<![CDATA[Manual identification of Semarang guava cultivars is prone to subjectivity. This study proposes a MobileNetV2 model based on transfer learning to classify 12 cultivars. As an initial study, the main limitation of this research is the very small dataset size, consisting of 192 images with a balanced distribution of 16 images per class. The data were acquired under varied in-the-wild conditions, including differences in background, lighting, and shooting angles. The dataset was divided using a 70 percent training and 30 percent validation ratio. The testing results showed that the validation accuracy reached 94 percent, with an average F1-score of 0.94. However, analysis using the confusion matrix and per-class evaluation showed that the model still experienced difficulties in fine-grained misclassification among visually similar fruits. Considering the small dataset size and the absence of testing using an independent test set or cross-validation, the model’s performance should only be regarded as an initial indication with limited generalizability. In addition, the fine-tuning stage was found to be less significant. As a recommendation, future research should increase the data volume, apply cross-validation testing, and explore architectures with attention mechanisms.]]>
Copyrights © 2026
