<p>Real-time weed detection is a crucial technology for enabling smart agriculture and precise weed management, helping to boost crop yields and reduce pesticide use. However, field-based weed recognition still faces several practical challenges, such as inadequate model lightweight and real-time performance, background interference, high similarity between weed species, and difficulty detecting objects at multiple scales. To address these issues, this study proposes a lightweight, real-time weed detection model called SCR-DETR. The model incorporates a Split Reparameterized Convolution (SRC) module into the backbone to reduce computational redundancy and improve detection efficiency. Additionally, a Unified Statistical Attention Reduction (USAR) module based on subspace optimization is introduced to enhance the model’s global representation and robustness in complex environments. Furthermore, a context-guided scale reconstruction (CGSR) framework is introduced to fuse multi-scale semantic features and local details, improving detection accuracy for visually similar weeds. Experiments on the CottonWeedDet3 and CottonWeedDet12 datasets show that the proposed model reduces FLOPs and parameter count by 40% and 33.5%, respectively, while improving mAP@0.5 by 2.9% on one dataset and maintaining performance on the other. The proposed model achieves a real-time inference speed of 15.4 FPS on the edge computing device Jetson Nano, demonstrating its practical potential for real-time deployment in weed detection applications.</p>