Aprendizaje Incremental para la clasificación de objetos en un mundo real y dinámico

Classifiers have significantly increased their performance with the advent of deep learning methods. However, these classifiers require a large number of examples to operate, tend to be fragile in the presence of out-of-distribution examples, and are static in the sense that they are unable to learn...

Descripción completa

Detalles Bibliográficos
Autor: Yareli Aburto Sanchez
Tipo de recurso: tesis de maestría
Estado:Versión aceptada para publicación
Fecha de publicación:2024
País:México
Institución:Instituto Nacional de Astrofísica, Óptica y Electrónica
Repositorio:Repositorio Institucional del INAOE
Idioma:español
OAI Identifier:oai:inaoe.repositorioinstitucional.mx:1009/2529
Acceso en línea:http://inaoe.repositorioinstitucional.mx/jspui/handle/1009/2529
Access Level:acceso abierto
Palabra clave:info:eu-repo/classification/Incremental Learning/Incremental Learning
info:eu-repo/classification/Machine Learning/Machine Learning
info:eu-repo/classification/Convolutional Networks/Convolutional Networks
info:eu-repo/classification/Hierarchical Tree/Hierarchical Tree
info:eu-repo/classification/Classifier/Classifier
info:eu-repo/classification/Imagery/Imagery
info:eu-repo/classification/Object Classification/Object Classification
info:eu-repo/classification/Support vector machine/Support vector machine
info:eu-repo/classification/Open World/Open World
info:eu-repo/classification/cti/7
info:eu-repo/classification/cti/33
info:eu-repo/classification/cti/3399
Descripción
Sumario:Classifiers have significantly increased their performance with the advent of deep learning methods. However, these classifiers require a large number of examples to operate, tend to be fragile in the presence of out-of-distribution examples, and are static in the sense that they are unable to learn to classify examples of new classes. Incremental learning has been developed as a way to gradually increase the number of classes supported by a classifier, without degrading its performance on previously learned classes. However, they still require a large number of examples to train to know when new classes are given to the classifier. In this thesis, we propose a new incremental learning classifier that can be trained with a single unlabeled example, does not need to know whether the example comes from an existing or known class, and is able to automatically obtain new instances and the name of the example class. Specifically, the proposed system uses a segmentation algorithm to obtain information of possible objects from an initially unknown environment, automatically obtains training examples and the class name of the segmented objects from the Internet, and builds an incremental classifier, using a hierarchical k-means algorithm and support vector machines. It is shown that the proposed system is able to operate under these open-world conditions with competitive performance under open-world conditions with competitive performance.