Visão computacional aplicada a análise de frutos de C. Arabica

Abstract

At coffee research centers, various traits are phenotyped by researchers. Some of these traits are directly determined by fruit phenotyping, such as ripening. Fruit ripening is an important trait to measure, because it allows for cultivars release with different ripening cycles, which is essential for farmers as it allows for the scaling of production and maximization of efficiency and profitability. However, measuring this trait in breeding programs presents several challenges. This study was divided into three chapters that present a comprehensive evaluation of coffee fruit and aspects associated with the selection and evaluation of ripening in Coffea arabica. In the first chapter, coffee fruits obtained from a phenotyping platform were thoroughly evaluated, by examining their morphological and color characteristics using computer vision. To achieve it, a classification model based on convolutional neural networks was created to classify the different stages of ripening. In the second chapter, images were synthesized from the generated image dataset to train a computer vision model based on the YOLO neural network architecture for direct classification and detection of coffee fruits in numerous scenarios and environments. In chapter 3, the objective was to establish the ideal sample size for ripening fruit evaluation and to verify the associated errors in adopting each sample size, as well as to demonstrate that the K-means clustering method can be an alternative to assist researchers in making decisions about the constituent genotypes of the breeding population. Detailed analysis was conducted on fruits from 21 cultivars, providing valuable information to researchers about their morphological and color characteristics. A total of 36.879 images of coffee fruits at different ripening stages were created. The use of the YOLO architecture allows for the direct evaluation of coffee fruits in different scenarios and environments, reducing and facilitating the process of phenotyping the trait. It was found that samples larger than 500 ml of fruits demonstrate an excellent sample size, and the use of the Kmeans technique to group data into different ripening cycles can be an excellent alternative for researchers, allowing for precise and efficient analysis.