Ricetection: Implementasi Arsitektur Vision Transformer untuk Klasifikasi Varietas Beras dengan Presisi Tinggi

Abstract

This research presents Ricetection, a rice variety classification system developed using the Vision Transformer (ViT) architecture to achieve high-precision rice grain recognition. Productivity and quality control in rice post-harvest handling remain crucial challenges, particularly in cases where rice varieties are visually similar. Rice quality assessment that is still performed manually often leads to inconsistency and misclassification. To address this, this study proposes an image-based classification approach utilizing a publicly available rice grain dataset consisting of four Indonesian varieties: IR64, Inpari 32, Mekongga, and Ciherang. The image preprocessing stage includes normalization, resizing, and augmentation to enhance data variation and improve model learning. TheViT architecture is applied as the core classifier due to its ability to capture global image features more effectively compared to traditional CNN-based models. Model performance evaluation shows that ViT achieves superior accuracy, reaching above 98% in the testing phase. Additionally, a web-based application prototype is implemented to provide real-time prediction through an intuitive user interface, enabling users to upload rice grain images and obtain classification results immediately. The proposed system is expected to assist rice farmers, quality control institutions, and supply chain stakeholders in improving decision-making related to rice standardization and variety detection. Future research may explore broader datasets, multi-class classification expansion, and integration into industrial-scale quality inspection systems.