Comparative performance analysis of convolutional neural network-architectures on coffee-bean roast classification

Abstract

The classification of coffee bean roast levels using Agtron standards has evolved from traditional subjective methods to technology-driven approaches employing advanced artificial intelligence. Recent advancements in computer vision have demonstrated the capability of convolutional neural networks (CNNs) in providing objective and consistent roast level classification compared to human visual assessment, which is prone to variability and subjectivity. This research presents a performance analysis of five CNN architectures (AlexNet, ResNet, MobileNet, VGGNet, and DenseNet) for classifying coffee beans into eight distinct Agtron roast levels. The comprehensive methodology encompasses four phases: i) data acquisition, ii) image preprocessing, iii) model training and validation, and iv) evaluation metric. During training-validation, DenseNet outperformed other models, achieving 99.702% training accuracy and 77.68% validation accuracy. In the testing evaluation, DenseNet also led with an average testing accuracy of 93.8%, followed by ResNet at 92.6%, VGGNet and AlexNet both at 92.4%, and MobileNet at 89.7%. The results show that the DenseNet shows promise in classifying Agtron coffee-bean roast classification.