关键词: 目标检测, YOLOv3, 卷积神经网络, 马铃薯芽眼, 深度学习

Abstract:  In order to quickly and accurately identify potato buds and improve the germination rate of seed potatoes, a potato buds recognition method based on convolutional neural network is proposed. The database of the potato buds images with multiple perspectives and different degrees of overlap is established by data augmentation and image preprocessing. Based on this, the rapid and accurate recognition of the potato buds is achieved by using the highperformance feature extraction characteristics of the YOLOv3 network. The result shows that the YOLOv3 network can achieve good and stable recognition for samples containing a single unobstructed bud, samples containing multiple obstructed buds, and samples containing mechanical damage as well as bug eyes and impurities. The final detection accuracy P is 97.97%, the recall rate R is 96.61%, the harmonic mean F1 is 97%, the mean recognition accuracy mAP is 98.44%, and the single detection time is 0.018 s. After comparing and analyzing networks such as YOLOv4-tiny and SSD, the YOLOv3 model can simultaneously meet the accuracy and speed requirements of potato buds recognition. Therefore, the YOLOv3 network has good robustness to the potato buds recognition, which lays the foundation for the realization of an automatic potato seed cutting machine.

Key words: target detection, YOLOv3, convolutional neural network, potato buds, deep learning