Construction of Lightweight Model for Cotton Top Sprout and Research on Targeted Cotton Topping Device

Abstract

To mitigate potential risks associated with terminal bud inhibitors on cotton plant growth, ecological environment, and human health during cotton topping operations, a lightweight model by integrating the GhostNetV2 network with an end-side neural network architecture based on the improved YOLOv7-tiny algorithm for cotton terminal bud detection. This approach reduces both the model’s parameter count and reasoning speed while minimizing accuracy loss. Additionally, an end-effector execution scheme and workflow for contactless targeted spraying is proposed using this model. The deployed model on a Jetson TX2 embedded computer achieved a computation load of only 1.0 G with an average accuracy of 98.2%. Moreover, the number of drug droplets attached to the end-effector per square centimeter meets national standards for targeted spraying. The experimental results show that this method is suitable for cotton topping operations.

Keywords

• Cotton topping
• deep learning
• YOLOv7-tiny
• precision topping
• spatial localization