Research And Application Of A Deep Learning-Based Diagnosis System For Common Potato Leaf Diseases

The potato is the fourth largest food crop in the world,after maize,wheat and rice in terms of production.Widely grown in China for its short growth cycle and rich nutritional content,it has made a huge contribution to increasing China’s food production and boosting farmers’ incomes.In potato cultivation,disease is an important factor limiting its yield.Traditional disease diagnosis mainly relies on farmers’ visual identification,which is highly subjective and less accurate,and it is difficult to distinguish accurately between some diseases with similar symptoms.In this study,common diseases of potato leaves were used as research objects,disease images were collected to construct disease datasets,and the datasets were expanded by data enhancement methods to improve the number and diversity of the datasets.Four classical deep learning image classification networks were used to train the dataset,and the classification network with the highest accuracy in disease recognition was selected as the base research network.The performance of the base network was analysed and improved to obtain a diagnostic network for common potato leaf diseases based on deep learning algorithms.In order to further improve the performance of the diagnostic network,this study used response surface analysis to construct a local hyperparameter spatial portrait of the network,and carried out hyperparameter tuning and hyperparameter configuration optimisation.The experimental results showed that the hyperparameter tuning improved the diagnostic performance of the network and enabled more accurate diagnosis and classification of common potato leaf diseases.Finally,this study deployed the common leaf disease diagnostic network on a GUI so that the diagnostic network can be operated more conveniently through GUI interaction buttons.The specific research in this thesis is as follows:(1)Production of potato leaf disease image datasets and selection of diagnostic algorithmsThe potato leaf images were collected by a collection device and the dataset was expanded using data enhancement.The dataset was constructed using common potato leaf without disease,potato early blight leaf and potato virus leaf as the research objects.Four mainstream image classification networks,VGG16,Mobilenet V1,Resnet50 and Vit,were built and trained based on the potato leaf disease dataset.The VGG16 network with the highest accuracy was selected as the base network for potato leaf disease diagnosis.(2)Improvement research based on the VGG16 networkIn-depth analysis of the diagnostic performance of the VGG16 network was carried out,and in response to its shortcomings,the network structure of "convolutional layer +spreading layer + fully connected layer" was changed to "convolutional layer + global average pooling" by incorporating the CBAM attention mechanism,introducing the Leaky Re LU activation function to participate in the learning training and changing the VGG16 network from "convolutional layer + spreading layer + fully connected layer" to "global average pooling".The improved VGG16 S network has 15 M parameters,and the recognition accuracy of the test set is 97.98%.Compared with the number of network parameters before the improvement,the number of network parameters is only 1/10 of it,and the accuracy is improved by 0.75 percentage points.The need for a lightweight diagnostic network is met with improved accuracy.(3)Hyperparameter search and comparative evaluation of the diagnostic networkFor the optimization problem of hyperparameter adjustment of VGG16 S network before training,this thesis uses response surface analysis to design the hyperparameter seeking test of the network.A Plackett-Burman screening test is designed to select hyperparameters with strong correlation with the recognition accuracy of the diagnostic network;then a steepest climb test is designed for the strongly correlated parameters to further narrow down the optimal parameter setting range;then a central composite(CCD)test is designed to solve the functional relationship between the hyperparameter combination and the recognition accuracy of the network,to construct a 3D response surface of the relationship between hyperparameters and the recognition accuracy of the diagnostic network The 3D response surface of the network is inscribed,and the optimal hyperparameter combination is predicted based on the parameter space;finally,the effectiveness of the proposed parameter finding method is verified by comparing VGG16 S with VGG16,Mobilenet V1,Resnet50 and Vit mainstream image classification networks for quantitative and qualitative analysis.It was verified that the optimal parameter combinations obtained by the method yielded reliable results and the recognition accuracy of the network was effectively improved.(4)Deployment of a network for diagnosis of common potato leaf diseasesUsing the modular programming ideas of Py Qt5 and Py Charm,the running logic of the GUI user side,data side and server side was designed to deploy the potato leaf common disease diagnosis network to the human-machine interface.This deployment allows farmers to quickly diagnose common potato leaf diseases in a simple and convenient way.After obtaining the diagnosis results,farmers can guide their potato disease control practices based on the corresponding descriptive information and expert control recommendations.