Performance Comparison of CNN Transfer Learning Models for Coffee Bean Quality Classification

Nur Muhammad Fadli (Politeknik Negeri Jember)
Prawidya Destarianto (Politeknik Negeri Jember)
Hendra Yufit Riskiawan (Politeknik Negeri Jember)
Bekti Maryuni Susanto (Politeknik Negeri Jember)
Satrio Adi Priyambada (National Research and Innovation Agency)
Wawan Hendriawan Nur (National Research and Innovation Agency)
Mukhamad Angga Gumilang (Politeknik Negeri Jember)

Jurnal Teknologi Informasi dan Terapan (J-TIT)
Vol 12 No 2 (2025): December

According to SNI Standard No. 01-2907-2008, accurate sorting of coffee beans is crucial for improving export value. Manual sorting is time-consuming, subjective, and error-prone, especially when visual differences are subtle between roast levels. This study proposes and evaluates an automatic, machine-learning based system to support quality assurance in coffee production. We compare three transfer-learning CNN architectures: Xception, MobileNetV2, and EfficientNet-B1 on a publicly available dataset of 1,600 coffee bean images divided into four classes (dark, medium, light, green). All models were trained with the same preprocessing and hyperparameter settings. EfficientNet-B1 achieved the highest test accuracy (100%), followed by Xception (99.5%) and MobileNetV2 (97%). We discuss trade-offs between accuracy and computational efficiency and recommend EfficientNet-B1 for high-accuracy applications and MobileNetV2 for edge/mobile deployment.