Agrupamento de faces em coleções de fotografias digitais.

Abstract

AS a result of the popularization of the Internet (motivated by social networks like Facebook, Orkut, MySpace) and the increasing proliferation of digital cameras and mobile devices, the automatic organization of large digital photo albums has become an extremely relevant resource. Traditional systems use only simple information (such as date, file and folder name) to help with the organization task. However, for large collections, typically formed by millions of images, this information is insufficient to achieve good leveis of organization and satisfaction. Most advanced techniques in this area aim to analyze image content and to extract high levei information, e.g., faces. In this sense, faces occupy a preponderant role, given their importance to human relations. Therefore, within the scope of photographs containing faces, face clustering is a very relevant topic. Within this context, this dissertation aims to address the problem of face clustering, while seeking: (i) to obtain better performance over the state of the art techniques in face clustering, and (ii) to investigate ways to minimize degradation usually associated with variations in face images (such as lighting, facial expressions and pose). The proposed approach to reach the above goals is composed of a preprocessing step followed by SURF (Speeded Up Robust Features) feature extraction and clustering steps. From an experimental study and statistical tests, in which the proposed approach and three commercial applications were compared, statistically significant differences between the generated results were inferred, with better results obtained by the proposed approach.