Abstract: A target detection model for mulberry fruit sclerotiniose disease severity was constructed based on YOLOv3 deep learning algorithm combined with transfer learning by training on 10 000 images of mulberry fruit with five different disease severity levels. To verify the robustness of the YOLOv3 model, comparative experiments were conducted with the EfficientDet, Faster R-CNN and YOLOv4 that also used transfer learning. The results showed that the average precision rate of the YOLOv3 model could reach 0.79 for detecting healthy fruits and sclerotinia fruit, which was 6.76%~54.90% higher than that of the other models. The average precision rate of the YOLOv3 model for detecting disease severity levels of sclerotinia fruit was 7.04%~80.95% higher than that of the other models. The detection precision rate and recall rate of the YOLOv3 model were optimal or sub-optimal. The detection and recognition system constructed by Flask+Vue technology could obtain disease severity, fruit size and confidence information within 1 s, and could also realize dynamic recognition of video. This system could provide a reliable software processing platform for automated disease monitoring and fast, efficient, and precise fungicide application during mulberry cultivation.

Key words: mulberry fruit sclerotiniose disease, deep learning algorithm, transfer learning, YOLOv3, detection of disease severity