Research On Crop Disease Image Recognition Method And Application Based On Deep Learning

Crop disease is one of the main hazards that affect the yield and quality of agricultural cash crops.How to identify and judge the diseases from the beginning when crop diseases occur and prevent them in time plays a vital role in agricultural production.The traditional disease identification method completely relies on personal work experience and visual observation,and has the disadvantages of low efficiency,strong subjectivity,low accuracy and poor real-time performance,etc.With the continuous development and improvement of information technology,it is an urgent need to use a technical means to assist in crops diseases identification.In recent years,owing to its excellent generalization performance,deep learning technology and methods have developed vigorously and received widespread attention,which provides a new technical way for research and application of crop diseases identification.This paper focuses on theoretical research on crop disease images recognition methods based on the deep residual network and lightweight models in deep learning,and applied these lightweight models in the application development of crops diseases recognition.Specifically,our research includes an improved deep residual network,an application model distillation scheme,and further improvement of lightweight models.A more effective and reliable crop diseases identification algorithm is proposed,and the deployment of the algorithm on the mobile terminal is completed.In this thesis,the main work and research results are as follows:(1)In view of the drawbacks of the residual network(ResNet),this paper finetunes and improves its convolutional structure,and proposes a crop diseases identification method based on deep residual variant network(VRNet).In this method,the combination of 1×1 convolution and average pooling is used to replace the original 3×3 convolution process,which improves the expressive ability of the network model;at the same time,the squeeze-and-excitation(SE)module of the attention mechanism is added into the improved network model,obtaining the deep residual variant network(SE-VRNet),which effectively enhances the selective feature extraction ability of the network,and better solves the problems of scattered locations of crop diseases and difficulty in extracting features.Experiments show that the Top-1accuracy of this method on the Plant Village and Self Data datasets of crop diseases reaches 99.53% and 94.72%,respectively.(2)For the sake of mobile application,the SE module is added into the lightweight model Mobile Net(MNet)to obtain an improved lightweight model SE-MNet,and the knowledge distillation(KD)is used on the trained model to further improve the accuracy of the model to get a distilled SE lightweight network KDSE-MNet.In the end,the Top-1 accuracy of this model on the Self Data datasets of crop diseases reaches96.89%,an increase by 0.52% to 1.11% compared with other models,and the model parameters is only 2.35 M.