Research On Visual Recognition Technology Of Sugarcane Stem Nodes In Natural Environment Based On Deep Learning

Sugarcane is not only an important sugar crop in China,but also a key economic crop in Guangxi Zhuang Autonomous Region.In order to meet the needs of the intelligent development of sugarcane harvesting machinery in Guangxi,we must first solve the problem of sugarcane stem node recognition in the natural environment.Sugarcane stem node recognition is not only the key technology to realize the fine operation of sugarcane harvesting,but also the prerequisite for the development of small intelligent sugarcane harvester.This can not only liberate a lot of manpower,save the cost of sugarcane industry and improve the income of farmers,but also avoid the physical damage caused by sugarcane stem nodes to harvesting tools in the harvesting process.In view of the complex background of the sugarcane field in the natural environment,the large difference in the characteristics of the identified target sugarcane stem nodes,the sugarcane stem nodes are wrapped by sugarcane leaves to varying degrees,the large variation of the field illumination in the daytime,the high complexity of the algorithm,and the requirements of high memory and high computing power for the deployment platform.The research has put forward data enhancement methods,deep learning methods,lightweight structure design methods and network slimming methods related to field sugarcane stem node recognition to solve the above difficult problems,which provides a theoretical basis for the next sugarcane fine operation.The main contents and conclusions of this paper are as follows:(1)This paper establishes a perfect field sugarcane stem node image data set,which covers large target stem nodes,medium target stem nodes and small target stem nodes,and classifies the wrapped stem node images without sugarcane leaves.(2)Aiming at the problems of large changes in field illumination and different degrees of inclined growth of sugarcane,we adopted the data enhancement method to expand the original data to 6 times through brightness adjustment,mirror image reversal and left-right rotation.By using the control variable method,we can find that this technology can effectively improve the accuracy and robustness of the detection algorithm.(3)A field sugarcane stem node recognition model based on deep learning is proposed,which fills the gap in the field sugarcane stem node recognition at home and abroad.The accuracy performance of different types of target detection algorithms based on deep learning in field sugarcane stem node recognition is discussed,and the superiority of YOLO v5 s algorithm in field sugarcane stem node detection is highlighted.(4)In order to reduce the complexity of the algorithm and the memory and computing power requirements of the deployment equipment,we use the lightweight network of Mobile Net series to design the lightweight structure of YOLO v5 s algorithm.The accuracy of stem node test set recognition after the combination of Mobile Net and YOLO v5s: Mobile Net v2-YOLO v5 s >Mobile Net v1-YOLO v5 s > Mobile Net v3-YOLO v5s;Their complexity(from low to high): Mobile Net v3-YOLO v5 s > Mobile Net v1-YOLO v5 s >Mobile Net v2-YOLO v5 s.(5)Without changing the network structure of the algorithm,the trained sugarcane stem node recognition model based on YOLO v5 s is compressed by the network slimming method,and the unimportant channel layers are found and deleted by applying L1 regularization to the batch normalization layer,so as to reduce the complexity of the algorithm.The results show that when the pruning rate reaches 80%,the accuracy and complexity of the model achieve the optimal balance.