Shi Yaying, Li Jiayi, Yu Zeyun, Li Yin, Hu Yangpingqing, Wu Lushen
School of Mechatronic Engineering, Nanchang University, Nanchang 330047, China.
Department of Electrical Engineering and Computer Science, University of Wisconsin-Milwaukee, Milwaukee, WI 53201, USA.
Foods. 2022 Nov 6;11(21):3531. doi: 10.3390/foods11213531.
As a raw material for beer, barley seeds play a critical role in producing beers with various flavors. Unexcepted mixed varieties of barley seeds make malt quality uncontrollable and can even destroy beer flavors. To ensure the quality and flavor of malts and beers, beer brewers will strictly check the appropriate varieties of barley seeds during the malting process. There are wide varieties of barley seeds with small sizes and similar features. Professionals can visually distinguish these varieties, which can be tedious and time-consuming and have high misjudgment rates. However, biological testing requires professional equipment, reagents, and laboratories, which are expensive. This study aims to build an automatic artificial intelligence detection method to achieve high performance in multi-barley seed datasets. There are nine varieties of barley seeds (CDC Copeland, AC Metcalfe, Hockett, Scarlett, Expedition, AAC Synergy, Celebration, Legacy, and Tradition). We captured images of these original barley seeds using an iPhone 11 Pro. This study used two mixed datasets, including a single-barley seed dataset and a multi-barley seed dataset, to improve the detection accuracy of multi-barley seeds. The multi-barley seed dataset had random amounts and varieties of barley seeds in each image. The single-barley seed dataset had one barley seed in each image. Data augmentation can reduce overfitting and maximize model performance and accuracy. Multi-variety barley seed recognition deploys an efficient data augmentation method to effectively expand the barley dataset. After adjusting the hyperparameters of the networks and analyzing and augmenting the datasets, the YOLOv5 series network was the most effective in training the two barley seed datasets and achieved the highest performance. The YOLOv5x6 network achieved the second highest performance. The mAP (mean Average Precision) of the trained YOLOv5x6 was 97.5%; precision was 98.4%; recall was 98.1%; the average speed of image detection reached 0.024 s. YOLOv5x6 only trained the multi-barley seed dataset; the trained performance was greater than that of the YOLOv5 series. The two datasets had 39.5% higher precision, 27.1% higher recall, and 40.1% higher mAP than when just using the original multi-barley seed dataset. The multi-barley seed detection results showed high performance, robustness, and speed. Therefore, malting and brewing industries can assess the original barley seed quality with the assistance of fast, intelligent, and detected multi-barley seed images.
作为啤酒的原料，大麦种子在生产具有各种风味的啤酒中起着关键作用。意外混入的大麦种子品种会使麦芽质量无法控制，甚至会破坏啤酒风味。为确保麦芽和啤酒的质量与风味，啤酒酿造商会在麦芽制造过程中严格检查大麦种子的合适品种。有各种各样尺寸小且特征相似的大麦种子。专业人员可以通过视觉区分这些品种，但这可能既繁琐又耗时，而且误判率高。然而，生物检测需要专业设备、试剂和实验室，成本很高。本研究旨在构建一种自动人工智能检测方法，以在多大麦种子数据集上实现高性能。有九个大麦种子品种（CDC Copeland、AC Metcalfe、Hockett、Scarlett、Expedition、AAC Synergy、Celebration、Legacy和Tradition）。我们使用iPhone 11 Pro拍摄了这些原始大麦种子的图像。本研究使用了两个混合数据集，包括单大麦种子数据集和多大麦种子数据集，以提高多大麦种子的检测准确性。多大麦种子数据集中每张图像中的大麦种子数量和品种是随机的。单大麦种子数据集中每张图像有一粒大麦种子。数据增强可以减少过拟合并最大化模型性能和准确性。多品种大麦种子识别部署了一种有效的数据增强方法来有效扩展大麦数据集。在调整网络的超参数并对数据集进行分析和增强后，YOLOv5系列网络在训练这两个大麦种子数据集方面最有效，并且性能最高。YOLOv5x6网络性能次之。训练后的YOLOv5x6的平均精度均值（mAP）为97.5%；精确率为98.4%；召回率为98.1%；图像检测平均速度达到0.024秒。YOLOv5x6仅训练了多大麦种子数据集；其训练后的性能优于YOLOv5系列。与仅使用原始多大麦种子数据集相比，这两个数据集的精确率提高了39.5%，召回率提高了27.1%，mAP提高了40.1%。多大麦种子检测结果显示出高性能、鲁棒性和速度。因此，麦芽制造和酿造行业可以借助快速、智能且经过检测的多大麦种子图像来评估原始大麦种子质量。
