Investigação do problema de detecção de faces com variações de orientação.

Abstract

In this thesis, the problem of detecting faces that present high variations of orientation is investigated. Some factores were identified that may infiuence the detection results when some evaluation metrics are used. For example, if the applied metric takes in consideration the detected áreas obtained by the classifiers and the human labeled áreas (groundtruth), the way as the detected images are marked will interfere in the computed results. In relation to the face image cropping aspect, the experimental results show that if externai regions of the faces are included for training, the detection results will be better. To deal with ali those factors, it was proposed and implemented an approach to face detection that explores the invariance by training to yield classifier tree with lower computational complexity than other approaches in the state of the art, and able to deal with high angle in-plane orientations. To make the training of the cascades of classifiers feasible, a hybrid parallel approach of the training method of Viola e Jones (2004) was proposed. The parallel approach is able to achieve superlinear speedup, as it was demonstrated in the experiments. The face detection approach obtained higher results than those obtained by Rowley, Baluja e Kanade (1998a), and Viola e Jones (2004). Only one of the evaluated approaches obtained higher results, that proposed by Huang et al. (2007). However, the approach proposed in this thesis has lower computational complexity in terms of quantity of leveis in the classifier tree, and quantity of processing nodes.