Aprendizaje de conceptos visuales basado en múltiples clasificadores

Object recognition is usually based on learning from a large dataset of previously selected training images; however, not all objects have one associated dataset. Nowadays it is possible to find images on Internet of virtually any object, only by launching a query with the object's name in a se...

Descripción completa

Detalles Bibliográficos
Autor: DULCE JAZMIN NAVARRETE ARIAS
Tipo de recurso: tesis de maestría
Estado:Versión aceptada para publicación
Fecha de publicación:2012
País:México
Institución:Instituto Nacional de Astrofísica, Óptica y Electrónica
Repositorio:Repositorio Institucional del INAOE
Idioma:español
OAI Identifier:oai:inaoe.repositorioinstitucional.mx:1009/763
Acceso en línea:http://inaoe.repositorioinstitucional.mx/jspui/handle/1009/763
Access Level:acceso abierto
Palabra clave:info:eu-repo/classification/Reconocimiento de objetos/Object recognition
info:eu-repo/classification/Funciones/Sift features
info:eu-repo/classification/Características locales/Local features
info:eu-repo/classification/Affine transforms/Affine transforms
info:eu-repo/classification/Filtros Gabor/Gabor filters
info:eu-repo/classification/cti/1
info:eu-repo/classification/cti/12
info:eu-repo/classification/cti/1203
Descripción
Sumario:Object recognition is usually based on learning from a large dataset of previously selected training images; however, not all objects have one associated dataset. Nowadays it is possible to find images on Internet of virtually any object, only by launching a query with the object's name in a search engine. Nevertheless, this strategy introduces computational challenges: (i) the object's name can have more than one meaning; (ii) the object can have different appearances, and (iii) without prior knowledge of the object it is hard to identify which visual features to use in order to train a classifier. In this thesis a method to tackle the above problem is proposed. The method incorporates an ensemble of classifiers that builds several object models treating intraclass variability. Responses of every classifier are combined to determine the presence or absence of the object. The method begins with a small training set obtained via Web, and a series of image transformations is applied in order to identify invariant features. The weight of the local and global features is optimized, in order to recognize different category types. We also develop an algorithm based on sliding windows to identify the object's position within an image. We evaluated our method on images from the Web, the Caltech-7 dataset and real environments. We compared our method against related work, obtaining competitive classification performance in the recognition of general and specific objects.