As one of the important raw materials for sugar production，sugar cane extracted is not only a necessity but also belongs to national strategic reserve materials. At present，the sugarcane industry as a whole is inefficient，especially in terms of machine seeding，and the efficiency of manual seed cutting is low. The developed methods for sugarcane stem node identification are all based on clean background conditions，but the working environment of agricultural machinery is disgusting. A large amount of debris，miscellaneous leaves and other dirt produced by seed cutting will lead to background pollution in image acquisition area and reduce the recognition ability of algorithm. Therefore，a sugarcane stem node identification method is put forward based on improving YOLOv5 to solve the problems of low recognition accuracy of sugarcane stem node under complex background. To optimize the neck structure and enhance the information fusion capability among different levels by using cross-level connection. At the same time，the model loss function is improved. On the one hand，EIoU loss is introduced to replace original CIoU loss to improve the precision of boundary box regression； on the other hand，Focal loss function is used to replace the cross-entropic loss function to solve the problem of unbalanced proportion of positive and negative samples. Finally，Ghost module is introduced to lightweight network model. The experimental results show that compared with the original model，the average precision value of the model proposed in this study is increased to 97.80%，the single detection time is 16.9 ms and the memory of the model is only 11.4 Mb，which realizes the identification of sugarcane stem joints in different chaotic scenes and reduces the impact of background chaos when cutting injurious buds.