Development of Android application for coffee roast level classification using CNN mobilenetev3 based on digital image analysis

Abstract

Coffee is one of Indonesia's key commodities with high economic value in both domestic and international markets. Roasting is a crucial stage in coffee processing, as it directly influences the final taste and aroma. However, until now, the roasting level has been determined visually and subjectively, making it prone to human error. This study aimed to develop an automatic classification system for coffee roasting levels using a Convolutional Neural Network (CNN) approach integrated into an Android application. The CNN model was built using the MobileNetV3 architecture and applied with a transfer learning method on Google Colab. The dataset consisted of 1,600 coffee images with three roasting levels: light, medium, and dark. The trained model was then converted into TensorFlow Lite (TFLite) format for integration into an Android application called RoastScan. This app features automatic classification through the camera or gallery and provides real-time predictions with confidence levels. The evaluation results show that the model has a classification accuracy of 97% on the test data, with high precision, recall, and F1-score across all classes. Further testing via the application showed that the best accuracy reached 93.55%. These findings suggest that integrating CNN with mobile applications has the potential to be a practical, efficient, and accurate solution for determining coffee roasting degrees, as well as supporting the standardization of coffee product quality at both industrial and MSME levels.