Comparative Analysis of Lightweight Pre-Trained CNN Models for Coffee Bean Roasting Level Identification

Coffee is a widely consumed beverage across various societal strata. The quality of coffee taste depends on the roasting process applied to the coffee beans. To ensure consistent quality and maturity of coffee beans, monitoring during the roasting process is essential. This study presents a comparative analysis of lightweight pre-trained CNN models that can effectively maintain coffee quality and maturity without human intervention. Specifically, four lightweight pre-trained CNN models-SqueezeNet, ShuffleNet, MobileNetv2, and NasNet Mobile-are chosen. The “Coffee Roast Intelligence” dataset is utilized, comprising 1200 training images and 400 testing images across four classes. The results indicate that all the pre-trained model achieved 100% accuracy. SqueezeNet and ShuffleNet can be regarded as superior models compared to MobileNetv2 and MNasnet due to their smaller number of learning parameters and model sizes. The learning parameters of ShuffleNet and SqueezeNet is less than one million, whereas the other two models have over two million parameters. Based on these findings, it can be concluded that SqueezeNet and ShuffleNet are the most suitable embedded model for coffee bean roasting level identification.