by 8 times over the Elastic Bunch Graph Matching for the fully automatic case. Index Terms— Eigenfaces, Elastic Bunch Graph Matching, Face Recognition, Gabor wavelets, Hybridization, Particle Swarm Optimization, and Principal Component Analysis I. I NTRODUCTION Face recognition is a challenging problem in pattern recognition research. Many face recognition methods have been proposed in the past few years, and Elastic Bunch Graph Matching (EBGM) [1]–[5] is considered as one of the successful methods. In EBGM, a face is represented by a face graph (FG). The local features of a facial landmark are represented by a jet, where a jet is a set of Gabor wavelet features. A Face Bunch Graph (FBG) is created as a generalized representation of faces of various individuals, thus, it consists of a ‘bunch’ of jets corresponding to a landmark. To obtain the optimal FG to represent a new face, a two-step approach is adopted. The first step is to create a new FG for the new face, which will be fitted to the face in the second step. The fitting of the FG to the face is made by an iterative face graph matching procedure, where its geometrical structure is deformed until the graph similarity between the FG and the FBG is maximized. When maximizing this graph similarity, the best matching jet is selected from the bunch. Even though EBGM is considered as a successful technique, it has several deficiencies in obtaining the optimal FG: • It uses only few sizes of FGs to estimate the size and the location of the face in an image (three sizes in stage 1, and two sizes in stage 2 [2]). The ability to use an FG of any size is more desirable as it may facilitate more accurate estimation of the size and the location of a face, so that the range of different sizes of faces that would exist in a generic face database can be tolerated. • It initially scans the image only at selected locations (In stage 1, the image is scanned at locations on grid points of a lattice with a spacing of 4 pixels [2]). A non-exhaustive method that is capable of placing the FG at any location, rather than only at the selected locations, would be more effective. • It is a computationally intensive algorithm [5]–[7], as it has several stages with extensive searches. A more efficient approach to obtain the optimal FG for a new face would be to optimize the face graph matching procedure using a suitable optimization technique. To overcome the above problems, inspired by both EBGM and Active Shape Model [8], a new method named Landmark Model Matching (LMM) [9], [10] was pro- posed, which incorporates an evolutionary optimization method known as Particle Swarm Optimization (PSO) [11]. LMM disregards the concept of using the entire ‘bunch’ of jets to locate a landmark. EBGM compares against the best matching jet in the bunch, whereas LMM compares against only the average of the jets. By using the entire bunch of jets, it would be possible to cover a wide variety of features that would exist in a face. Therefore, in this work, we propose a new method named EBGM PSO , which is an extension of EBGM that exploits its unique strength of using a bunch of jets to locate a landmark. In EBGM PSO , the face graph matching procedure of EBGM is optimized using PSO. Above methods use Gabor wavelets for feature extrac- tion to locate the landmarks.In this work, we extended Landmark Model Matching by combining Gabor wavelet 204 JOURNAL OF MULTIMEDIA, VOL. 4, NO. 4, AUGUST 2009 © 2009 ACADEMY PUBLISHER