Research On Identification And Model Compression Of Plant Diseases Based On Relational Knowledge Distillation

In recent years,deep learning has made great progress in plant disease identification.However,these researches are difficult to be widely used in recognition owing to the problems of large model volume and complex computing cost.The above problems are more prominent when using UAV,Io T equipment and other equipment with scarce computing resources for plant disease detection.For the key problems in recognition and model simplification of plant diseases,this thesis proposes a structured model compression method based on knowledge distillation,which is to further reduce the parameters and model volume of convolutional neural network on the premise of ensuring the accuracy of the model.Our model achieves higher accuracy and better performance,and is easier to deploy to devices with limited memory and computing resources.The main work and contributions of this thesis are as follows:1.We design a plant disease recognition model based on multi feature fusion.Firstly,the median filter method is used to denoise the image.Then the color,texture and contour features of the plant leaf lesion are extracted by computer vision technology to establish the feature database.On the basis of using a large number of sample data in the feature database,BP neural network algorithm is used to optimize and adjust the neuron weight and learning rate of the model for many epoches.Finally,our experiment achieve more than 93% recognition accuracy.2.We propose a knowledge distillation method called LIKD based on layer interaction,and compress the above plant disease recognition model based on multi-feature fusion and other advanced convolutional neural networks.We conducte experiments on the data set of mango powdery mildew and anthracnose provided by the Plant Protection Institute of Guangxi Academy of Agricultural Sciences.The results show that the Distilled-Mobile Net model compressed by LIKD achieves an average recognition accuracy of 94.62% and greatly shortens the recognition time.3.For the problems of declining accuracy and single type of knowledge of student model in traditional knowledge distillation methods,this thesis proposes a knowledge distillation method called IRMKD based on instance relationship matrix.The IRMKD considers the correlation of samples in the process of knowledge distillation and combines the consistency between instance feature maps,then models the correlation through instance relationship matrixs(IRMs).As mentioned above,the algorithm propose in this thesis can effectively reduce the parameters of the model and ensure the recognition accuracy of the model.It provides a new method for recognizing plant diseases for devices with limited memory and computing resources.