Research On Segmentation Technology Of Maize Leaf Disease In Complex Background Based On DeepLabV3+

As the main food crop in my country,corn is an important feed and industrial processing raw material,and its output directly affects the sustainable development of agriculture,animal husbandry and the national economy.However,in the process of maize planting,maize leaves are often disturbed and damaged by various pests and diseases,and because of the wide variety of maize leaf diseases,some disease symptoms are similar,and it is difficult to identify them by naked eye observation.Traditional disease identification methods rely on human subjective judgment and require strong professional knowledge background.In large-scale identification scenarios,the identification speed is slow and the efficiency is low.With the development of computer hardware and software,deep learning is more and more widely used in crop disease identification,but it is less used for corn disease identification.Based on the Convolutional Neural Network(CNN)algorithm,this paper aims at the current corn leaf disease During the segmentation process,the segmentation accuracy of lesions is low,the data set is insufficient,and the training speed is too slow.A method for disease segmentation of maize leaves based on the improved DeepLabV3+network was constructed.The main research contents include:(1)Image collection and preprocessing of maize leaves.In this study,the original images of leaves with diseases such as maize leaf blight,maize gray spot,maize rust,and maize army worm were collected manually.In order to balance the number of 4 types of disease images and avoid overfitting of the model due to too few images,the research uses data enhancement methods such as image rotation,mirroring,image enlargement and reduction to expand the original image data,and use annotation tools to perform data augmentation.The maize leaves and diseased areas were marked,and finally the data set for the study was formed.(2)Semantic segmentation training based on DeepLabV3+network.This study uses the DeepLabV3+network for semantic segmentation training,in which the feature extraction network uses ResNet101,MobileNetV2,and HRNet three structures respectively.At the same time,according to the image semantic segmentation accuracy evaluation indicators PA(Pixel accuracy,pixel accuracy),mPA(mean Pixel Accuracy,average pixel accuracy),mIoU(mean Intersection over Union,average intersection ratio),FWIoU(Frequency Weighted Intersection over Union,Frequency weight crosscombination ratio)to evaluate the segmentation accuracy.The evaluation results show that when ResNet101 is used as a feature extraction network,the PA value is 91.53%,the mPA value is 85.77%,the mIoU value is 81.15%,and the FWIoU value is 84.72%.The segmentation effect of DeepLabV3+with ResNet101 as the feature extraction network is better.(3)Optimization design and implementation of DeepLabV3+network.Due to the poor generalization ability of the model trained on the imbalanced data set,it is prone to overfitting,and the model has serious bias.Therefore,this paper adds transfer learning based on the optimal DeepLabV3+model,adds an attention mechanism to the ResNet101 feature extraction network,and replaces the cross-entropy loss function in the DeepLabV3+model with the Focal Loss loss function.In the comparison of test results,the effect of the network model trained by transfer learning is obviously better than the original model,and PA,mPA,mIoU,and FWIoU can be increased by 0.53,0.62,0.68,and 0.85 percentage points respectively;for adding focalloss on the basis of using transfer learning The loss function increases PA by 0.92%,mPA by 2.64%,mIoU by 0.56%,and FWIoU by 0.86%;on the basis of using transfer learning,an attention mechanism is introduced into the backbone network ResNet101,PA is increased by 2.71%,mPA is increased by 5.59%,mIoU increased by 2.03%,FWIoU increased by 2.44%,indicating that the attention mechanism can effectively improve the accuracy of feature extraction;multi-path feature extraction can further increase PA by 2.38%,mPA by 3.34%,mIoU by 0.24%,and FWIoU by 1.56%.The experimental results show that the introduction of the optimization module can realize the optimization of the DeepLabV3+network to a certain extent.