<![CDATA[Manual assessment of papaya ripeness was prone to inconsistency because visual differences between ripeness stages could be affected by color similarity, lighting, and transitional fruit characteristics. This study compared MobileNetV2 and EfficientNet-B0 for classifying papaya ripeness from RGB images by integrating dataset deduplication and model validation. The initial dataset contained 500 images in three classes: mature, partially mature, and immature. After duplicate images were removed using SHA-256, 246 unique images were obtained. The models were trained using transfer learning with ImageNet weights, stratified data splitting, and five-fold stratified cross-validation. Testing on the deduplicated dataset showed that MobileNetV2 achieved 95.92% accuracy, while EfficientNet-B0 achieved 93.88%. MobileNetV2 was more efficient in terms of model size and inference time, whereas EfficientNet-B0 was more stable in cross-validation. These findings showed that dataset deduplication, efficiency evaluation, and visual interpretation produced a more representative model assessment.Keywords: Papaya ripeness classification; Dataset deduplication; Transfer learning; MobileNetV2; EfficientNet-B0 AbstrakPenilaian manual terhadap kematangan buah pepaya dapat menghasilkan ketidakkonsistenan karena perbedaan visual antartingkat kematangan dipengaruhi oleh kemiripan warna, pencahayaan, dan karakter buah pada fase transisi. Penelitian ini membandingkan MobileNetV2 dan EfficientNet-B0 untuk klasifikasi tingkat kematangan buah pepaya berbasis citra RGB dengan mengintegrasikan deduplikasi dataset dan validasi model. Dataset awal terdiri dari 500 citra dalam tiga kelas, yaitu mature, partially mature, dan immature. Setelah citra duplikat dihapus menggunakan SHA-256, diperoleh 246 citra unik. Model dilatih menggunakan transfer learning dengan bobot ImageNet, stratified split, dan 5-fold stratified cross-validation. Hasil pengujian pada dataset hasil deduplikasi menunjukkan MobileNetV2 memperoleh akurasi 95,92%, sedangkan EfficientNet-B0 memperoleh 93,88%. MobileNetV2 lebih efisien dari sisi ukuran model dan waktu inferensi, sedangkan EfficientNet-B0 lebih stabil pada validasi silang. Temuan ini menunjukkan bahwa deduplikasi dataset, evaluasi efisiensi, dan interpretasi visual menghasilkan penilaian model yang lebih representatif. ]]>
