Servicios Personalizados
Revista
Articulo
Indicadores
Citado por SciELO
Accesos
Links relacionados
Similares en SciELO
Compartir
Ingeniería, investigación y tecnología
versión On-line ISSN 2594-0732versión impresa ISSN 1405-7743
Resumen
GANTE-DIAZ, Saulo Abraham et al. Integration of a multispectral camera and machine learning for apple sorting. Ing. invest. y tecnol. [online]. 2022, vol.23, n.4, e1935. Epub 01-Mayo-2023. ISSN 2594-0732. https://doi.org/10.22201/fi.25940732e.2022.23.4.031.
This paper presents a system capable of performing a binary classification (ripe and rotten) of red apples, which is achieved through the use of the Normalized Difference Vegetation Index (NDVI) and an Xception Neural Network. In order to carry out this process a multispectral camera is used to observe details outside the scope of the human eye, reducing the presence of errors during their classification. The selection of NDVI is the result of its comparison with the GNDVI, GVI, NDRE, NDVIR, NG, NGRDI, and RVI vegetation indices applied to an image bank obtained by means of the multispectral camera. Furthermore, classification results are displayed when using Xception, ResNet, and MobileNet neural networks, thus justifying the use of the Xception network. Finally, the instrumentation and lighting used in a prototype, which emulates the actual classification process using a conveyor belt, are described. This method allows to experimentally validate the proposed classification system, in this case, 73 % success in the online classification was achieved. The development of this work is based on the following methodology: a multispectral camera is used to create a database, the obtained images are sent to a processing stage, which involves alignment and RGB reconstruction. Subsequently, two or more bands are used for the calculation, comparison and analysis of different vegetation indices. Once the vegetation index is established, the training and comparison among different neural network architectures is carried out. Regarding the training stage, transfer learning is used to reduce the need for a large database. Finally, experimental tests are carried out to validate the behaviour.
Palabras llave : Apple sorting; deep learning; machine learning; multispectral camera; vegetation index.