Identification Of Wokan Pests And Diseases Based On Deep Learning

The tangerine industry in Wuming,Guangxi is developing rapidly,with its output and sales increasing steadily.However,the traditional methods of relying on experience and artificial diagnosis of pests and diseases are too subjective,which may lead to a large amount of pesticide spraying blindly,resulting in pesticide residues in citrus.This is contrary to people’s scientific and healthy concept of diet.With the wide application of deep learning technology in various industries,people pay more and more attention to its application in the field of agriculture.This thesis aims to apply the deep learning technology to the detection of Tangerine diseases and insect pests,laying a solid foundation for the control of tangerine diseases and insect pests,and is of great significance for the sustainable and healthy development of Guangxi Tangerine industry.In this thesis,the identification and diagnosis methods of five kinds of diseases and pests,including woodlouse,yellow dragon disease,leaf miner moth,canker disease and sand derma,were studied,and two effective identification networks were designed.Compared to the base YOLOv7 network,the performance is improved by 3.2 percentage points.The details are as follows:(1)Since there is no previous research on citrus diseases and insect pests,and no corresponding data set is available,the students conducted field research in the citrus plantation in Wuming,Guangxi.4634 data on five species of citrus pests and diseases were collected and preprocessed to obtain a high quality data set.And named it CitrusPD(the Chinese name for it is the Mandarin Disease Data Set).(2)Considering that stacking convolutional layers in CNN network may cause information loss and thus affect accuracy,an improved model based on YOLOv7 was proposed,named YOLOv7 Mix.Combined with CNN and Transformer,this model introduces multi-head attention mechanism,which improves the recognition accuracy and convergence speed while enlarging the feature extraction field of vision.(3)Aiming at the problems of low resolution,too low pixel of the research object,and the background noise can easily cover the small target information in complex scenes,an improved disease and insect recognition model YOLOv7 HMix based on YOLOv7 Mix was proposed.By adding a cross-scale feature layer connection to the trunk feature extraction module and adopting the pixel-level feature fusion method,the model adds the pixel values of the feature points of the two branches to calculate the new pixel value feature map according to the weight,thus effectively improving the extraction ability of detailed features such as edges and textures,so as to achieve higher accuracy.