Research On Weed Identification And Segmentation In Corn Field Based On Deep Learning

As an important food product in my country,corn is seriously affected by grass damage.Under extreme conditions,corn will be reduced by more than 20%due to weeds.The traditional large-area full-spray pesticide weeding method cannot be applied to weeds on demand,which is easy to cause a lot of waste of pesticides,higher production costs,and serious environmental pollution,which affects the yield and quality of corn.In order to solve the above problems,it is necessary to apply precise variable application of corn weeds.Accurate identification and detection of weeds is the first prerequisite to achieve weed variable spraying.In this paper,for the problem of low accuracy of weed identification in complex fields,we selected corn and its associated weeds as the research object,combining deep learning and convolutional neural network features Extracted features,using the feature propagation ability of graph convolutional networks to improve the accuracy of weed identification,positioning,and segmentation,launched a deep learning-based corn field weed identification and segmentation technology research,which was the follow-up corn field weed The precise prevention and control of grass provides theoretical basis and technical support.The main research contents are as follows:(1)A field weed recognition network based on CNN and GCN was studied.This network uses CNN to extract the weed image sample features,build a graph model based on Euclidean distance based on the weed features,use Laplace transform to optimize the graph model,use GCN atlas convolution for feature propagation,and put the fused features into the classifier.Realize classification of weeds.The classification accuracy on the three types of corn,lettuce and radish weed datasets were 97.8%,99.37%,and 98.93%,respectively.(2)A weed segmentation model was constructed based on Mask R-CNN method.This model combines Faster R-CNN target detection and FCN semantic segmentation.First,we use the ResNet-101 network to extract image feature maps of weeds,use RPN(Region Proposal Network)to extract regional coordinate features,and then use the RoIAlign layer to obtain fixed-size feature maps.Use the output module to perform classification regression segmentation calculation on the feature map to complete the calculation of the specific orientation,category and contour of weeds.When IoU is 0.5,this model has a mAP of 0.853on the corn weed dataset,which is better than SharpMask and DeepMask’s 0.816 and 0.795.In order to test the segmentation effect of the weed segmentation model based on Mask R-CNN in the corn field,the model was used to test on the data set.The mAP value of the model is0.785,which can realize the weed segmentation in the corn field.(3)A variable spraying system based on Mask R-CNN field weed segmentation was constructed.The system obtains the weed image through the camera,calculates the position,label and pixel of the weed based on the Mask R-CNN segmentation model.According to the calculation result,the microprocessor adjusts the opening and opening time of the solenoid valve through PWM duty The grass is sprayed with variables.In the field variable spray test,the accuracy rate of the weed identification was 91%,the accuracy rate of the weed identification and accurate spraying was 85%,and the density of the spray of the weed pesticide sprayed accurately was 55/cm~2..It can meet the control requirements of herbicide variable spraying.