Sistema de visión artificial para el reconocimiento de enfermedades y plagas en hojas de yuca (Manihot esculenta Crantz) por medio de redes neuronales convolucionales
Loading...
Date
2021
Authors
Gomez Pupo, Santiago Maria
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Detectar las enfermedades en los cultivos de yuca a tiempo, puede hacer la diferencia para obtener ganancias económicas, fortaleciendo la seguridad alimentaria en regiones de Colombia y el mundo donde la yuca es un cultivo de interés agrícola. Es por eso, que en este trabajo proponemos un sistema de visión artificial para el reconocimiento de enfermedades y plagas de la yuca, por medio de redes neuronales convolucionales como una solución que puede ayudar a prevenir las perdidas económicas, evitando la propagación y favoreciendo la toma de decisiones para un manejo adecuado de estas enfermedades. Las redes neuronales convolucionales están a la vanguardia en el reconocimiento de imágenes para tareas complejas de visión artificial, es una técnica que ha demostrado ser muy eficientes en comparación con las redes neuronales ordinarias. El modelo propuesto en el presente trabajo, basado en este tipo de redes, se entrenó con un conjunto de datos formado a partir de dos bases de datos provenientes de dos competencias de la página web kaggle, constituido por 5 categorías con enfermedades en la hoja de la yuca identificadas como: añublo bacterial, rayado marrón, acaro verde, mosaico y hojas sanas. El entrenamiento se llevó a cabo implementando el enfoque de transferencia de aprendizaje, muy recomendado cuando la cantidad de imágenes es limitada. Además, permite alcanzar un buen rendimiento de la red. De los tres modelos seleccionados por haber obtenido un buen rendimiento en problemas de clasificación de enfermedades en plantas, según los antecedentes consultados, el mejor fue Xception entrenado durante un periodo de 35 épocas con 6120 imágenes de hojas de yuca, logrando una exactitud (accuracy) de 94,56 %. Este modelo proporciona una opción para detectar las enfermedades de la hoja de la yuca in situ de manera temprana, confiable y a bajo costo.
Detecting diseases in cassava crops in time can make a di_erence to obtain economic gains, strengthening food security in regions of Colombia and the world where cassava is a crop of agricultural interest. That is why in this work we propose an arti_cial vision system for the recognition of cassava diseases and pests, by means of convolutional neural networks as a solution that can help prevent economic losses, avoiding the spread and favoring the taking decision-making for an adequate management of these diseases. Convolutional neural networks are at the forefront of image recognition for complex machine vision tasks, it is a technique that has proven to be very e_cient compared to ordinary neural networks. The model proposed in the present work, based on this type of networks, was trained with a data set formed from two databases from two competencies on the kaggle website, consisting of 5 categories with diseases in the sheet of cassava identi_ed as: bacterial blight, Brown streaking, green mite, mosaic, and healthy leaves. The training was carried out implementing the learning transfer approach, highly recommended when the number of images is limited. In addition, it allows to achieve good network performance. Of the three models selected for having obtained a good performance in problems of classi_cation of diseases in plants, according to the consulted antecedents, the best was Xception trained during a period of 35 epochs with 6,120 images of yucca leaves, achieving an accuracy of 94.56 %. This model provides an option to detect cassava leaf diseases in situ early, reliably and at low cost.
Detecting diseases in cassava crops in time can make a di_erence to obtain economic gains, strengthening food security in regions of Colombia and the world where cassava is a crop of agricultural interest. That is why in this work we propose an arti_cial vision system for the recognition of cassava diseases and pests, by means of convolutional neural networks as a solution that can help prevent economic losses, avoiding the spread and favoring the taking decision-making for an adequate management of these diseases. Convolutional neural networks are at the forefront of image recognition for complex machine vision tasks, it is a technique that has proven to be very e_cient compared to ordinary neural networks. The model proposed in the present work, based on this type of networks, was trained with a data set formed from two databases from two competencies on the kaggle website, consisting of 5 categories with diseases in the sheet of cassava identi_ed as: bacterial blight, Brown streaking, green mite, mosaic, and healthy leaves. The training was carried out implementing the learning transfer approach, highly recommended when the number of images is limited. In addition, it allows to achieve good network performance. Of the three models selected for having obtained a good performance in problems of classi_cation of diseases in plants, according to the consulted antecedents, the best was Xception trained during a period of 35 epochs with 6,120 images of yucca leaves, achieving an accuracy of 94.56 %. This model provides an option to detect cassava leaf diseases in situ early, reliably and at low cost.
Description
Keywords
Citation
Gomez Pupo, S. (2021). Sistema de visión artificial para el reconocimiento de enfermedades y plagas en hojas de yuca (Manihot esculenta Crantz) por medio de redes neuronales convolucionales (tesis de maestría). Universidad Tecnológica de Bolívar.