Una gramática visual para la detección de rostros

Several methods for face detection have been developed with certain success, these methods typically include features like texture, skin color, some predefined templates or deformable templates, etc., however these tend to fail under “difficult” conditions such as partial occlusions and changes in o...

ver descrição completa

Detalhes bibliográficos
Autor: AUGUSTO MELÉNDEZ TEODORO
Formato: tesis de maestría
Estado:Versión aceptada para publicación
Fecha de publicación:2011
País:México
Recursos:Instituto Nacional de Astrofísica, Óptica y Electrónica
Repositorio:Repositorio Institucional del INAOE
Idioma:español
OAI Identifier:oai:inaoe.repositorioinstitucional.mx:1009/703
Acesso em linha:http://inaoe.repositorioinstitucional.mx/jspui/handle/1009/703
Access Level:acceso abierto
Palavra-chave:info:eu-repo/classification/Redes bayesianas/Bayesian networks
info:eu-repo/classification/Gramáticas/Grammars
info:eu-repo/classification/Reconocimiento facial/Face recognition
info:eu-repo/classification/cti/1
info:eu-repo/classification/cti/12
info:eu-repo/classification/cti/1203
Descrição
Resumo:Several methods for face detection have been developed with certain success, these methods typically include features like texture, skin color, some predefined templates or deformable templates, etc., however these tend to fail under “difficult” conditions such as partial occlusions and changes in orientation and illumination. We propose a novel technique for face detection based on a visual grammar. We first define a symbol relational grammar for faces, representing the visual elements of a face and their spatial relations. This grammar is then transformed to a Bayesian network representation. The structure of the Bayesian network is derived from the grammar, and its parameters are obtained from data, i.e., from positive and negative examples of faces. Then the Bayesian network is used for face detection via probabilistic inference, using as evidence a set of weak detectors for different face components. We evaluated our method on a set of sample images of faces under “difficult” conditions, and contrasted it with a simplified model without spatial relationships, and the AdaBoost face detector. The results show a significant improvement when using our method based on a visual grammar. Although the grammar is restricted to the representation of faces, it is possible to extend it to represent a complete person or other object.