Abstract: Four typical weeds Xanthium strumarium,Setaria viridis, Chenopodium album,Ambrosia artemisiifoliain in rapeseed fields were taken as the research objects. Key challenges in weed detection, including small seedling targets, weak features of withered weeds, and difficulty in identifying highly overlapping areas, were addressed by proposing an improved detection method based on the RT-DETR (Region transformer DETR) model. The asymptotic feature pyramid network (AFPN) replaced the original cross-scale context fusion module (CCFM) in the RT-DETR model, effectively resolving the imbalanced feature distribution issue in withered weeds caused by blurred texture and feature sparsity. The SPD-Conv module was introduced into the backbone network to enhance the feature representation capability for small-target weeds. The convolutional block attention module (CBAM) was integrated at the end of the backbone network, effectively mitigating feature information loss under low-resolution targets and occlusion conditions. Systematic ablation experiments and comparative experiments verified that the improved RT-DETR+AFPN+SPD-Conv+CBAM (RW-DETR) model demonstrated significant advantages in both detection accuracy and robustness. The RW-DETR model achieved recognition precision and mean average precision of 85.2% and 82.5%, respectively, for weeds in rapeseed fields, significantly outperforming the RT-DETR model, Faster R-CNN model, SSD model, YOLOv5m model, and YOLOv8m model. While maintaining real-time detection performance, the RW-DETR model significantly improved weed recognition effectiveness in complex scenes, meeting the accuracy and efficiency requirements of modern agriculture for weed detection systems.

Key words: rapeseed, improved RT-DETR model, RT-DETR model, weed recognition