Research And Application Of Corn Disease Degree Grading Based On Convolutional Neural Networks

Corn is one of the main food crops in China.It has a significant impact on ensuring national food security and improving people’s living standards.During the growth of corn,it is susceptible to various diseases,especially common leaf diseases such as corn rust,large spot disease,and gray spot disease.These diseases have a high incidence rate and a wide range of infections,which seriously affect the yield and quality of corn.Therefore,rapid,efficient and intelligent disease identification and disease severity grading are of great significance for the precise prevention and control of corn diseases.With the development of computer vision technology,more and more scholars use convolutional neural networks to solve image recognition and classification problems and try to deepen the network structure to improve classification accuracy.However,this leads to a large number of model parameters,long training time and difficulty in deploying on mobile devices.To solve these problems,this thesis takes corn leaf disease images as the research object and focuses on the research of corn leaf disease identification based on convolutional neural networks and corn leaf disease image semantic segmentation based on deep learning from the aspects of model lightweighting,model depth transfer learning and disease severity quantitative evaluation.The main research contents and conclusions of this thesis are as follows:(1)The basic structure of convolutional neural networks is explored.Convolutional neural networks are combined with corn disease identification tasks.Various convolutional neural networks such as AlexNet,VGG,Goog LeNet,ResNet and EfficientNet V2 are trained using the constructed corn leaf disease dataset.Since most convolutional neural networks have a large number of parameters and occupy a large space,they have high requirements for hardware equipment,making it difficult to deploy models in practical applications.Therefore,this thesis adopts deep transfer learning and data enhancement strategies to further consider the model’s parameter quantity and space occupancy while ensuring the model’s recognition accuracy.The comparison results show that the lightweight network EfficientNet V2 has higher classification accuracy and smaller space occupancy,which is more suitable for building mobile corn disease identification systems.(2)The performance of semantic segmentation networks such as U-Net,PSPNet,SegNet and Deep Lab V3+ in complex background corn leaf segmentation tasks is compared.The optimal network is trained under different hyperparameter combinations to obtain a better hyperparameter combination.Experimental results show that Deep Lab V3+ performs better than other networks in complex background corn leaf image semantic segmentation tasks.The average intersection-over-union ratio of corn leaf segmentation reaches 84.17%.In comprehensive consideration,EfficientNet V2 is finally selected as the backbone network of Deep Lab V3+ to achieve model lightweighting while achieving good segmentation results.(3)In order to improve the position information extraction ability of Deep Lab V3+network,CA attention mechanism is introduced to improve semantic segmentation effect.The cross-entropy loss function in Deep Lab V3+ is replaced by Cyclical Focal Loss loss function to increase the weight of target part and solve the problem of data class imbalance in corn leaf disease spot segmentation dataset.The average intersection-over-union ratio of CA＿Deep Lab V3+ for spot segmentation reaches 83.21%,an increase of 2.07% compared with Deep Lab V3+.(4)Design of a corn leaf disease identification and severity grading system.A corn leaf disease identification and grading system was developed using the Spring Boot development framework,with the designed CA＿Deep Lab V3+ as the backend model,and the disease severity diagnosis was performed based on the national standard on the basis of the lesion segmentation results.Test results show that the system’s disease identification accuracy reached 95.77%,and the disease severity grading accuracy reached 85.70%.The system interface is simple and easy to use,and can achieve the expected disease identification and severity grading functions as well as common disease prevention and control science popularization functions.