In order to address the challenges posed by the high model complexity and low detection efficiency of existing sugarcane node detection algorithms, a lightweight detection network YOLOv5s-SG2E was designed. Firstly, this study enhanced the network's capability to detect small targets by refining the structure of the small target network, which involved the removal of medium and small-scale detection heads and the incorporation of super-large-scale detection headed to enhance the model's perception of small targets. Secondly, GhostNetV2 replaced the C3 module, and GSConv replaced standard convolutions in the neck network to reduce the models complexity. Finally, the channel attention mechanism ECA was introduced at the end of the backbone network to enhance the model's learning abilities and strengthen the network's ability to extract the sugarcane nodes features. The self-made sugarcane node dataset was tested, and the experimental results showed that the YOLOv5s-SG2E model achieved a 96.4% accuracy rate for sugarcane node recognition, a 96.8% recall rate, and an average accuracy mAP@0.5 of 98.4%. These metrics represented improvements of 0.6%, 2.4% and 1.0%, respectively, over the original YOLOv5s model. Furthermore, YOLOv5s-SG2E exhibited significant reductions in model size, with an 89.8% volume reduction, a 95.03% reduction in parameters, and a 55.06% decrease in computational workload. Additionally, detection time was shortened by 31.6%. When compared to other mainstream one-stage target detection algorithms, YOLOv5s-SG2E outperforms them, which can realize the efficient identification and detection of sugarcane stem nodes.