Classification Of Plant Disease Based On Improved EfficientNet And Adversarial Attack

Agriculture obtains products through manual training.Many plant diseases and pests will reduce the yield and quality of crops,and plant disease is a serious problem.Therefore,the classification of plant pests and diseases has attracted extensive attention in the academic world in recent years.There are many ways to classify plant diseases.Artificial diagnosis of disease types has strong subjectivity,and time consuming.Later,deep learning was applied to the identification of pests and diseases,and some achievements were made.The current deep neural networks have some problems such as low recognition rate or low efficiency.The paper proposed Sim AM-Efficient Net,which improves the attention module,the convolution module and the activation function.The recognition accuracy of Sim AM-Efficient Net is high,and training and recognition are fast.The existing attention modules generally have the following two problems.The first problem is that they can only refine features through channels and Spaces,which leads to inflexible attention weights.The second problem is that the structure is complex and the performance is influenced by a number of factors.It considers both space and channel,inferring three-dimensional attention weights without adding parameters in the original network.Therefore,this paper integrates Sim AM into Efficient Net.Features in feature maps generated by standard convolution are highly redundant.Therefore,this paper uses Ghost module to optimize a part of the standard convolution,with the standard convolution operation and linear operation combined to reduce the amount of calculation in the process of feature extraction.In addition,the original Efficient Net shows that the Swish activation function applies to modules 1 and9 of the model,as well as MBConv modules 2 to 8.While Swish activation functions perform better than Re LU on deeper models,they are less efficient.Therefore,the Hard Swish activation function is used instead of Swish.The Hard Swish activation function uses piecewise linear functions to reduce the number of memory accesses.In the training process,transfer learning and learning rate attenuation are introduced to improve the performance of the model.The experimental results show that the accuracy of the improved model is 99.31% in Plant Village.Res Net50 has an accuracy rate of 98.33%.Res Net18 has an accuracy of 98.31%.Dense Net’s accuracy rate was 98.90%.However,DNN is very vulnerable,and its adversarial examples in image classification deserves attention.It is very important to detect the robustness of DNN by adversarial examples.In this paper,GP-MI-FGSM is proposed by improving the GP-MI-FGSM algorithm through gamma correction and image pyramid.Then the adversarial examples generated by this method is used to adversarial train DNN.The error rate of the model proposed in this paper is 87.6% under GP MI-FGSM adversarial attack algorithm.The attack success rate of GP-MI-FGSM proposed in this paper is higher than that of other adversarial attack algorithms,including FGSM,I-FGSM and MI-FGSM.The adversarial examples generated by this method can be used for adversarial training.After adversarial training,the robustness of the model is further enhanced,and the performance can be improved.The accuracy of the Sim AMEfficient Net reaches 99.78% after adversarial training.