Abstract
<title>Abstract</title> <p>Saffron (Crocus sativus L.) is a high-value spice prone to economically motivated adulteration due to its distinctive morphology and premium price. This study investigates the potential of convolutional neural networks (CNNs) as a rapid, non-destructive, and cost-effective tool for distinguishing authentic saffron stigmas from visually counterfeit materials. A curated dataset of 1,020 expert-labeled images (719 authentic, 301 counterfeit) was used to evaluate three CNN architectures—AlexNet, ResNet, and VGG16—through transfer learning. The dataset was split into 80% training and 20% validation, with a separate test set for final evaluation. AlexNet achieved the highest accuracy (96.49%) with perfect sensitivity (100%) and high specificity (93.75%), outperforming ResNet (94.74%) and VGG16 (87.72%). These results demonstrate the feasibility of CNN-based visual authentication for saffron and provide a foundation for practical deployment in postharvest quality control pipelines. Future work should explore larger datasets, domain generalization, and lightweight on-device inference for real-world applications.</p>