Face Recognition Using SIFT and Binary PSO Descriptors Lanzarini Laura, La Battaglia Juan, Maulini Juan, Hasperué Waldo III- LIDI (Institute of Research in Computer Science III-LIDI) School of Computer Science UNLP - La Plata, Bs.As. - Argentina {laural, juanlb, jmaulini, whasperue}@lidi.info.unlp.edu.ar Abstract. In this paper, a strategy for face recognition based on SIFT descriptors of the images involved is presented. In order to reduce the number of false positives and computation time, a selection of the most representative feature descriptors is carried out by applying a variation of the binary PSO method. The results obtained show that the strategy proposed is better than the direct application of SIFT descriptors. Keywords. Face recognition; SIFT Features; Swarm Intelligence 1. Introduction Face recognition is a biometric technique that is widely used in various areas such as security and access control, forensic medicine, and police controls. It involves determining if the image of the face of any given person matches any of the face images stored in a database. This problem is hard to solve automatically due to the changes that various factors, such as facial expression, aging and even lighting, can cause on the image. In this paper, a method using only those SIFT descriptors that best represent the image is proposed. Good recognition results are achieved while solving the two major problems of this characterization method: false positive detection and the time required for the recognition process. The selection of SIFT descriptors is carried out by means of a variation of binary PSO (Particle Swarm Optimization), and it is applied only to database image descriptors. Therefore, SIFT descriptors processing is done before the recognition stage of the process. This paper is organized as follows: In Section 2, a brief description of previous related works using similar techniques is included; in Section 3, the method that allows obtaining SIFT descriptors from an image is described; whereas in Section 4 some clarifications regarding the binary PSO variation used are presented. In Section 5, implementation details are provided, and in Section 6 the results obtained are described. Finally, in Section 7 the conclusions obtained are presented. 2. Related work There are currently various solutions to this problem that use SIFT descriptors. It has been shown [1] that using SIFT descriptors for the face recognition process is better than Eigenfaces and Fisherfaces algorithms. Training datasets were of various sizes, which allowed establishing that performance decreases as dataset size decreases. As regards the significant number of SIFT descriptors required for a reliable comparison, it was observed that, with a lower number of descriptors, performance is better than that obtained with Eigenfaces and Fisherfaces. In order to tackle the issue of comparing very long feature vectors for all images in a database, a biased classification of the features that make SIFT descriptors, is proposed and used to reduce the length of SIFT descriptors used for face recognition [6]. Thus, the number of comparisons is reduced and the recognition process is faster. This process also filters out those descriptors that are irrelevant for face recognition, thus increasing recognition accuracy. On the other hand, a face recognition algorithm that uses the binary PSO algorithm to explore the solution space for an optimum subset of features in order to increase recognition rate and class separation is presented in [8]. This algorithm is applied to feature vectors extracted using the Discrete Cosine Transform (DCT) and the Discrete Wavelet Transform (DWT). 3. SIFT Features In [5], Lowe defined a method to extract features from an image and use them to find 557 Proceedings of the ITI 2010 32 nd Int. Conf. on Information Technology Interfaces, June 21-24, 2010, Cavtat, Croatia