Deep learning approach and adaptive data augmentation technique for non-destructive quality recognition of exported saffron

Pouya Bohlol, Ali Bakherad, Mahmoud Omid, Mahmoud Soltani Firouz, Mohammad Hosseinpour-Zarnaq,
Deep learning approach and adaptive data augmentation technique for non-destructive quality recognition of exported saffron,
Journal of Food Composition and Analysis,
Volume 144,
2025,
107758,
ISSN 0889-1575,
https://doi.org/10.1016/j.jfca.2025.107758.
(https://www.sciencedirect.com/science/article/pii/S0889157525005733)
Abstract: Saffron is highly valued for its medicinal and culinary uses. Its global market value and characteristics of saffron have led to the design of quality management and intelligent systems for industrial and postharvest processes. In this study, a computer vision system was integrated with advanced deep learning algorithms and data augmentation techniques to evaluate the quality of saffron, which was categorized into 6 classes based on market demand and business opinions. Raw dataset consisted of 3152 images, with classes including Daste, Poshal Grade Two, Poshal Grade One, Poshal Peresi, Peresi Grade One, and Peresi Negini. Adaptive data augmentation was then implemented in 2 phases, and raw data increased to 22064. Subsequently, hyperparameters of the deep learning model, such as initial learning rate, batch size, image size, optimization, and number of epochs, were assessed, and the best settings for the final model were selected based on model performance. Six popular deep learning models with varying architectures and parameters were evaluated to identify the most optimal structure. The proposed model was InceptionV3, which achieved 100 % accuracy and 0.04 loss in recognizing the quality stage of saffron. The development took 3.20 hours, with a recognition time of 0.055 seconds. The model also had 100 % precision, F1-score, and sensitivity, with an MSE/RMSE/MAE of 0. This model and dataset are highly effective for quality detection and can serve as a basis for developing robotic and remote sensing systems in industrial and quality management systems.
Keywords: InceptionV3; Adaptive data augmentation; Saffron quality; Deep learning; Computer vision; Quality mangement