Método de reconhecimento de face baseado em estatísticas de ordem superior

Abstract

Face recognition is one of the most effective non-intrusive methods in biometrics. On the other hand, it is a major challenge for researchers in the area, as it involves factors that include the ambient lighting aspects, the individual pose, image quality, occlusion, disguises, among others. Face recognition systems have broad applicability, especially the security systems, performing important task in society. It is key access to systems and locations, for example, personal computers, smartphones, access to specific rooms of the banking system, among other systems linked to human-machine interface. This master’s thesis presents contributions in two aspects: (i) it explores the hither-order statistics to build compact signature of faces; (ii) it considers a scenario whose the goal is to detect and identify criminals automatically with face recognition to assist the military police. The algorithm proposed in this thesis for face recognition was developed in three stages. In the first stage the feature extraction using higher-order statistics (second-, third- and fourth-order cumulants) is performed. The next step comprises the feature selection, through the Fisher’s discriminant, and redundancy analysis with linear correlation. In the last stage, the classification using the Bayes classifier is performed. To check the performance of face recognition algorithm proposed in this work it was carried out tests using the database ORL, which is a well-known dataset in the image processing area. Promising results were achieved in which detection and classification rates over 70% were reached, which shows the potential of higher-order statistics on building compact feature vector signatures of faces.