Abstract: Based on deep learning technology, SSD, YOLOv5 and Fast RCNN models with the same basic network （vgg16-Fast R-CNN, darknet53 -Fast R-CNN） were built to detect round spot, gray mold, powdery mildew and viral diseases of Panax notoginseng under different conditions. The results showed that the m-weight model of YOLOv5 performed the best among all weight models of YOLO, with accuracy rate of 88.62%, recall rate of 89.59%, F1 precision of 89.10%, and average precision of 83.55%. The detection time of a single image was only 0.031 s. Compared with vgg16-Fast R-CNN, which performed better in the two-stage model, the accuracy rate, recall rate, F1 precision, and average precision were only reduced by 1.69 percentage points, 3.92 percentage points, 2.78 percentage points, and 3.47 percentage points respectively, but the detection speed of a single image was increased by 451.4%. Compared with the SSD model, the accuracy rate, recall rate, F1 precision, and average precision of YOLOv5m were improved by 1.06 percentage points, 1.32 percentage points, 1.19 percentage points, and 0.61 percentage points respectively, and the detection speed of a single image was improved by 83.52%. In addition, through the analysis of the confidence and robustness test, it could be seen that YOLOv5m had better disease detection ability in small areas and stronger anti-interference ability in complex environment, and was more conducive to deployment in embedded devices, which met the requirements of real-time detection of Panax notoginseng disease.

Key words: Panax notoginseng, disease detection, YOLOv5, deep learning