Research On Image Classification Of Tree Diseases And Pests Based On Image Enhancement And Feature Fusion

Urban greening construction is a great project undertaken by the state to improve people’s quality of life and benefit the people.With the improvement of people’s material living standards and the vigorous implementation of the national spiritual civilization construction,the people’s urgent expectations for a better life make urban greening development face higher requirements.Green trees are often infested by diseases and pests,which not only seriously affects the beauty of the city,but also brings a series of economic losses.Therefore,accurate and efficient classification of tree diseases and pests is a necessary prerequisite for targeted prevention and treatment of trees,and it is also a key part of improving the level of urban greening.Aiming at the problems of low efficiency,strong subjectivity and high misjudgment rate of traditional manual observation and empirical diagnosis classification methods,the use of deep learning technology to automatically classify images of tree diseases and pests is considered.However,some of the collected pictures of tree diseases and pests may have problems such as low light,which brings a lot of challenges to the subsequent feature extraction.In addition,due to the small area of diseases and pests,it is difficult to use traditional image classification algorithms to ensure the accuracy of tree disease and pest classification.In view of this,this paper takes the collected images of tree diseases and pests as the research object,and studies a low-light image enhancement algorithm based on double branch fusion,as well as a disease and pest image classification algorithm based on feature reuse and channel attention mechanism,so as to improve the classification accuracy of disease and pest images.The main research contents are as follows:(1)Research on low-light image enhancement method based on double-branch fusion.Aiming at the problem that it is difficult to obtain high-quality information due to the lack of light in the pictures of tree diseases and pests collected in the natural environment,a low-light image enhancement algorithm based on double-branch fusion is proposed.First,the input image is decomposed into two components,reflectance and illumination,which are processed separately.On the one hand,the illumination mapping function is learned from real data,allowing flexible definition of exposure levels.On the other hand,the illumination map is used to guide different regions to perform differential reflectance recovery,and finally the two components after the operation are fused and output to achieve effective enhancement of low-light images.(2)Research on image classification algorithm of tree diseases and pests based on feature reuse and channel attention.Aiming at the problem that the traditional convolutional neural network model often loses some shallow detail information after multi-layer convolution,which leads to the problem that the classification accuracy of fine-grained images such as diseases and pests is not ideal,an image classification algorithm based on feature reuse and channel attention is proposed.Select Inception V4 as the backbone network,on the one hand,the output of the basic feature extraction module Stem is spliced and fused with the shallower features,and then the basic fusion features are output for multi-scale convolution to improve the model’s utilization of shallow features.On the other hand,the global semantic information obtained by multiscale convolution is used to calculate the attention weight,which guides the basic fusion feature map for channel attention weighting.Then,the corresponding elements of the weighted basic fusion features and the original multi-scale features are added to obtain advanced fusion features,which further improves the feature expression ability of the classification model for pest images.Finally,a new loss function is introduced to solve the imbalance problem of dataset samples.(3)Design and implementation of tree disease and pest image classification system.On the basis of the above related methods and technologies,a C/S-based image classification system for tree diseases and pests is designed and implemented.The system includes functions of user management,low-light image enhancement,image classification of tree pests and diseases,query of disease and pest prevention and control opinions,and data storage.Finally,the system is tested experimentally to verify the availability and efficiency of the system.