Electronic Theses and Dissertation

Abtrak
Deteksi objek berbasis deep learning telah menjadi teknologi penting dalam otomasi industri, termasuk pengklasifikasian biji kopi berdasarkan kualitas dan jenisnya. Biji kopi merupakan salah satu komoditas unggulan dengan nilai ekonomi tinggi, terutama di Indonesia. Proses penyortiran biji kopi secara manual seringkali tidak efisien dan rentan terhadap kesalahan. Penelitian ini berkontribusi pada pengembangan teknologi berbasis kecerdasan buatan untuk meningkatkan efisiensi dan akurasi proses penyortiran biji kopi di industri. Metode yang dilakukan membandingkan kinerja model YOLOv8 dan YOLOv11 dalam mendeteksi biji kopi multikelas menggunakan visualisasi interaktif melalui Streamlit. Model dilatih menggunakan dataset USK-Coffee yang mencakup kelas biji kopi seperti Defect, Peaberry, Longberry, dan Premium. Model YOLOv8 dan YOLOv11 di-fine-tuning dan dievaluasi menggunakan metrik seperti mean average precision (mAP) dan confusion matrix seperti presisi dan recall. Hasil penelitian menunjukkan bahwa YOLOv11 unggul dalam akurasi dengan mAP sebesar 82%, dibandingkan YOLOv8 dengan mAP sebesar 80%. Visualisasi interaktif berbasis Streamlit terbukti efektif dalam menampilkan hasil deteksi biji kopi dan memudahkan pengguna dalam menganalisis performa model.
Kata kunci: Deteksi Biji Kopi, Klasifikasi Multikelas, Deep learning, YOLO, Streamlit, Visualisasi Interaktif

Abstract Deep learning-based object detection has become a crucial technology in industrial automation, including the classification of coffee beans based on their quality and types. Coffee beans are a high-value commodity, particularly in Indonesia, due to their economic importance. The manual sorting process for coffee beans is often inefficient and prone to errors. This study contributes to the development of artificial intelligence-based technology to improve the efficiency and accuracy of the coffee bean sorting process in the industry. The proposed method compares the performance of YOLOv8 and YOLOv11 models in detecting multiclass coffee beans using interactive visualizations through Streamlit. The models were trained using the USK-Coffee dataset, which includes coffee bean classes such as Defect, Peaberry, Longberry, and Premium. Both YOLOv8 and YOLOv11 models were fine-tuned and evaluated using metrics such as mean average precision (mAP) and confusion matrix metrics like precision and recall. The results show that YOLOv11 outperformed YOLOv8 in terms of accuracy, achieving an mAP of 82%, compared to YOLOv8's mAP of 80%. Interactive visualizations powered by Streamlit proved effective in presenting coffee bean detection results and facilitating users in analyzing the model's performance. Keywords: Coffee Bean Detection, Multiclass Classification, Deep Learning, YOLO, Streamlit, Interactive Visualization