Image-based identification of onion varieties using deep learning techniques

Abstract

Onion is a crop of immense economic and dietary importance, widely cultivated and traded globally. Accurate identification of onion varieties is critical for pricing, quality assurance, traceability, and consumer preference, yet remains challenging due to high morphological similarity across cultivars. This study is the first reported attempt to classify Indian onion varieties using deep learning applied to bulb images. We evaluated the performance of four pre-trained convolutional neural networks, DenseNet121, InceptionV3, MobileNetV2 and Xception, on a curated image dataset of 10 popular onion varieties. The result showed that DenseNet121 outperformed all models, achieving the highest precision (95.76%), recall (94.92%), F1 score (94.82%) and the lowest mean squared error of 0.94, demonstrating exceptional reliability and accuracy. Its dense connectivity architecture effectively captured subtle features, making it the most suitable for practical applications. InceptionV3 and MobileNetV2 also showed competitive results, with MobileNetV2 offering computational efficiency but facing challenges with certain misclassifications. Xception, despite its efficiency, had the lowest performance metrics, with precision and recall of 91.04% and 88.14%, respectively, and significant misclassification issues. These findings highlight the potential of DenseNet121 for automated onion variety identification and its superiority in addressing the intricate variability within agricultural datasets. These findings demonstrate the potential of deep learning for automating onion variety identification and supporting sorting, grading, and seed chain verification systems. Future work should extend to broader varietal coverage and seasonal datasets for real-world applications.