Diseño de un modelo basado en redes neuronales artificiales para la clasificación de palta hass
No hay miniatura disponible
Fecha
2020-10-28
Autores
Título de la revista
ISSN de la revista
Título del volumen
Editor
Pontificia Universidad Católica del Perú
DOI
Resumen
Perú se ha convertido en uno de los principales productores de palta Hass, en este aspecto una
etapa fundamental es la clasificación, esta situación conllevó al planteamiento del presente trabajo de
investigación el cual tuvo por objetivo diseñar un modelo basado en Redes Neuronales Artificiales que
permita la clasificación de dicha fruta considerando como criterios el estado de madurez fisiológica y
la evaluación de los daños y defectos que presente, dichas consideraciones están contempladas en la
Norma Técnica Peruana NTP 011.018-2018.
En la etapa inicial se diseñó un entorno controlado con un nivel de luminosidad frío-día, el cual
permitió la adquisición de imágenes, construyendo un dataset de 310 imágenes etiquetadas, sobre el
cual se aplicó Data Augmentation.
Luego se procedió a definir la parametrización de una arquitectura de red neuronal
convolucional, obteniendo un modelo de CNN sobre el cual se fueron evaluando 4 criterios, la
resolución de las imágenes de entrada, la cantidad de capas de convolución y pooling, el factor de
aprendizaje y la cantidad de épocas de entrenamiento.
Finalmente se mostraron los resultados obtenidos, definiendo la resolución de la imágenes de
entrada en 64 x 64 pixeles, 3 capas de convolución acompañas de pooling, con máscaras de 3x3 y 2x2
respectivamente y con funciones de activación ReLU, pasando luego a una capa capa fully connected,
la cual se conectó a una capa oculta y ésta a la capa de salida, la cual constó de 4 neuronas bajo la
representación One Hot Encoding, con una función de activación softmax, y un factor de aprendizaje
de 0.001, utilizando en su entrenamiento 50 épocas. Luego de evaluar el modelo parametrizado se
alcanzó una identificación correcta de las imágenes de palta Hass con una exactitud de 87.5%.
Peru has become one of the main producers of Hass avocado, in this aspect a fundamental stage is the classification, this situation led to the approach of this research work which aimed to design a model based on Artificial Neural Networks that allows the classification of said fruit considering as criteria the state of physiological maturity and the evaluation of the damages and defects that it presents, said considerations are contemplated in the Peruvian Technical Standard NTP 011.018-2018. In the initial stage, a controlled environment was designed with a cold-day light level, which allowed the acquisition of images, building a dataset of 310 labeled images, on which Data Augmentation was applied. Then we proceeded to define the parameterization of a convolutional neural network architecture, obtaining a CNN model on which 4 criteria were evaluated, the resolution of the input images, the number of convolution and pooling layers, the learning factor and the number of training seasons. Finally, the results obtained were shown, defining the resolution of the input images in 64 x 64 pixels, 3 convolution layers accompanied by pooling, with 3x3 and 2x2 masks respectively and with ReLU activation functions, then moving to a fully layer layer connected, which was connected to a hidden layer and this to the output layer, which consisted of 4 neurons under the One Hot Encoding representation, with a softmax activation function, and a learning factor of 0.001, using in its training 50 epochs. After evaluating the parameterized model, a correct identification of the Hass avocado images was achieved with an accuracy of 87.5%.
Peru has become one of the main producers of Hass avocado, in this aspect a fundamental stage is the classification, this situation led to the approach of this research work which aimed to design a model based on Artificial Neural Networks that allows the classification of said fruit considering as criteria the state of physiological maturity and the evaluation of the damages and defects that it presents, said considerations are contemplated in the Peruvian Technical Standard NTP 011.018-2018. In the initial stage, a controlled environment was designed with a cold-day light level, which allowed the acquisition of images, building a dataset of 310 labeled images, on which Data Augmentation was applied. Then we proceeded to define the parameterization of a convolutional neural network architecture, obtaining a CNN model on which 4 criteria were evaluated, the resolution of the input images, the number of convolution and pooling layers, the learning factor and the number of training seasons. Finally, the results obtained were shown, defining the resolution of the input images in 64 x 64 pixels, 3 convolution layers accompanied by pooling, with 3x3 and 2x2 masks respectively and with ReLU activation functions, then moving to a fully layer layer connected, which was connected to a hidden layer and this to the output layer, which consisted of 4 neurons under the One Hot Encoding representation, with a softmax activation function, and a learning factor of 0.001, using in its training 50 epochs. After evaluating the parameterized model, a correct identification of the Hass avocado images was achieved with an accuracy of 87.5%.
Descripción
Palabras clave
Redes neuronales (Computación), Procesamiento de imágenes, Palto