Purpose <p>With the rapid advancement of weeding robot technology, accurate identification of crops and weeds has become a prerequisite for autonomous weeding. To address the challenges of low accuracy and poor real-time performance in corn (<i>Zea mays</i> L.) and weed recognition under complex field conditions, a lightweight detection model, RT-DETR-ME, is proposed based on an improved RT-DETR framework.</p> Methods <p>A lightweight backbone network, MEIE Backbone, is introduced, which incorporates a multi-scale edge information enhancement module to improve feature extraction capability while reducing parameter count and computational cost. Furthermore, an efficient discriminative frequency domain–based feedforward network (EDFFN) is integrated into the attention-scale intra-feature interaction (AIFI) module to reduce spatial information loss and better capture the relationships between different regions in the image. Finally, a gConvC3 module is designed to replace the fusion block in the cross-scale feature fusion (CCFF) module, aiming to enhance the model’s ability to extract fine-grained features.</p> Results <p>Experimental results demonstrate that the proposed RT-DETR-ME model achieves a 0.4% improvement in mAP@0.5 compared to the original model, while reducing the number of parameters and FLOPs by 35.1% and 27.7%, respectively. And the model reaches an inference speed of 56.9 FPS.</p> Conclusion <p>The model proposed in this paper can meet the requirement of quickly distinguishing multiple weeds with similar morphology from the corn crop under the complex growing environment of corn, which provides a reference for the design of the subsequent intelligent weeding equipment.</p>