Construcción de una Base de Datos de Frutas Cítricas A Partir del Aumento de Datos para un Modelo de Clasificación con Aprendizaje Profundo

Abstract

Every day millions of images are stored, either by websites or by social networks, generating datasets with labels. However, they are not always representative, of great variety or in the public domain. This article presents the construction of a particular dataset of citrus fruits distributed in Colombia that are sparsely found in repositories, these information are the input records to a single-label classification model with deep learning. The images were obtained through the burst option on a 48 MP camera (f2.0 with Quad Pixel technology, 1.6 µm). Exposure conditions such as position were evaluated through rotation and lighting of the fruit with the purpose to obtain the increase of records in the dataset, previous defining the features that identify each of the fruits, the selection of a representative limited quantity is made. Through the use of different techniques, a spatial resizing is applied to the images at the software level. In this way, 20738 samples belonging to 8 classes of citrus fruits are obtained and they are separated by implementing the division rule 90% to 10%, in two directories called; train and test.