Semantic Segmentation And Diagnosis Of Apple Leaf Diseases Under Natural Conditions Diagnostic System Development

While the apple industry is closely linked to rural development and farmers’ personal interests,the production of apple foliar diseases can lead to the drying and shedding of large numbers of leaves,inhibit trunk growth and weaken immunity to external diseases,thus affecting the yield and quality of the orchard.In view of the damage caused by foliar diseases in apple cultivation,it is important to identify and protect against them accurately and scientifically.In this study,four common foliar diseases,namely rust,grey spot,brown spot and spotted leaf drop,were selected as the research objects,and the methods of apple foliar disease data enhancement,foliar disease spot semantic segmentation and the design and implementation of a disease diagnosis system were investigated.This study provides new ideas,methods and approaches for leaf disease segmentation and identification,and also provides technical and theoretical support for the subsequent segmentation of fruit and vegetable diseases.The research content and conclusions of the thesis are as follows:(1)Research on CycleGAN-based data enhancement method for apple leaf images.A CycleGAN-based method for apple leaf disease data enhancement is proposed to address the problem of unbalanced data between disease classes and the difficulty of obtaining certain disease data.The CycleGAN network is used as the backbone network to transform from healthy leaf images to disease images.To address the problem of poor quality of disease spot generation in the converted CycleGAN network,the CBAM attention mechanism is invoked to exclude the influence of interference information on the generated images,and the residual structure in the CycleGAN network is replaced with a densely connected convolutional structure to achieve high-quality conversion of diseased leaf data.The U-Net semantic segmentation model was used to perform semantic segmentation experiments on the dataset obtained from the traditional data enhancement method,the dataset generated by the original CycleGAN network and the dataset generated by the improved CycleGAN network respectively.The average segmentation accuracy was 69.25% and classification accuracy was76.4%;the average segmentation accuracy on the improved CycleGAN network enhanced dataset was 71.75% and classification accuracy was 79.2%,which improved the segmentation accuracy by 5% and 2.5%,and the classification accuracy by 2.88% and 2.8%,respectively,compared with the traditional data enhanced and original CycleGAN network enhanced datasets.The experimental results show that using the improved CycleGAN network method to construct the dataset can effectively improve the performance of the semantic segmentation network.(2)Study on semantic segmentation method for apple leaf spots.In order to improve the segmentation and recognition accuracy of apple leaf diseases under natural conditions,an improved U-Net semantic segmentation network structure was designed in this study.To address the problem of gradient disappearance and gradient explosion,a residual module is used in the encoder part to protect the integrity of feature information.To address the problem of too much information about the outdoor environment,the spatial attention mechanism(CBAM)was fused with the Convolutional Block Attention Module(CBAM)to focus on the more valuable feature information.The experimental results showed that the average segmentation accuracy,pixel accuracy and average classification accuracy of the improved UNet network reached 87%,94.2% and 91.75%,respectively,which were 14.5%,7.04% and12.5% higher than those of the original U-Net network.(3)Design and implementation of an image diagnosis system for apple leaf diseases.This thesis designs the apple leaf disease segmentation interface based on Py Qt5,and implements user registration and login,weight selection,model initialization,disease detection and control strategy recommendation functions.The diagnostic system is simple and straightforward,allowing users to segment and identify the four apple leaf diseases,understand the characteristics of the spots,the status of the disease and the control strategies,and to deal with the disease scientifically and accurately.