The Detection Of Cassava Leaf Diseases Based On Deep Convolutional Neural Network

Cassava is widely used,but cassava is susceptible to viral diseases.If diseased plants and disease types can be detected and accurately identified in time,cassava yields can be improved.Usually,cassava growers can seek the help of agricultural experts for visual inspection and diagnosis,but this method has disadvantages of low efficiency and high cost.With the fast development of deep learning in the field of image classification,it becomes possible to use convolutional neural networks to detect cassava leaves and quickly identify disease types.If a cassava leaf disease detection system based on deep convolutional neural network can be developed,cassava growers can realize automatic identification of cassava leaf disease types at home,thereby improving detection efficiency and reducing losses as much as possible.However,there are limitations to develop the detection system:(1)The size of the deep learning model at this stage is large.If the model is deployed to the cloud server,there will be slow response speed when multiple users access it at the same time.If the model is deployed locally to the users,there are certain requirements for the local device running the detection system;(2)Due to the difference in the angle of shooting cassava plants by different users,the presentation angle of the cassava leaf lesion area and the area ratio on the image will be different,and the above deviation will affect the generalization performance of the detection model;(3)The difference in appearance between cassava leaves with different diseases is small,and users may only use simple devices such as mobile phone to take samples of cassava plants for analysis by the detection system,but the acquisition accuracy is far less than professional equipment.The low image clarity and contrast make this difference more difficult to visualize.For solving the above problems,this thesis proposes the following countermeasures:(1)In order to make the detection model have a lower amount of parameters and computation,the paper does not use traditional convolution but depth-wise separable convolution to design the network structure according to the specific characteristics of cassava leaf images;(2)In order to enable the detection model to recognize cassava images from multiple angles and ranges,this thesis proposes a multi-scale inverted residual structure based on inverted residual and atrous convolution to enhance the multi-scale feature fusion ability of the model;(3)In order to enable the detection model to distinguish the small differences between cassava leaves with different diseases,combined with the actual characteristics of the inverted residual structure,this thesis proposes the split attention inverted residual structure based on versatile split attention mechanism,making the model adapt to the fine classification task of cassava leaf disease detection;(4)This thesis integrates the multi-scale inverted residual and the split inverted residual structures to introduce the multi-scale split attention inverted residual structure MSIR,and finally proposes the cassava leaf disease detection model based on Mobile Net backbone.The proposed methods are evaluated on the public dataset of cassava leaf disease detection,and the ablation experiment adopts the five-fold cross-validation method.The experimental results show that the proposed method is effective.In addition,this thesis develops a cassava leaf disease detection system running on the PC terminal based on the Tkinter GUI library and evaluate the execution efficiency of the system to verify the effect of the proposed model in practical applications.The experimental results show that cassava growers can efficiently realize single and batch detection of cassava leaf images through the above applications,thereby improving detection efficiency.