Research On The Methods Of Identification And Lesion Segmentation Of Maize Leaf Disease Based On Deep Learning

Maize is the main cash crop in China,diseases and insect pests are one of the main reasons affecting the yield and quality of maize,rapid and accurate identification of maize diseases,timely and effective measures can minimize the losses caused by diseases.Traditional disease identification methods mainly rely on experts to assess the disease type and disease severity,which is time-consuming,laborious and subjective.The crop disease identification method based on deep learning has the advantages of high precision and fast speed,which provides an effective technical means for the scientific prevention and control of crop diseases.Therefore,this paper takes maize leaf disease as the object,studies the method of maize leaf disease identification and spot segmentation based on deep learning,and evaluates the severity of maize disease on this basis.The main research content of this article is as follows:(1)Dataset construction and preprocessing.First,the Syn-Plant Village dataset(Syn-PV)of plant disease images with complex backgrounds and the image dataset of maize Northern Leaf Blight captured by drones in the natural field are obtained through the network;The maize Northern Leaf Blight images were randomly cropped to construct a data set(NLB data set)containing maize Northern Leaf Blight leaves and healthy leaves,as the maize leaf disease identification data set.Finally,data enhancement and manual labeling of the lesion area were performed on the leaf images of the NLB dataset to construct a segmentation dataset(NLB-Seg dataset)for maize Northern Leaf Blight lesions.(2)Research on the identification method of maize leaf diseases based on the improved Swin Transformer network.First,a Multi-resolution Overlapping Attention module is added to the Swin Transformer network to build a plant leaf disease recognition model(Swin T-MOA)based on the improved Swin Transformer network.Then,use the Syn-PV dataset obtained from the network to pre-train the Swin T-MOA model,and test the identification effect.Finally,using the Swin T-MOA pre-training model to perform transfer learning on the NLB dataset,and combine it with Res Net50,Google Net,Vision Transformer,Swin Transformer four identification networks for comparison.The results show that the Swin T-MOA model achieves the best identification performance on the NLB dataset.(3)Study the segmentation method of maize Northern Leaf Blight image.First,the YOLACT++ instance segmentation network is improved,Res Net-101 is used as the feature extraction network,and the Convolutional Block Attention Module(CBAM)is added before and after the feature pyramid network(FPN)to build a lesion segmentation model.Then,the segmentation model is trained and tested with the NLB-Seg dataset and compared with the Mask R-CNN and YOLACT++ models.The results show that the improved segmentation model can achieve accurate segmentation of maize Northern Leaf Blight lesions.Finally,using the segmentation result of maize Northern Leaf Blight,the area of leaf diseased area and healthy area was calculated.Referring to the grading standard of maize Northern Leaf Blight,an experiment was carried out to evaluate the severity of maize Northern Leaf Blight at leaf scale and plant scale.The results show that the feasibility and effectiveness of the evaluation method for the severity of maize Northern Leaf Blight.