$Saf\text{-}2^{10}$: A Novel Benchmark Dataset for Crocus Sativus L. Flower Detection

Saffron is a spice derived from Crocus sativus L. flowers. This kind of production is currently one of the few agricultural fields for which no significant attempts of mechanization and automation have been proposed, despite its high economic value. This is mainly due to the fragility of these flowers, for which traditional harvesting methods are not applicable. In recent years, many robotic and artificial intelligence systems have been applied in agriculture, allowing them to face problems that traditional approaches have shown to be unable to solve. A reliable in-field saffron flower detection system constitutes the first step toward any kind of production automation. In this work, a novel saffron image dataset featuring two classes, i.e., saffron flower and bud, is presented and named $Saf\text{-}2^{10}$ dataset. The dataset has been used to train three widespread object detection and classification frameworks based on deep convolutional neural networks: YOLOv8, Faster R-CNN, and RetinaNet. The models were trained using both the original version and an augmented version of the $Saf\text{-}2^{10}$ dataset, making a detailed performance comparison. The best results were provided by the YOLOv8 medium-size model and by Faster R-CNN with a ResNet-101 backbone, both trained on the augmented dataset. This study shows that automatic saffron flower detection and classification is indeed viable.