Klasifikasi Warna Biji Kopi menggunakan Algoritma EfficienNet

Abstract

Determining the roasting level of coffee beans is generally done manually by roast masters, making it subjective and prone to inconsistencies. To address this issue, this study developed a Convolutional Neural Network (CNN)-based automatic classification system using the EfficientNet B0–B5 architecture to identify four roasting levels: green, light, medium, and dark. The dataset used consisted of 1,600 evenly distributed coffee bean images and was processed through preprocessing, normalization, and light augmentation stages. All EfficientNet models were trained using ImageNet pretrained weights to maximize the ability to extract visual features such as color, texture, and surface gloss of coffee beans. The test results showed that EfficientNet-B1, B2, B4, and B5 achieved the highest accuracy of 100%, while variants B0 and B3 achieved 99%. EfficientNet-B1 was considered the optimal model because it provided the highest accuracy with a relatively more efficient training time compared to B2, B4, and B5. These findings demonstrate that CNN methods, particularly EfficientNet, are capable of detecting visual differences between roasting levels with high accuracy and have the potential to be used as an automation system for roasting quality standards in the coffee industry.