Comparação de métodos descritores e métodos de aprendizado de máquina aplicados no reconhecimento de imagens de frutas

Abstract

The quality of a fruit is an important factor for producers and consumers. Manual separation, that is a harvesting or post-harvest stage is a great labor effort, and if not carried out correctly, can cause damage to both sides. Computer vision plays an important role in making fruit classifications, based on features extracted from images, with techniques that are continuously improved over time. This work proposes to compare different classification methods, combining them with different characteristics extracted from the images, such as color, shape and texture, in order to identify the best combination, and achieve the best possible accuracy. In this way, this work portrays the use of algorithms descriptors of images in fruits, with the objective of extracting their characteristics. With the aid of computer vision, the tests and the training are carried out based on the characteristics of each image, aiming at a comparison between the proposed methods that best help in the classification of a particular fruit. The set of images used is called Fruits-360, obtained through the work of Muresan e Oltean (2018), where 5279 images of fruits that have similarities between them, such as round and red, were selected. The extraction of characteristics from fruit images is based on algorithms that describe color (GCH and CCV), shape (Hu Moments) and texture (LBP and Haralick). It was found that the IBK classification algorithm obtained, in most experiments, the best result of accuracy and standard deviation for similar fruit images.