REAL-TIME TOMATO QUALITY DETECTION SYSTEM USING YOU ONLY LOOK ONCE (YOLOv7) ALGORITHM

Sistem Deteksi Mutu Tomat Secara Real-time Menggunakan Algoritma You Only Look Once (YOLOv7)

Abstract

Real-time object detection is a crucial aspect of computer vision. With the increasing prominence of the big data field, it has become easier to gather data from various sources. Over the past few decades, computer vision inspection systems have become essential tools in agricultural operations, and their usage has seen a significant rise. Computer vision automation-based technology in agriculture is increasingly being employed to enhance productivity and efficiency. Tomato is a widely utilized crop commodity, finding applications in food, cosmetics, and pharmaceuticals. Consequently, tomato farming continues to evolve and has become one of the nation's export commodities. YOLO is an algorithm capable of real-time object detection and recognition. In this study, the YOLOv7-tiny architecture, which has lower computational overhead, was utilized. For quality detection of tomatoes, they were categorized into three classes: ripe, unripe, and defective. The trained model yielded a recall score of 0.97, precision of 1.0, a PR-curve of 0.838, and an F1-score of 0.81, indicating that the model learned effectively. The research achieved an accuracy of 90.6% on original images with an average IoU of 0.90 and a detection time of 2.7 seconds. In images with added light disturbance, the average accuracy was 91.2%. Images with reduced light yielded an average accuracy of 92%, while images with blur disturbance had an average accuracy of 78.2%. In real-time testing, ripe tomatoes were detected up to a maximum distance of 90cm, unripe tomatoes at 90cm, and defective tomatoes at 70cm.