Contagem de laranjas em diferentes métodos

Abstract

This undergraduate thesis presents a comparative analysis of methods for the automatic counting of oranges in images, utilizing both classical computer vision algorithms and artificial intelligence models. The computer vision approach, implemented with the OpenCV and scikit- image libraries, involved image preprocessing to enhance the orange color and remove noise, followed by the detection of connected components representing the oranges. In contrast, an object detection model based on artificial intelligence, YOLO, was trained using an extensive dataset built and annotated with the help of the Roboflow platform. This dataset, comprising hundreds of images, enabled the training and evaluation of different YOLO model versions, including versions 5, 8, 9, and 11. Counting accuracy served as the primary metric for comparing the methods. The results demonstrate the superiority of the YOLO model compared to computer vision techniques, with an accuracy of 52.23% for YOLOv5 and reaching 58.08% for YOLOv9, highlighting the effectiveness of artificial intelligence in object counting tasks in images. However, the research identifies the need to optimize parameters such as confidence thresholds and the number of training epochs to maximize YOLO's performance.