Penerapan Arsitektur TL-MobileNetV2 Untuk Klasifikasi Pisang Lokal Berbasis Mobile

Abstract

Bananas are one of the horticultural commodities with numerous varieties and diverse morphological characteristics. However, several banana varieties share similar visual features such as shape, color, and peel texture, making manual identification difficult, particularly for individuals without specialized knowledge of banana varieties. This condition may lead to misidentification when relying solely on visual observation. Therefore, an automated system is required to support accurate and efficient banana variety identification. This study aims to develop an image classification model for banana variety recognition using the Transfer learning MobileNetV2 (TL-MobileNetV2) architecture and to implement the model in an Android-based mobile application. The dataset used in this study consists of 400 banana images representing eight varieties, namely agung, ambon, barlin, candi, cavendish, embug, mas, and susu bananas. Two model architectures were evaluated, namely the standard MobileNetV2 and TL-MobileNetV2, with variations of optimizers (Adam and SGD), Learning rates (0.001 and 0.0001), and three dataset types (center cropping, object cropping, and without cropping). Model performance was evaluated using a confusion matrix with four evaluation metrics: accuracy, precision, recall, and F1-Score. The experimental results show that the best-performing model was obtained using the TL-MobileNetV2 architecture with the Adam optimizer and a Learning rate of 0.001 on the object cropping dataset, achieving a score of 1.000 for all evaluation metrics. Grad-CAM heatmap visualization indicates that the model focuses on relevant banana object regions during the classification process. The best model was subsequently converted into the TensorFlow Lite (.tflite) format and deployed in an Android-based mobile application. The implementation results demonstrate that the model performs inference quickly and reliably on mobile devices. These findings indicate that TL-MobileNetV2 is effective for banana variety identification and can be practically implemented as an image-based classification system on mobile platforms.