Aplicación de técnicas de Inteligencia Artificial para identificar el porcentaje de suciedad en lotes de aceitunas en almazaras
Fecha
2022-07-01
Autores
Título de la revista
ISSN de la revista
Título del volumen
Editor
Jaén: Universidad de Jaén
Resumen
El objetivo de este trabajo es identificar y estimar el porcentaje de suciedad en lotes de aceitunas en almazaras, mediante
técnicas de inteligencia artificial. Para ello, se han probado técnicas clásicas de análisis sobre imágenes de infrarrojo
cercano (NIR), que alcanzan resultados aceptables en imágenes con pocos elementos, pero que no funcionan
adecuadamente en imágenes complejas. Por ello, se ha recurrido a distintas técnicas de segmentación y clasificación a
partir de imágenes RGB, desde métodos no supervisados (k-means) a métodos supervisados crecientes en complejidad:
mínima distancia, aprendizaje automático (árboles de decisión, k-NN y SVM), redes neuronales de tipo perceptrón (MLP)
y redes neuronales convolucionales (CNN). Los resultados son progresivamente mejores, hasta llegar a las redes
convolucionales que proporcionan una exactitud superior al 90% y una baja confusión entre aceitunas, hojas (suciedad)
y fondo. A partir de ahí se ha realizado una aplicación para estimar el porcentaje de suciedad.
The main goal of this work is to identify and estimate the percentage of dirt in batches of olives in oil mills, using artificial intelligence techniques. Then, classical analysis techniques have been tested on near infrared (NIR) images, which achieve acceptable results in images with few elements, but do not work properly in complex images. For this reason, different segmentation and classification techniques have been used with RGB images, from unsupervised methods (k-means) to supervised methods that are increasingly complex: minimum distance, machine learning (decision trees, k-NN and SVM), perceptron neural networks (MLP) and convolutional neural networks (CNN). The results are progressively better, until reaching the convolutional networks that provide an accuracy higher than 90% and a low confusion between olives, leaves (dirt) and background. From there, an application has been made to estimate the percentage of dirt.
The main goal of this work is to identify and estimate the percentage of dirt in batches of olives in oil mills, using artificial intelligence techniques. Then, classical analysis techniques have been tested on near infrared (NIR) images, which achieve acceptable results in images with few elements, but do not work properly in complex images. For this reason, different segmentation and classification techniques have been used with RGB images, from unsupervised methods (k-means) to supervised methods that are increasingly complex: minimum distance, machine learning (decision trees, k-NN and SVM), perceptron neural networks (MLP) and convolutional neural networks (CNN). The results are progressively better, until reaching the convolutional networks that provide an accuracy higher than 90% and a low confusion between olives, leaves (dirt) and background. From there, an application has been made to estimate the percentage of dirt.
Descripción
Palabras clave
Automática