Procesamiento de imágenes de foto-trampeo con técnicas de aprendizaje profundo para el descarte automático de imágenes vacías
Fecha
2022-07-01
Autores
Título de la revista
ISSN de la revista
Título del volumen
Editor
Jaén: Universidad de Jaén
Resumen
Se propone el diseño de un procedimiento automático que permita determinar si en imágenes de foto-trampeo
(tomadas de forma automática ante la presencia de movimiento) aparecen animales o no, descartando aquellas
en las que no estuviesen presentes. Para ello, se incorpora un sistema de clustering que permite separar la base
de datos en siete grupos de imágenes con características similares. Sobre cada grupo, un modelo de red
neuronal, concretamente un RAE (robust autoencoder), será el encargado de reconstruir imágenes vacías. Bajo la
premisa de que el error de reconstrucción será mayor en imágenes con animales, se ha implementado un
modelo de red neuronal densamente conectada que, tomando dicho error, actúa como clasificador, etiquetando
la imagen como vacía o con animales.
The design of an automatic procedure is proposed to determine if in photo-tramping images (taken automatically in the presence of movement) animals appear or not, discarding those in which they were not present. In order to do this, a clustering system that allows the database to be separated into seven groups of images with similar characteristics is developed. In each group, a neural network model, specifically a RAE (robust autoencoder), will be in charge of reconstructing empty images. Under the premise that the reconstruction error will be greater in images with animals, a densely-connected neural network model has been implemented so that, taking this error, acts as a classifier, labeling the image as empty or with animals.
The design of an automatic procedure is proposed to determine if in photo-tramping images (taken automatically in the presence of movement) animals appear or not, discarding those in which they were not present. In order to do this, a clustering system that allows the database to be separated into seven groups of images with similar characteristics is developed. In each group, a neural network model, specifically a RAE (robust autoencoder), will be in charge of reconstructing empty images. Under the premise that the reconstruction error will be greater in images with animals, a densely-connected neural network model has been implemented so that, taking this error, acts as a classifier, labeling the image as empty or with animals.
Descripción
Palabras clave
Tecnologías de la Información / Tratamiento inteligente de la información