Lightweight detection method for weeds in complex paddy fields based on the improved YOLOv5

Weeds are a significant factor that influences the growth of rice, and accurate weed identification is crucial for field management. With the development of image recognition technology, this technology has been gradually applied to the field of paddy field weeding. To address the problem of the poor weed recognition effect of this technology in the process of intelligent paddy field weeding, this study establishes a paddy field weed dataset in a complex environment and proposes a lightweight paddy field weed recognition model (FGE-YOLO). The model is based on the YOLOv5s network, replacing the YOLOv5s backbone network with a lightweight network (FasterNet), introducing the GSConv module to replace some traditional convolutional layers in the neck network, and incorporating an efficient channel attention (ECA) module in the C3 module of the neck network. The experimental results show that the FGE-YOLO model maintains high performance with a lightweight improvement, achieving a precision (P) of 91.1%, a recall (R) of 89.0%, and a mean average precision (mAP50) of 91.9%. Compared to the base YOLOv5s model, the model parameters are reduced by 34.5%, GFLOPs are reduced by 39.4%, and the memory usage is 66.8% of the preprocessing one. Finally, the FGE-YOLO model achieved the inference time per image of 118.7 milliseconds after deployment on the embedded Jetson Nano device. Compared to the base YOLOv5s model, it was reduced by 16.9%. The research findings are beneficial for achieving precise field management and provide some methodological reference for the research of intelligent paddy field weeding machinery.