Research On Apple Leaf Disease Detection Method Based On Deep Learning

Apple has become one of the most productive fruits in the world due to its rich nutritional value.The development of apple industry not only meets the needs of people’s living needs,but also brings huge benefits to the local economy.However,in the process of apple growth,there will be a variety of diseases in the apple leaf because of the influence of natural environmental factors.These diseases seriously hinder the normal growth of apple,affect the quality of apple,and cause huge economic losses eventually.Therefore,the construction of apple leaf disease detection model is great significance to improve the apple quality and reduce the economic loss of farmers.In this study,five kinds of common diseases of apple leaves were taken as the research object,and the research was carried out from three aspects of apple leaf image preprocessing,apple leaf segmentation method and apple leaf disease detection method.The apple leaf disease detection model was constructed,and the apple leaf disease detection small program was realized.The main work of this study is as follows:(1)Apple leaf image preprocessing.The original image of apple diseased leaves in this study consists of three parts: manual crawling,manual shooting and collection and partial public dataset.For the purpose of balancing the number of five types of disordered images,reducing the influence of natural factors on the pictures taken,and avoiding over-fitting due to the small number of images in the model.In this study,a variety of image augmentation methods were used to expand the original images,Then,the expanded images were scaled by bilinear interpolation method,and labeling tools were used to mark the apple leaf disease area.Finally,the PASCAL VOC format dataset and the mask dataset were produced.(2)Research on apple leaf segmentation model.The background of apple leaf image under natural conditions is complex,which will affect the detection effect model.Therefore,this paper builds an apple leaf segmentation model based on DeepLabV3+.The model adopts encoder decoder structure,in which the encoder part is mainly composed of feature extraction network and atrous spatial pyramid pooling.In order to reduce model parameters and improve model training efficiency,Xception with detachable depth was selected as the model feature extraction network.At the same time,in order to expand the receptive field of the model,make the model capture more apple leaf pixels from the complex background and enhance the segmentation ability of the model,make dense connection of the Atrous Convolution in Atrous Spatial Pyramid Pooling module in the model.The experimental results indicated that the Mean Intersection over Union and the Mean Pixel Accuracy of apple leaf segmentation model were 84.27% and 85.43% respectively.(3)Research on apple leaf disease detection model.The area of apple leaf disease spot is small and contains less characteristic information.The model is mainly composed of feature extraction network,region proposal network and classification regression network.For the sake of extracting more abundant features of apple leaf spots,the feature extraction network in the Faster R＿CNN model was changed into a Feature Pyramid Network which could integrate feature information of different scales.At the same time,the ROI Pooling in the Faster R＿CNN model was replaced with the Pr ROI Pooling to avoid the precision loss caused by quantization rounding in the ROI pooling.The experimental results showed that the mean Average Precision of apple leaf disease detection model can reach 84.45% on the dataset added with segmentation image.This paper designed and implemented apple leaf disease detection applet.Users can detect five kinds of apple leaf diseases and understand the relevant disease knowledge and prevention measures through this small program.