Dragon fruits quality grading system using machine learning approach

Abstract

Dragon fruit production has gained popularity in Bangladesh's agricultural sector in recent years because of its high economic worth and health advantages. Dragon fruit quality grading is still done by hand, which is time-consuming and prone to mistakes. This results in inconsistent and ineffective supply chain management. In order to close this gap, this study suggests an automated dragon fruit quality assessment system based on machine learning that divides fruits into two groups: fresh and defective. To enhance model generalization, a custom dataset of 5,011 photos was gathered and preprocessed utilizing data augmentation methods as flipping, zooming, and rotation. Three deep learning models—a baseline Convolutional Neural Network (CNN), ResNet-50 with transfer learning, and EfficientNet-B0—were investigated and contrasted. With a validation accuracy of 99.5% and an AUC-ROC score higher than 0.95, the suggested EfficientNet-B0 model outperformed models found in the literature. This study shows that the suggested method provides a highly accurate, scalable, and economical solution for practical agricultural applications. It guarantees uniformity in the distribution of superior produce, reduces the need for human intervention in quality evaluation, and improves the accuracy of fruit grading. By enhancing decision-making, cutting losses, and preserving customer happiness, it may also benefit farmers, importers, and other agricultural stakeholders. In addition to laying the foundation for future research combining multi-class fruit grading and mobile deployment, the study's findings greatly advance AI-driven precision agriculture.