Recognition Of Wheat Stripe Rust Disease Grade Based On Deep Learning

Crop pests and diseases are important biological disasters in agricultural production and have always been one of the main factors limiting high yield,high quality,high efficiency,ecology and safety in agriculture.The Food and Agriculture Organization of the United Nations(FAO)estimates that the annual loss due to pests,diseases and weeds worldwide accounts for about one-third of the total food production,including 10% loss due to diseases,14% loss due to pests and 11% loss due to weeds.Stripe rust is one of the most important diseases of wheat,and a moderate occurrence of wheat stripe rust can cause a 20-30% yield reduction in wheat,and in severe cases can lead to a yield reduction of more than60%.At present,the control of wheat rust is still based on chemical control,but the long-term use of chemical pesticides can lead to resistance,environmental pollution,land caking,and increased production costs.And disease surveys are mostly manual identification,timeconsuming and subjective,resulting in more pesticide application measures.In recent years,a variety of remote sensing data are widely used in the monitoring of wheat,rice,corn and other bulk food crop diseases and pests,to provide technical support for the accurate application of crop diseases,but because of the high cost of equipment,data acquisition when the weather requirements are high,the monitoring of the disease has a certain limitation of time.With the continuous development of machine learning algorithms,a deep algorithm that can autonomously find a large number of image features and learn them has been widely used.Using images taken by cell phones combined with deep learning algorithms to identify crop diseases has become a more convenient and rapid means,which can provide more powerful technical support for accurate application of wheat.In this paper,we use RGB images taken by cell phones as the data source and deep learning as the basic framework to conduct the following research on the method of identifying the disease level of wheat stripe rust.(1)Wheat leaf image acquisition and enhancement pre-processing.In this paper,we use cell phones for image acquisition of wheat stripe rust at different disease levels in the wheat breeding experiment field of Henan Agricultural University.In order to restore the wheat growth environment,the images are acquired on both cloudy and sunny days,and the shooting process is easily affected by different lighting shooting angles.In order to improve the model generalization ability and prevent the overfitting problem of the model in the training process,the images are first subjected to wavelet transform,Gaussian low-pass filter,and median.To improve the model’s generalization ability and prevent overfitting problems in the training process,wavelet transform,Gaussian low-pass filtering,median filtering,scaling,and geometric transformation were first applied to the images.Then a scripting program was written in python to uniformly scale the images without deformation of the leaves in the figure.Label Img tool was used to label the target areas;finally,the wheat stripe rust disease level dataset was constructed.(2)Research on complex background segmentation methods.To reduce the influence of complex backgrounds in wheat leaf images on disease level detection and to address the problem that it is challenging to segment complex backgrounds such as other unrelated leaves,bare ground,and weeds,s included in the images,a segmentation model combining YOLOv5 s and Grab Cut is constructed in this paper.In order to verify the effectiveness of the segmentation method in this paper,Io U and PA are mainly used as segmentation effect evaluation indexes for comparison and analysis.The PA of K-means and Deep Lab V3 is above80%,and Io U is around 80%,and both PA and Io U of the segmentation algorithm in this paper can reach above 90%.The Grab Cut algorithm combined with YOLOv5 s can achieve automated leaf segmentation,and compared with traditional methods of background removal such as K-means clustering threshold segmentation,the algorithm proposed in this paper can still obtain a high segmentation rate when there are many plants with the same characteristics in the background or other wheat leaves.(3)A study on the disease level recognition method of wheat stripe rust.In this paper,Res Net50 is used as the base network model architecture.The original model is modified by adding the Inception module for multi-scale feature extraction to expand the perceptual field and extract finer-grained image features.The revised model has improved evaluation indexes such as accuracy,recall,and average accuracy.The experimental results showed that the recognition rate of the B-Res Net50 network model(Better-Res Net50)on the wheat stripe rust leaf dataset was 97.3%,which was a significant increase in accuracy compared with Inception V3(87.8%),Dense Net121(87.6%),and Res Net50(88.3%).