© 2015 ICECE May 10 - 13, 2015, Žilina, SLOVAKIA IX International Conference on Engineering and Computer Education 22 AN EMPIRICAL STUDY OF THE BEHAVIOR OF A FACE RECOGNITION SYSTEM BASED ON EIGENFACES AND K- NEAREST NEIGHBORS TECHNIQUES Fabio Abrantes Diniz 1 , Thiago Reis da Silva 2 , Francisco Eduardo Silva Alencar 3 1 Fábio Abrantes Diniz, Federal University of Campina Grande – UFCG/PPgCC, Brazil, fabio.abrantes.diniz@gmail.com 2 Thiago Reis da Silva, Federal University of Rio Grande do Norte – UFRN/PPgSC, Brazil, thiagoreis@ppgsc.ufrn.br 3 Francisco Eduardo da Silva Alencar, Rural Federal University of the Semi-Arid – UFERSA/PPgCC, Brazil, eduardu.dudu@gmail.com Abstract  Developing a face recognition computational model is a hard task. Extracting facial features from facial images becomes a hard when the images have different dimensions, especially in the steps of extraction and classification. In this paper, we propose an empirical study of optimization on the rate accuracy results from a facial recognition based on Eigenfaces and K-Nearest Neighbors techniques. It was investigated the following topics: images with three different dimensions, number of features (Eigenfaces), k values from K-Nearest Neighbors technique and three distance measures. Addressing the problems of image dimensionality for facial recognition, understanding which parameters are more relevant from the addressed techniques in order to enhance the accuracies rate of facial recognition were the goals of this study. Following this, it was proved from the experiments that images with 12x9 sizes produce the best facial recognition accuracies rate, using the normalized Euclidean distance and a number of Eigenfaces equals to twenty. Index Terms  Empirical Study, Recognition System, Eigenfaces and K-Nearest Neighbors Techniques. INTRODUCTION Face recognition is one of the most used identification process by humans, allowing then to quick identify any individual. Although facial recognition is a simple task for humans, it is not trivial to implement this process in a machine. Modeling a face that abstracts features that differentiate one face to another faces is the most difficult step in the implementation, since they have few substantial differences [1]. In addition, images with different dimensions make the recognition process a hard task, especially in the steps of extraction and classification of the facial features. Many algorithms have been proposed to solve the facial recognition problems [2]. In this paper we present an empirical study that optimizes the accuracy rate for a face recognition system [3] based on Eigenfaces techniques [4] and K-Nearest Neighbors (K-NN) [5]. The covered variable techniques which were analyzed are the following: (a) three- dimensional images; (b) number of features ranging from 15 to 20; (c) the value of k of K-NN technique ranging from 1 to 10 and (d) the use of the three measures distances (Euclidean, Manhattan and the Normalized Euclidean). This study was important in the image dimensionality problems for facial recognition in the following way: verifying which of the dimensions of the images provides the best significant facial features. Furthermore, it provided an analysis of the values of the relevant parameters for the techniques were used. An experiment was done using a database containing 1280 images from a total of 64 individuals. Each individual of this database was represented by 20 images in five different positions of the individual. According to the main results of experimental tests, it was found that images of different sizes produce different rates accuracies. In addition, the best accuracy in facial recognition system reached was had the following combination of parameters: size 12x9, normalized Euclidean distance, value of k equal to one and characteristic number equal to twenty. Also, the images with the smaller dimension analyzed (12x9) produced the best accuracies rates facial recognition than the others studied. RELATED WORK There are two basic approaches for face recognition [6]. The first is based on the extraction of feature vectors of the basic parts of a face, such as: eyes, nose, mouth and chin, the second approach is based on information theory concepts [6]. In the first approach, Shu Liao and Chung [9] proposed a new way to formulate the face recognition problem. Each facial image tested was deformed to a face image training in the fixed space for a predefined deformation model. However, in the second approach, Agarwal et. al. [1] it was used the PCA technique, which extracts the most important features of the face of an individual. These characteristics were used as inputs of a neural network classifier of the face. Therefore, this work uses the second approach, conducting an empirical study on the results of accuracies rates of a face recognition system based on the technical Eigenfaces and K-NN.