Research Of Cassava Leaf Disease Recogtion Based On Attention Mechanism And Data Enhancement

As the sixth largest food crop in the world,cassava has great application value,which can be processed into food in various ways and can also be applied to the economic market.However,cassava is vulnerable to viral diseases,making it difficult to guarantee yield and quality.In the early days,experts were required to go to the field to judge and treat the disease types of cassava leaves,but this method was time-consuming and labor-intensive,and experts were required to have rich diagnostic experience.At present,recognition technology based on deep learning can greatly improve the efficiency of identifying disease types,but it is difficult to deal with dataset with complex backgrounds and unbalanced samples,what’s more,the recognition accuracy of network models is still not high enough.Therefore,this thesis proposes an improved residual neural network model to improve the recognition accuracy.First,an attention module is embedded into the model to enhance the network model’s attention;and then,data augmentation techniques are used to solve the problem of too few disease samples in the dataset and improve the generalization ability of the model.In view of the problems of high noise,complex background,and poor focus in the dataset,this thesis proposes to embed SE,CBAM,and PSA attention modules behind the convolutional layer of the Res Net50 network,to improve the model’s ability to extract disease features.Experimental results show that the recognition accuracy of the model is increased by 1.3% by embedding the PSA attention module behind the Conv4 layer in the Res Net50 network architecture.In order to deal with the problem of too few samples of Cassava Bacterial Blight（CBB）disease category in the dataset,this thesis uses Cycle Generative Adversarial Network（Cycle GAN）to process healthy cassava leaf images to generate images of cassava leaf with CBB disease features.The experimental results show that after increasing the number of samples of the CBB disease category,its test accuracy is improved by 13%,and the recognition accuracy of other disease categories is also improved to varying degrees.Among them,the recognition of Cassava Green Mottle（CGM）disease category is particularly obvious,at 7%.The recognition accuracy of the network model is improved by 3.5% up to91.3%.This thesis uses the attention mechanism combined with the data augmentation method to effectively improve the performance of the network model for cassava leaf disease recognition.The model in this thesis can accurately identify the disease types of cassava leaves,which helps farmers to better treat cassava leaf diseases symptomatically,thereby improving farmers’ operational efficiency.The research results of this thesis promote the research development of cassava leaf disease recognition,and provide research ideas for other similar plant disease recognition.