The precise identification of corn and weeds plays an important role in precise spraying. This paper proposed a lightweight model based on the improved yolov5s and built a precision spraying robot. Firstly, we used a data augmentation method based on category balance and agronomic characteristics to solve the data imbalance problem. Then, compared with yolov5s, yolov5l, yolov5m, and yolov5x, we found that yolov5s has both real-time and accuracy and is easier to deploy the model on edge devices. Through the feature map visualization experiment, we found that the feature extraction network can’t pay close attention to the important feature of the target and suppress the feature of the noise. Therefore, we added the attention mechanism. In order to improve the real-time performance of the model, we designed the C3-Ghost-bottleneck module. Finally, we built a precision spraying robot. Compared with the original model, the value of map@0.5 is increased by 3.2%, the model file is reduced by 3.6 MB, the AP value for corn is increased from 93.2% to 96.3%, and the AP value for weeds is increased from 85.6% to 88.9%. Finally, the precision spraying experiment of weeds was carried out. The recognition accuracy of weeds is 83%, the probability of the spraying robot correctly identifying weeds and accurately spraying is 81%, and the detection speed is 30ms/f. The experimental results verify the feasibility of precision spraying weeding and the effectiveness of the improved model, which can provide a reference for the engineering application of precision weeding.
Wang et al. (2023) studied this question.