Publication: DETECCIÓN Y SEGMENTACIÓN DE MÓDULOS PV EN IMÁGENES RGB AÉREAS PARA ANÁLISIS EN CONDICIÓN DE OPERACIÓN
Loading...
Date
2022-08
Authors
HURTADO ASTORGA, CLAUDIO SEBASTIÁN
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
En el presente trabajo se desarrolla un modelo basado en técnicas de machine learning para poder detectar y segmentar módulos fotovoltaicos en imágenes RGB. Tras un breve estudio del estado del arte en análisis de imágenes se opta por desarrollar un modelo basado en Mask R-CNN, ya que es capaz de detectar y segmentar objetos en imágenes. Para lograr este objetivo se genera un conjunto de ejemplos de 110 imágenes con ambientes o fondos complejos, segmentando manualmente los módulos PV dentro de estas, con las cuales se logra un entrenamiento efectivo del modelo desarrollado. A partir de este entrenamiento el modelo es capaz de detectar y segmentar paneles fotovoltaicos en ambientes complejos logrando una precisión de 99.6%, una puntuación F1 de 0.97 y un parámetro de pérdida de área segmentada “loss” del 15%, lo que demuestra la capacidad del modelo y su fiabilidad para detectar y segmentar imágenes futuras.
In the present work, a model based on machine learning techniques is developed to be able to detect and segment photovoltaic modules in RGB images. After a brief study of the state of the art in image analysis, it’s decided to develop a model based on Mask R-CNN, since it’s capable of detecting and segmenting objects in images. To achieve this goal, a set of examples of 110 images with complex backgrounds is generated, manually segmenting the PV modules in the images, with which an effective training of the developed model is achieved. Based on this training, the model is capable to do an instance segmentation of photovoltaic panels in complex environments, achieving an accuracy of 99.6%, an F1 score of 0.97 and a segmented area loss parameter call “loss” of 15%, which demonstrates the ability of the model and its reliability to detect and segment future images.
In the present work, a model based on machine learning techniques is developed to be able to detect and segment photovoltaic modules in RGB images. After a brief study of the state of the art in image analysis, it’s decided to develop a model based on Mask R-CNN, since it’s capable of detecting and segmenting objects in images. To achieve this goal, a set of examples of 110 images with complex backgrounds is generated, manually segmenting the PV modules in the images, with which an effective training of the developed model is achieved. Based on this training, the model is capable to do an instance segmentation of photovoltaic panels in complex environments, achieving an accuracy of 99.6%, an F1 score of 0.97 and a segmented area loss parameter call “loss” of 15%, which demonstrates the ability of the model and its reliability to detect and segment future images.
Description
Keywords
APRENDIZAJE AUTOMATICO , INTELIGENCIA ARTIFICIAL , FOTOVOLTAICO , CONVOLUTIONAL NEURAL NETWORK (CNN)