<![CDATA[The classification of coffee bean roasting levels is an important aspect of ensuring coffee product quality. This study compares the performance of two deep learning architectures, ResNet50 and VGG16, in classifying coffee bean images into three roasting levels: light, medium, and dark. The dataset consists of 1,800 images with a resolution of 224×224 pixels, divided into training, validation, and testing sets. Both models were trained with identical configurations using transfer learning and partial fine-tuning. The evaluation results show a very small accuracy difference of only 0.01 point, with ResNet50 slightly outperforming VGG16. This indicates that both models are equally reliable for roast level classification. However, ResNet50 is more time-efficient, requiring only about 10 minutes of training compared to over 25 minutes for VGG16. This difference is suspected to be related to the complexity of VGG16’s architecture. The study concludes that ResNet50 offers high efficiency with competitive accuracy. Further research is recommended to optimize VGG16’s architecture to improve computational efficiency without compromising accuracy.Keywords: Machine learning; Restnet50; VGG16; Coffee bean roasting AbstrakKlasifikasi tingkat roasting biji kopi merupakan aspek penting dalam penjaminan mutu produk kopi. Penelitian ini membandingkan performa dua arsitektur deep learning, ResNet50 dan VGG16, dalam mengklasifikasikan citra biji kopi pada tiga tingkat roasting: light, medium, dan dark. Dataset berisi 1.800 citra beresolusi 224×224 piksel, dibagi menjadi data latih, validasi, dan uji. Kedua model dilatih dengan konfigurasi identik menggunakan transfer learning dan fine-tuning parsial. Hasil evaluasi menunjukkan selisih akurasi sangat tipis, hanya 0,01 poin, dengan ResNet50 sedikit unggul. Hal ini menunjukkan kedua model sama-sama andal untuk klasifikasi tingkat roasting. Namun, ResNet50 lebih efisien secara waktu, hanya memerlukan sekitar 10 menit pelatihan dibandingkan VGG16 yang lebih dari 25 menit. Perbedaan ini diduga terkait kompleksitas arsitektur VGG16. Disimpulkan bahwa ResNet50 menawarkan efisiensi tinggi dengan akurasi kompetitif. Penelitian lanjutan disarankan mengevaluasi optimasi arsitektur VGG16 untuk meningkatkan efisiensi komputasi tanpa mengorbankan akurasi. ]]>
