Kualitas produk merupakan faktor penting dalam industri keramik, khususnya pada piring yang rentan mengalami cacat permukaan seperti pinhole, bump, dan crack, yang dapat menurunkan nilai jual serta membahayakan pengguna. Metode inspeksi manual yang masih dominan bersifat subjektif, memakan waktu, dan berisiko tidak konsisten, sehingga diperlukan sistem deteksi otomatis yang cepat dan akurat. Penelitian ini bertujuan mengembangkan sistem deteksi cacat permukaan piring keramik berbasis deep learning menggunakan algoritma YOLOv8. Dataset terdiri atas 2.081 citra hasil augmentasi dari gambar piring keramik yang mengandung tiga jenis cacat yaitu Pinhole, Bump, dan Crack. Model dilatih dalam dua skenario pembagian data, yaitu 80% pelatihan–10% validasi–10% pengujian dan 70% pelatihan–20% validasi–10% pengujian. Evaluasi performa dilakukan menggunakan metrik precision, recall, mean Average Precision (mAP), dan F1-Score. Hasil menunjukkan bahwa scenario pemisahan data 80:10:10 menghasilkan precision 95%, recall 94,1%, mAP@50 sebesar 97,4%, dan F1-Score 94,4%, sedangkan scenario pemisahan data 70:20:10 menghasilkan precision 89,1%, recall 79,5%, mAP@50 sebesar 90,1%, dan F1-Score 83,8%. Model mampu mendeteksi cacat dengan akurasi tinggi, khususnya pada kelas crack, meskipun performa pada kelas pinhole masih memerlukan peningkatan. Sistem ini berpotensi diimplementasikan pada inspeksi kualitas produk keramik secara real-time untuk meningkatkan efisiensi dan konsistensi kontrol mutu.

Product quality is a crucial factor in the ceramic industry, particularly for plates prone to surface defects such as pinholes, bumps, and cracks, which may reduce market value and pose safety risks. Manual inspection methods, still predominantly used, are subjective, time-consuming, and potentially inconsistent, highlighting the need for a fast and accurate automated detection system. This study aims to develop a ceramic plate surface defect detection system based on deep learning using the YOLOv8 algorithm. The dataset comprises 2,081 augmented images of ceramic plates containing three defect types. The model was trained under two data split scenarios: 80% training–10% validation–10% testing and 70% training–20% validation–10% testing. Performance evaluation used precision, recall, mean Average Precision (mAP), and F1-Score metrics. Results indicate that the 80:10:10 split achieved 95% precision, 94,1% recall, 97% mAP@50, and an 94,4% F1-Score, while the 70:20:10 split achieved 89,1% precision, 79,5% recall, 90,1% mAP@50, and a 83,8% F1-Score. The model effectively detected defects with high accuracy, particularly in the crack class, though performance in detecting pinholes requires improvement. This system has potential for real-time implementation in ceramic product quality inspection to enhance efficiency and consistency in quality control.