Semantic Segmentation Algorithm For Ridge Image Of Paddy Field Based On Deep Learning

The environmental perception technology of agricultural machinery is the primary link in the research of agricultural machinery intelligence.In the unmanned intelligent agricultural machinery system for rice transplanters,machine vision can effectively distinguish the operation area and the ridge area in the paddy field scene,which can provide a basis for subsequent operation path planning.This paper takes the tilled paddy field as the research object,and realizes the semantic segmentation of paddy field images based on deep learning technology combined with supervised training.The research contents of this paper are as follows:1)Construction of paddy field ridge dataset.Taking the experimental field of Jiangxi Agricultural University,the Keli Village Farm of Xinjian County and the Ganzhu Farm of Yuanzhou District as the research areas,the paddy field data collection platform was built based on the 2ZGQ-80 D Yanmar rice transplanter,and the rice transplanter was driven along the field ridges in the morning,middle and evening.machine and collect images of paddy field ridges.Use labelme software to manually label paddy field boundary features to obtain semantic label maps.It is divided into training set,validation set and test set according to a fixed proportion,and the data set is expanded through data enhancement methods such as geometric rotation and random lighting.make.2)Construction and pre-experiment of semantic segmentation model of paddy field and ridge.After completing the production of the field ridge data set,the initial semantic segmentation models such as PSPNet network,UNet network and Multires UNet were constructed based on the Pytorch framework,and the training and testing operations of the models were completed.Through comparative analysis of evaluation indicators,the Multires UNet network with better segmentation effect and high robustness is used as the main framework for the research on paddy field ridge segmentation model in this study,and as the research basis for subsequent paddy field ridge segmentation work.3)Optimization of image semantic segmentation model for field and ridge scene.Based on the Multi Res Unet model,in view of the complex characteristics of the paddy field ridge environment,the Spatial Pyramid Pooling(ASPP)module is added after the fourth downsampling of the model,and a combination of smaller expansion rates is used to improve the extraction of small-scale ridge features;The attention gate mechanism is introduced at the front end of the skip connection to strengthen the characteristic response of paddy field ridges,forming the AM＿Unet model.Through experiments and data analysis,the F1 score and average intersection ratio of the AM＿Unet model are93.95% and 88.60%,respectively,which are 3.31% and 1.89% higher than the original model,respectively.Compared with the typical segmentation models of PSPNet and UNet,the F1 value is improved respectively.3.32% and 2.71%,and the mean m Io U increased by5.71% and 4.68%,respectively.The results show that the AM＿Unet model has higher accuracy and robustness for the segmentation of paddy field ridges,which provides an important basis for further realizing the autonomous identification of ridges and tracking operations by rice planting machinery.4)Semantic model deployment based on TX2 embedded platform.In order to test the time-consuming segmentation of this model in the actual field environment,this research is based on NVIDIA’s Jetson TX2 embedded development platform,using the Tensor RT engine to optimize and deploy neural network predictive inference,and transplant the system to an energy-efficient embedded computing platform.The online segmentation of field ridge images based on embedded devices is realized.The segmentation time of a single image is 813 ms,which can basically meet the navigation requirements of the forward-looking path under the normal operating speed of the rice planting machinery of 0.8m/s.