Sistema cuantificador de calidad de cultivo de manzana para monitoreo de la producción utilizando algoritmos de Aprendizaje Profundo con Visión Artificial y Segmentación de Instancias

Abstract

Nowadays, agriculture is an activity of marked influence in the economy world, therefore, in order to satisfy the progressive food needs, human beings have been introducing technological tools for the optimization of agricultural practices, this management is also known as Precision Agriculture (PA) Artificial Vision is a technology that has given greater support to Precision Agriculture (PA), granting a wide range of tools with the ability to reduce difficulties faced by the farmer during his hand labor. The aim of this project is to provide farmers a tool to improve the process for apple harvest management, by using Deep Learning (DL) algorithms and a Computer Vision system. The system development includes two study analyses: apple type detection and quality quantification for its inspection and validation using a non-invasive method. For apple type detection, SSD-MobileNet model was used and for apple quality segmentation, a fully convolutional network FCN-ResNet-18 was used. For both studies, networks were retrained with customized databases generated specifically for the development of this project. Lastly, evaluation parameters of the detection and segmentation systems are presented with metrics such as confusion matrices, and overlapping of objects on the IoU, respectively.