Deteksi dan Klasifikasi Cacat Biji Kakao Varietas Forastero Berbasis Multi-Class Object Detection Menggunakan YOLO V8

Fauzan Abhirama, Anggoro Cahyo Sukartiko, S.T.P., M.P., Ph.D.; Dr. Ir. Radi, S.T.P., M.Eng., IPU, ASEAN Eng.

2026 | Skripsi | TEKNOLOGI INDUSTRI PERTANIAN

Proses sortasi biji kakao yang masih dilakukan secara manual memiliki kelemahan berupa penilaian yang bersifat subjektif, kurang konsisten, dan memerlukan waktu yang lama. Selain itu, keberadaan berbagai jenis cacat biji kakao seperti biji berjamur, biji berkecambah, biji berserangga, biji pipih dan kotoran/waste menyebabkan proses identifikasi dan klasifikasi semakin kompleks. Oleh karena itu, penelitian ini bertujuan untuk memperoleh ukuran model deteksi dan klasifikasi terbaik dari cacat biji kakao berbasis YOLOv8 serta mendapatkan konfigurasi pelatihan model terbaik dalam mengklasifikasikan cacat biji kakao untuk mendukung proses sortasi biji kakao secara otomatis dan efisien. Penelitian dilakukan melalui tahapan pengambilan citra biji kakao menggunakan webcam yang dipasang pada Color Assessment Cabinet, kemudian dilanjutkan dengan preprocessing data berupa labeling, resizing, splitting dataset, dan augmentasi data. Dataset yang digunakan terdiri dari 1.019 citra dengan total 14.456 anotasi objek yang merepresentasikan beberapa kategori cacat biji kakao dan biji normal. Model YOLOv8 yang digunakan terdiri atas beberapa varian ukuran model, yaitu YOLOv8n, YOLOv8s, YOLOv8m, YOLOv8l, dan YOLOv8x. Evaluasi model dilakukan menggunakan parameter akurasi, presisi, recall, dan mean average precision (mAP) untuk mengukur kemampuan model dalam mendeteksi dan mengklasifikasikan objek secara tepat. Hasil penelitian menunjukan bahwa model terbaik diperoleh pada YOLOv8m menggunakan Dataset Augmentasi. Kemudian, dilakukan penyetelan hyperparameter berupa batch size, learning rate, dan epoch untuk memperoleh performa terbaik. Hasil penyetelan hyperparameter pada model YOLOv8m, diperoleh pada penggunaan batch size 24, learning rate 0.01, dan epochs 100. Konfigurasi tersebut menghasilkan nilai akurasi sebesar 98,0%, presisi sebesar 99,2%, recall sebesar 98,7%, mAP50 sebesar 99,3%, dan mAP50-95 sebesar 92,2%. Penerapan teknologi ini diharapkan dapat meningkatkan efisiensi sortasi, mengurangi ketergantungan pada pengamatan manusia, serta mendukung pengendalian mutu biji kakao berbasis kecerdasan buatan. Hasil penelitian juga diharapkan menjadi referensi pengembangan teknologi otomatisasi pada industri pengolahan kakao.

The cocoa bean sorting process, which is still carried out manually, has drawbacks such as subjective assessment, a lack of consistency, and time-consuming procedures. In addition, the presence of various types of cocoa bean defects such as moldy beans, sprouted beans, insect-infested beans, flattened beans, and impurities or waste makes the identification and classification process even more complex. Therefore, this study aims to determine the optimal detection and classification model for cocoa bean defects based on YOLOv8 and to identify the best model training configuration for classifying cocoa bean defects to support an automated and efficient cocoa bean sorting process. The research was conducted by first capturing images of cocoa beans using a webcam mounted on a Color Assessment Cabinet, followed by data preprocessing, including labeling, resizing, splitting the dataset, and data augmentation. The dataset consisted of 1,019 images with a total of 14,456 object annotations representing various categories of cocoa bean defects and normal beans. The YOLOv8 model used consists of several model size variants, namely YOLOv8n, YOLOv8s, YOLOv8m, YOLOv8l, and YOLOv8x. Model evaluation was conducted using the metrics of accuracy, precision, recall, and mean average precision (mAP) to measure the model’s ability to accurately detect and classify objects. The results of the study show that the best model performance was achieved with YOLOv8m using the Augmented Dataset. Subsequently, hyperparameter tuning was performed on batch size, learning rate, and epochs to achieve optimal performance. The optimal hyperparameter settings for the YOLOv8m model were a batch size of 24, a learning rate of 0.01, and 100 epochs. This configuration yielded an accuracy of 98.0%, a precision of 99.2%, a recall of 98.7%, an mAP50 of 99.3%, and an mAP50-95 of 92.2%. The implementation of this technology is expected to improve sorting efficiency, reduce reliance on human observation, and support AI-based quality control of cocoa beans. The research results are also expected to serve as a reference for the development of automation technology in the cocoa processing industry.

Kata Kunci : biji kakao, deep learning, deteksi objek, model, YOLOv8