The coffee industry in Indonesia continues to grow; however, the manual identification of Arabica and Robusta coffee bean varieties is prone to error. This study aims to develop an image classification model based on a Convolutional Neural Network (CNN) specifically the ResNet-50 architecture using a transfer learning approach to distinguish between the physical shape profiles of coffee beans. A dataset comprising 990 images sourced from Roboflow was divided into training (70%), validation (15%), and testing (15%) sets. The research process involved preprocessing (background removal and quality enhancement), data augmentation, modeling, and evaluation. Two configurations were tested: a base model acting as a feature extractor and a modified model incorporating additional Dense and Dropout layers. The results indicate that the base model achieved a test accuracy of 94.00%, while the modified model achieved 94.67%, with average precision and recall values of 95%. The addition of custom layers proved effective in stabilizing training, lowering validation loss, and mitigating the risk of overfitting. In conclusion, ResNet-50 is effective for classifying coffee varieties based on visual geometry and holds potential for implementation as an automated sorting system within the coffee industry.
Copyrights © 2026