Implementation of Coffee Bean Roasting Level Classification System Using CNN and Knn Models with Web-Based Real-Time Camera

Abstract

Manual coffee bean roasting assessment is still susceptible to operator subjectivity, variations in lighting conditions, and inconsistencies in results between assessors. This study aims to develop a real-time web-based coffee bean roasting classification system by integrating Convolutional Neural Network (CNN) and K-Nearest Neighbor (KNN) models. The study uses a quantitative experimental approach with a dataset of coffee bean images collected independently and expanded to 6,470 images, which are grouped into five classes: GreenRoasting, LightRoasting, MediumRoasting, DarkRoasting, and Unknown. All images are processed through a pre-processing stage including resizing to 160 × 160 pixels, normalization, data augmentation, and splitting training and test data with a ratio of 80:20 in stages. MobileNetV2 is used as a feature extractor in CNN, while KNN with a value of k = 7 and a cosine distance metric is applied for feature vector classification. The final prediction was obtained using a weighted ensemble method with a composition of 60% CNN and 40% KNN, then implemented in a Flask-based web application with support for real-time image upload and camera. Test results showed the ensemble model achieved an accuracy of 85.67% with an average response time of 1,247 ms. This system has the potential to support faster, more consistent, and objective coffee roasting level assessments, especially for small to medium-scale coffee businesses.