Recognition And Application Of Apple Leaf Diseases Based On Deep Learning

Apple planting process is easy to be infected by diseases,which affects the healthy development of Apple industry in China.Apple leaf disease is a kind of apple disease located on the fruit tree leaves.Apple leaf disease is a disease of apples whose diseased parts are located on the leaves of fruit trees.The leaves usually discolor or even fall off after diseased,thereby weakening the disease resistance of the tree,reducing the yield and quality of apples,and causing economic losses.Therefore,it is of great significance to quickly and accurately identify the categories and severity of apple leaf diseases for precise prevention and control of apple diseases and to reduce the economic losses caused by them.In the traditional sense,the most commonly used method of disease identification and severity estimation is identified by plant protection experts based on previous experience,but due to the subjectivity of identification means and the complex external morphology of crops,it is difficult to accurately identify them.The rapid development of computer technology makes it possible to automatic recognition of apple leaf disease,a large number of researchers use computer vision technology for identification,and most of them use traditional image processing methods.This method is mainly to manually select and extract the characteristics of diseased leaves,which is not only time-consuming and laborious,but also has poor universality and migration.In recent years,convolutional neural networks have achieved excellent results in the field of image recognition.Instead of manually selecting features,convolutional neural networks automatically extract image features through network training,which is more efficient than traditional image processing methods.This paper takes apple scab,black rot and rust leaves as the research objects,and uses convolutional neural network to identify and estimate the severity of the disease.The main research contents and conclusions of this paper are as follows:(1)An image data set containing three kinds of apple leaf diseases including scab,black rot and rust was established.The dataset is mainly composed of disease images with two different backgrounds,one is a simple background,which is taken in a laboratory scene,and the background is usually a single white and blackboard.A total of 892 images have been collected,and the other is a complex background,which is Shooting in the real orchard scene,representing the real growth environment of leaves in the orchard,a total of 998 photos were collected.Using the method of data enhancement,the image data was expanded to10020 pieces and normalized,and finally the image dataset of apple leaf disease was established.(2)Based on Shuffle Net v2 lightweight convolutional neural network model,an improved Shuffle Net＿ours apple leaf disease recognition model was proposed.The model first improves the original basic residual unit,by introducing a two-layer depthwise separable convolution and SE channel attention module to improve the feature extraction capability of the model,and by removing the point convolution at the end of the main branch to reduce the complexity.Using the improved basic residual unit to build the network model.In order to make the network extract enough features in the first convolutional layer,the number of output channels of the first convolutional layer and the maximum pooling layer is expanded from 24 to 32.The fully connected layer of the model is initialized,and the original Softmax classifier is replaced with the Softmax classifier of 3 targets.The results show that: The model can effectively identify the categories of apple scab,black rot and rust leaves,with an average recognition accuracy of 96.81% in the test set.It was compared with Mobile Net v2,Mobile Net v3 large,Ghost Net,Shuffle Net v2 1× and Shuffle Net v2 1.5× models under the same experimental conditions.This model has certain advantages in recognition accuracy,model complexity,number of model parameters,model size and other aspects,and has good comprehensive diagnosis performance,which can meet the requirements of fruit farmers for disease recognition accuracy.(3)Based on the YOLOv5 target detection model,an improved YOLOv5＿ours apple leaf disease detection model was proposed,and the severity of apple leaf disease was estimated based on the model.In this model,the part before SPP module in the original backbone feature extraction network was replaced by Shuffle Net＿ours model with Conv5 layer and FC layer removed,and the CSP2＿x module with a large number of parameters behind SPP module was replaced by depth separable convolution.In order to further reduce the computation of the network,the number of output channels in the neck network was setted to128.The results show that the m AP@0.5 of the model for the detection of three diseased spots and leaves reaches 79.5%,and the FPS reaches 151f/s.Under the same experimental conditions,the model was compared with the target detection models of YOLOv3,YOLOv5 s,YOLOv5m,YOLOv5 l and YOLOv5 x in m AP@0.5,FPS,model size and other aspects.This model has more excellent performance in all aspects.Finally,according to the result of target detection of the model,the ratio of the total area of detection frames of various spots and leaves in a single image was calculated,and the severity of disease was estimated by referring to the classification standard of disease severity.(4)In order to better realize the application of apple leaf disease recognition,an apple leaf disease recognition and severity estimation system was designed and developed.Based on Flask and Vue framework,the system integrates image uploading,disease recognition,disease severity estimation,result display and other functional modules,which can be used as a daily diagnosis tool for apple leaf disease by fruit farmers.The method and application of apple leaf disease identification and severity estimation proposed in this paper can realize effective identification and severity estimation of apple scab,black rot and rust,and can provide ideas and references for designing efficient and lightweight disease diagnosis models.