Evaluación empírica de la robustez de diferentes redes neouronales usadas para la detección de objetos.
Date
2019
Journal Title
Journal ISSN
Volume Title
Publisher
Universidad Tecnológica Nacional Regional Córdoba.
Abstract
Existen muchos algoritmos para la detección de objetos en imágenes pero dependiendo de las
necesidades computacionales, velocidad de respuesta y condiciones de trabajo, resulta difícil seleccionar
el que se ajuste a los requerimientos particulares. En este trabajo se presenta la evaluación de diferentes
redes neuronales convoluciones aplicadas a la detección de objetos. Se exploran sus comportamientos
bajo diferentes condiciones: cambios en los tamaños de objetos a detectar, en la iluminación y en los
ambientes. Se evalúan sus tiempos de cómputo y la posibilidad de su uso en tiempo real. Los resultados
demuestran la factibilidad del uso de estas redes para detección de objetos en ambientes industriales
pero de los experimentos surgen una serie de conclusiones sobre las condiciones de funcionamiento necesarias para lograr resultados óptimos. Estas están relacionadas con la red a usar dependiendo de la
velocidad, las condiciones de iluminación, el tamaño de los objetos y el entorno de trabajo. A futuro, se
espera que alguno de estos algoritmos sea utilizado como parte de un sistema de seguridad industrial.
There are many algorithms for object detection in images but depending on the computatio nal needs, response speed and working conditions, it is difficult to select the one that fits the particular requeriments. In this work is presented the evaluation of different convolutional neural networks applied to object detection. Their behaviors under different conditions are explored: changes in the size of object to be detected, in illuminations and in environments. Its computation times and the posibility of his use in real time are evaluated. The results demostrates the feasibility of using these networks for object de tection in industrial environments but from the experiments a series of conclusions about the operating conditions necessary to achieve optimal results arise. These are related to the network to be used depen ding on the speed, illumination conditions, object sizes and work environments. In the future, some of these algorithms are expected to be used as part of an industrial security system
There are many algorithms for object detection in images but depending on the computatio nal needs, response speed and working conditions, it is difficult to select the one that fits the particular requeriments. In this work is presented the evaluation of different convolutional neural networks applied to object detection. Their behaviors under different conditions are explored: changes in the size of object to be detected, in illuminations and in environments. Its computation times and the posibility of his use in real time are evaluated. The results demostrates the feasibility of using these networks for object de tection in industrial environments but from the experiments a series of conclusions about the operating conditions necessary to achieve optimal results arise. These are related to the network to be used depen ding on the speed, illumination conditions, object sizes and work environments. In the future, some of these algorithms are expected to be used as part of an industrial security system
Description
Keywords
Redes neuronales convolucionales, Detección de objetos
Citation
Mecánica Computacional Vol.XXXVII,2019.
Endorsement
Review
Supplemented By
Referenced By
Creative Commons license
Except where otherwised noted, this item's license is described as info:eu-repo/semantics/openAccess