Comparison of Mobilenetv3 Large and Small Architectures for Coffee Bean Roasting Level Classification Based on Web

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Conventional assessment of coffee roasting levels depends on the visual and olfactory judgment of a roaster, which makes it subjective and inconsistent. This study compares two convolutional neural network architectures, MobileNetV3-Large and MobileNetV3-Small, for classifying four roasting levels: green bean, light roast, medium roast, and dark roast. Both models were trained with transfer learning on 1,600 images (960 training, 240 validation, and 400 testing) and evaluated at batch sizes of 16, 32, and 64. MobileNetV3-Large with a batch size of 16 achieved the best accuracy of 99.50% with a loss of 0.0409, while MobileNetV3-Small reached 98.00% with a loss of 0.0764. However, MobileNetV3-Small was far more compact, with 2,122,430 parameters and a size of 8.10 MB, compared with 6,197,326 parameters and 23.64 MB for MobileNetV3-Large. MobileNetV3-Small was also more sensitive to batch size, dropping to 70.25% accuracy at a batch size of 64, whereas MobileNetV3-Large stayed above 94%. The best model was deployed in a web-based system with a Flask API backend, using CLAHE and test-time augmentation at inference to support prediction stability. The findings show a clear trade-off between accuracy and computational efficiency.