Reconhecimento de faces em poses arbitrária por meio de descritores locais

Abstract

Reliable face recognition still stands as a challenge to current automatic face recognition systems. Images acquired from subjects in uncontrolled environments may be strongly affected by numerous factors such as lighting variation, changes in pose and occlusion. This difficulty is due to a number of factors that occur in the images acquired in uncontrolled environments, such as change in illumination, rotation and occlusion. Among those, pose variation is one of the most difficult problems to be overcome, and classical face recognition algorithms assume that facial features are not parially or completely occluded. This works proposes an approach based on local descriptors. The pose variation have a significant effect on performance of the algorithms because most of approaches are based on features of the face that can be partially or completely hidden by rotation. In this paper, we propose an approach based on canonical three-dimensional models of the head and local descriptors for face recognition under arbitrary poses. The developed approach was evaluated in two areas of the face recognition process: identification and verification. Results show that this approach can be coupled in the current face recognition systems for successful face recognition in arbitrary poses.