Construction Methods Of Cucumber Disease Diagnosis Model Based On Deep Learning

Cucumber is one of the most widely planted vegetables in the world and is also an important greenhouse crop in China.Cucumber downy mildew and cucumber bacterial angular spots are easy to occur in greenhouses,which can easily cause great economic losses to farmers due to their short onset cycle and strong infectivity.With the rapid development of computer technology,the performance of plant disease diagnosis models has also been improved.Many diagnosis models designed based on convolutional neural networks have achieved excellent diagnostic results.However,many convolutional neural networks are trained on datasets with simple backgrounds or low complexity backgrounds,in this way these models have low applicability.In addition,disease diagnosis models designed based on convolutional neural networks often have relatively high parameters and calculation amounts,and also consume high computational energy,so they are difficult to apply in edge devices with limited computation,storage,and limited battery resources.Aiming at monitoring greenhouse cucumber diseases,this paper constructed cucumber disease diagnosis models for mobile phones and greenhouse equipment and realized the comprehensive monitoring of cucumber disease in the greenhouse by combining the two kinds of equipment.The main content of the research is divided into the following aspects:1.The cucumber disease dataset with complex backgrounds is constructed,and a method is proposed to divide this dataset according to the complexity of backgrounds.For the cucumber dataset,a cucumber disease detection model with high robustness in real fields is constructed.Firstly,Efficient Net B1 and Extra Layers are used as the backbone network of the model.Secondly,considering the cucumber disease leaves have various scales,six feature maps with different sizes are selected to participate in the detection procedure.Finally,to improve the semantic information content of the shallow feature maps and the detailed information of the deep feature maps,the adaptive feature fusion method is used to calculate the weight of each feature map for fusion.Many experiments show that the disease detection model constructed in this paper can effectively improve the robustness of the model in complex environments,and the m AP value of cucumber disease data sets with different background complexity has reached 85.52%.2.This paper constructed a lightweight cucumber disease classification model called CNNLight to reduce the model’s demand for computation and storage resources of edge equipment.Firstly,the basic modules of the lightweight model are constructed by depth separable convolution modules,conventional convolution modules,and squeeze-and-excitation modules.Then,the model structure suitable for this task is explored.Experimental results show that the lightweight model constructed in this paper has a classification accuracy of 90.2% on the cucumber disease dataset with complex backgrounds.The model size is 0.479 MB and the FLOPs is 0.03 GFLOPs.The low computational cost and small model size make the model have strong applicability to edge devices.3.An additive feature extraction method is used,that is,L1 normal form is used to measure the similarity between feature maps and convolution kernels,and the working time of the edge device is extended.Firstly,the general depth separable addition feature extraction modules and squeeze-and-excitation modules are constructed by the addition feature extraction method.Then,the above addition feature extraction modules are used to reconstruct the lightweight convolution disease classification model constructed in this paper and obtain the lightweight addition cucumber disease classification model ADDLight.The experimental results show that compared with the convolutional neural network with the same structure,the computational energy consumption of the additive feature extraction model is reduced by 96.1%,which can effectively prolong the working time of the edge device after a charge.