Deep Learning Based Recognition And Grading Of Crop Leaf Diseases

"The people are the foundation of the country,and cereals are the life of the people." The cultivation of crops is related to the livelihood of the people,and the production of crops is causally related to our country’s social development and economic construction.Currently,in the agriculture field,one of the most important research topics is to improve crop resistance,such as resistance to diseases and insect pests,drought resistance,cold resistance,and salt resistance,where phenotype is an indispensable part of the breeding.The phenotype can screen out poorly resistant varieties and select good varieties,which will help to cultivate better quality offspring.Regarding the resistance of plants to diseases,in the traditional breeding process,experts will manually identify the diseases of the leaves,and select seeds with strong disease resistance as the next generation objects according to the type and degree of disease on the leaves.But only relying on manual methods to identify diseases requires a lot of manpower and time.This way of working has low efficiency and suffers from heavy workload.Therefore,there is an urgent need for a crop disease automated identification platform and method to replace these manual tasks.Therefore,this article first builds an automated phenotyping platform for crop leaves.This platform includes a desktop phenotype platform and a mobile phenotype platform.Its main function is to collect and analyze plant leaf pictures.Aiming at the problem that traditional image processing methods and deep learning methods cannot effectively fight against noise and deal with complex backgrounds,this paper proposes a robust crop disease identification method based on domain adaptation.This method uses a small number of unlabeled images from complex domain to constrain the feature layer of the neural network,which makes it more capable of resisting noise and processing complex background.Empirical experiment showed that the classification accuracy of plant disease images in the complex domain is higher than traditional methods,satisfying practical requirements.Considering that powdery mildew is a common crop disease,and the severity of the disease usually needs to be characterized in the phenotype,it is of great significance to obtain the proportion of powdery mildew spots to healthy areas.Therefore,this paper proposes a powdery mildew segmentation method based on a fully convolutional neural network.Compared with the traditional segmentation method,this method has higher segmentation accuracy.In multiple evaluation indicators,such as precision,recall,Io U score,Dice score,and pixel accuracy,our method is superior to traditional segmentation methods and meets the pratical phenotypic needs.Considering that the proposed segmentation model is too large to run on mobile devices,this paper also proposes a lightweight segmentation method based on superpixel segmentation and Gaussian Mixture Model.Compared with the traditional lightweight model,the model has higher recognition accuracy,and the model occupies only 200 MB of memory,which can be easily deployed in mobile devices to meet the needs of portable platform phenotypes.Based on the developed phenotyping platform and the algorithm proposed above,this paper finally developed two sets of crop disease automated phenotyping software,one for the desktop phenotype platform and the other for the mobile phenotype platform.Experiments show that the developed automated phenotyping software can meet practical needs and provide important technical support for automated phenotyping.