Klasifikasi Kualitas Getah Kemenyan Menggunakan Arsitektur GoogleNet Inception V3 Berbasis Citra

Abstract

Proses penilaian kualitas getah kemenyan pada umumnya masih dilakukan secara visual sehingga berpotensi menimbulkan ketidakkonsistenan. Penelitian ini bertujuan merancang model klasifikasi kualitas getah kemenyan berbasis citra digital menggunakan arsitektur Convolutional Neural Network (CNN) Inception V3. Dataset yang digunakan berjumlah 1.500 citra yang dibagi ke dalam data pelatihan, validasi, dan pengujian dengan tiga kategori kualitas, yaitu tinggi, sedang, dan rendah. Empat konfigurasi pelatihan diterapkan, yakni training from scratch dan transfer learning, masing-masing dengan dan tanpa penerapan data augmentation. Hasil evaluasi menunjukkan bahwa konfigurasi training from scratch dengan data augmentation menghasilkan performa terbaik dengan akurasi pengujian sebesar 99% serta nilai rata-rata precision, recall, dan F1-score sebesar 0,99. Analisis confusion matrix memperlihatkan tingkat kesalahan klasifikasi yang sangat minimal pada seluruh kategori. Model terbaik kemudian dikonversi ke format TensorFlow Lite dan diintegrasikan ke dalam aplikasi mobile berbasis Android. Temuan ini menunjukkan bahwa arsitektur Inception V3 mampu memberikan kinerja klasifikasi yang tinggi dan mendukung proses penilaian mutu getah kemenyan secara lebih objektif. The quality assessment of benzoin resin is generally conducted through visual inspection, which may lead to inconsistencies and subjectivity. This study aims to develop an image-based classification model for benzoin resin quality using the Convolutional Neural Network (CNN) architecture Inception V3. The dataset consists of 1,500 images divided into training, validation, and testing sets across three quality categories: high, medium, and low. Four training configurations were implemented, namely training from scratch and transfer learning, each with and without data augmentation. The evaluation results indicate that the training from scratch configuration with data augmentation achieved the best performance, with a testing accuracy of 99% and average precision, recall, and F1-score of 0.99. The confusion matrix analysis demonstrates a very low misclassification rate across all categories. The bestperforming model was subsequently converted into TensorFlow Lite format and integrated into an Android-based mobile application. These findings demonstrate that the Inception V3 architecture provides high classification performance and supports a more objective quality assessment process for benzoin resin.