Study On Diagnostic Methods Of Citrus Huanglongbing On Microscopic Scale

Citrus is the largest fruit in the world,ranking first in area and production worldwide.Huanglongbing is the most devastating disease on citrus industry,causing huge economic losses.Its pathogenic bacteria are parasitic on the host bast and there is no cure yet,diagnosing the diseased trees and digging them up is one of the most effective means of huanglongbing prevention and control.However,plants infected with huanglongbing have no obvious symptoms in the early stage,and it is difficult to achieve accurate diagnosis of huanglongbing in citrus canopy or leaves.Therefore,in this paper,we conducted a study on huanglongbing detection based on microscopic images of citrus main leaf veins in order to achieve a more early diagnosis.On the other hand,there are phenotypic differences in the tissue regions such as bast,xylem and pith in the main leaf veins of citrus before and after the disease,and exploring the traits of different parts of citrus plants can resolve the effects of stress factors on their phenotypes from different scales.Therefore,this paper explores the method of segmentation of citrus main leaf veins instances,so as to achieve automatic and accurate quantification of each tissue region of the main leaf veins.Finally,image segmentation techniques are combined to enhance image features,which are used to construct a diagnostic model of huanglongbing with higher detection accuracy.The details of the study are as follows:Firstly,four diagnostic models of huanglongbing were constructed using ImageNet pre-trained light convolutional neural networks MobileNet-V2 and MobileNet-V3 and deep convolutional neural networks ResNeXt50 and ResNeXt101.Using independent sample testing,it was found that the diagnostic models of citrus huanglongbing based on lightweight convolutional neural networks were poorly diagnosed,with accuracy rates of 73.7%(MobileNet-V2)and 71%(MobileNet-V3),respectively,and a high rate of misclassification for undeveloped and developed diseased samples.The overall diagnostic accuracy of deep convolutional neural network was significantly higher than the former,and the 91.1% accuracy of ResNeXt101 was higher than the 87.8% accuracy of ResNeXt50,and the misclassification rate for the diseased samples was lower,which could achieve better diagnosis of huanglongbing.Secondly,the effect of traditional segmentation algorithms such as threshold segmentation,K-means clustering segmentation and interactive segmentation based on graph theory and the instance segmentation algorithm of mask region convolutional neural network(Mask R-CNN)based on deep learning on the instance segmentation of citrus main leaf veins was compared,and the the limitations of the image processing segmentation algorithm for segmenting each tissue region in the cross-section of citrus main leaf veins are discussed,and then an improved Mask R-CNN citrus main leaf vein instance segmentation model is proposed,using the feature pyramid network(FPN)and the residual network Resnet50 as the backbone feature extraction network,and adding a new region of interest alignment layer on the mask branch.A new region of interest align(ROI-Align)layer is added on the mask branch to improve the segmentation accuracy of the mask branch.The results showed that the average accuracy of segmentation AP(IoU=0.5)of Mask R-CNN model was 98.9%,89.8%,95.7% and 97.2% for pith,xylem,bast and cortex cells,respectively,and the average accuracy of segmentation mAP(IoU=0.5)value was 95.4% for the four tissue regions.The accuracy was improved by 1.6% compared to Mask R-CNN without adding ROI-Align to the Mask branch.Finally,a citrus huanglongbing diagnosis model with enhanced features is proposed by improving the model and pre-training method,which improves the model detection performance by unsupervised pre-training with the addition of enhanced feature network.The training method uses unsupervised learning mechanism combined with migration learning algorithm to solve the small sample training problem.In terms of model improvement,an enhanced feature network is designed based on the Yolact model,and the segmented features are fused with the original data set after segmentation of the huanglongbing feature region to achieve feature enhancement.On the detection side,a non-convolutional detect transformer model is used for detection output.After the independent sample validation,the model can locate and classify the huanglongbing regions more accurately,and the accuracy of huanglongbing diagnosis reaches 96.2%,which is better than 94.8% without the enhanced feature network,and has a better performance of citrus huanglongbing detection.