关键词: 大豆花朵, 生长状态, YOLOv5, 田间复杂环境, 注意力机制, 目标检测

Abstract:  In order to judge the fall of soybean flowers during the flowering period, soybean flowers in the field were accurately detected under four growth states such as flower bud, halfopening, fullopening and withering. Based on the YOLOv5 detection model, the backbone Bottleneck CSP structure was modified, the number of modules was reduced to preserve more shallow features and enhance feature expression ability. CA attention mechanism was introduced into the backbone network to obtain location information and help the model identify more accurately. Moreover, the size of anchor box was modified to improve the accurate identification of small target bud, and the improved YOLOv5 algorithm was compared with the selfbuilt data set of different growth states of soybean flowers in the field. The results showed that the accuracy rate of the model reached 93.4% and the recall rate reached 91.4%, which were increased by 0.8% and 2.1% respectively compared with the original model.

Key words: soybean flowers, growth state, YOLOv5, complex field environment, attention mechanism, object detection