Evaluasi Mutu Green Beans Kopi Arabika dengan Model Deep Learning YOLO

Abstract

Proses penilaian mutu biji kopi untuk memperoleh biji kopi berkualitas baik masih dilakukan secara manual. Proses tersebut cukup menyita waktu, sehingga memungkinkan terjadinya kesalahan akibat kelelahan mata pekerja. Tujuan penelitian ini membuat model deep learning untuk mengevaluasi mutu green beans kopi arabika menggunakan algoritma You Only Look Once (YOLO). Penelitian ini terdiri dari beberapa tahapan: (1) pengumpulan dataset, (2) pelabelan dataset, (3) training dataset, (4) uji fungsional model deep learning, dan (5) analisis hasil uji fungsional model untuk mengetahui nilai akurasi, presisi, recall, dan F1-Score. Training deep learning terhadap 11.591 dataset asli dilakukan menggunakan algoritma YOLOv5m dan YOLOv8m. Jumlah epoch yang digunakan pada algoritma YOLOv5 sebanyak 330, sedangkan algoritma YOLOv8 sebanyak 121 epoch. Nilai akurasi, presisi, dan recall pada algoritma YOLOv5 masing-masing sebesar 90,49%; 58,24%; dan 53,61%. Nilai akurasi, presisi, dan recall pada algoritma YOLOv8 sebesar 91,56%; 62,09%; dan 63,10%. Ditinjau dari nilai F1-Score algoritma YOLOv8 lebih baik daripada YOLOv5, yaitu sebesar 0,56%. YOLOv8 menggunakan nilai epoch yang sedikit dapat memberikan nilai akurasi, presisi, recall dan F1-Score yang lebih tinggi daripada YOLOv5. Hal tersebut dikarenakan adanya pembaharuan fitur dataloader mosaic dan anchor-free detection system pada algoritma YOLOv8 sehingga dapat meningkatkan kinerja model. Namun, kedua algoritma YOLO tersebut masih kurang baik dalam mendeteksi objek-objek kecil.

The process of assessing coffee beans quality to obtain good coffee beans is still carried out manually. The process is quite time-consuming, making it possible for errors due to workers' eyestrain. The purpose of this study was to create a deep learning model to evaluate the quality of arabica coffee green beans using the You Only Look Once (YOLO) algorithm. This research consists of several stages: (1) dataset collection, (2) dataset labeling, (3) dataset training, (4) deep learning model functional test, and (5) analysis of model functional test results to determine the value of accuracy, precision, recall, and F1-Score. Deep learning training on 11,591 original datasets was conducted using YOLOv5m and YOLOv8m algorithms. The number of epochs used in the YOLOv5 algorithm is 330, while the YOLOv8 algorithm is 121 epochs. The accuracy, precision, and recall values of the YOLOv5 algorithm were 90.49%; 58.24%; and 53.61%, respectively. The accuracy, precision, and recall values of the YOLOv8 algorithm were 91.56%; 62.09%; and 63.10%. Judging from the F1-Score value, the YOLOv8 algorithm is better than YOLOv5, which is 0.56%. YOLOv8 using a small epoch value can provide higher accuracy, precision, recall, and F1-Score than YOLOv5. This is due to the update of the mosaic data loader feature and anchor-free detection system on the YOLOv8 algorithm so that it can improve model performance. However, both YOLO algorithms are still not good at detecting small objects.