<![CDATA[Distinguishing coffee bean varieties remains a significant challenge in the agricultural industry due to high inter-class similarity and the subtle morphological differences between species. This study aims to conduct a comparative evaluation of MobileNetV2 and EfficientNetB0 for fine-grained coffee bean classification, specifically investigating how efficiency-oriented architectural mechanisms such as depthwise separable convolution and compound scaling influence feature extraction. The research employed a quantitative experimental method using a private dataset of 2,400 images comprising Arabica, Robusta, and Liberica varieties. Data preprocessing included resizing to 224×224 pixels and augmentation, followed by training the two architectures using transfer learning under a controlled experimental framework. The results showed that EfficientNetB0 achieved superior performance with a testing accuracy of 99.17%, while MobileNetV2 attained a competitive accuracy of 98.33% with lower computational complexity. These results demonstrate that while EfficientNetB0 is optimal for high-precision industrial sorting, MobileNetV2 offers a highly efficient alternative for resource-constrained mobile applications. This study provides a scalable framework for automating quality control, effectively balancing architectural efficiency with the sensitivity required for accurate coffee variety identification.]]>
Copyrights © 2026
