Research On Apple Leaf Disease Recognition Based On Generative Adversarial Network And Attention Mechanism

China is the country with the highest apple production and the most extensive planting area,but there is still a gap between the overall level of apple growing industry and developed countries.Apple leaf disease can cause leaf color change and a large number of shedding,which can lead to fruit quality decline or yield reduction.Therefore,it is important to build an apple leaf disease detection model to effectively and accurately identify apple leaf diseases,to accurately prevent and control apple diseases,to improve fruit quality,and to ensure stable economic income of fruit farmers.In this paper,we implement a data enhancement method based on generative adversarial network and an apple leaf disease identification model combining attention mechanism and residual network as the research object.Translated with www.Deep L.com/Translator（free version）（1）To address the problem of insufficient apple leaf disease data samples and inter-class imbalance,this paper analyzes the applicability of several typical algorithmic solutions,adopts a DCGAN-based data enhancement method to generate new sample data of apple leaf disease,and uses the generator of DCGAN to learn the data distribution of different disease type images,so as to generate new disease data and achieve the effect of expanding the apple leaf disease data set and achieving inter-class balance.The effect is to expand the apple leaf disease dataset and achieve inter-class balance.Finally,through a series of experiments,we found that this method overcomes the problem of insufficient samples and class imbalance while generating data that is more conducive to model feature extraction and improves network learning stability compared with traditional data enhancement methods.（2）To address the problems of low efficiency and complicated operation of traditional apple leaf disease recognition methods,and the difficulty of convergence of traditional deep convolutional networks by stacking invalid layers,the attention mechanism is added to deep residual networks to build apple leaf disease recognition models.In the residual block of Res Net50,the SE module is introduced so that the model rescales the inter-feature weight relationship,enhances the channel features of leaf disease spots and suppresses the channel features of healthy parts of leaves,improves the network’s ability to extract disease features,and enhances the robustness of the model.（3）By comparing multiple SE module introduction location strategies and comparing the accuracy rates of different strategies to determine the optimal location of the SE module,the Adam optimizer with high computational efficiency and low memory requirements was selected by controlled tests,and the tuning superparameter initial learning rate was 0.001 and the canonical coefficient was 0.0005 to achieve the optimization of the SE-Res Net50 network structure,and the final accuracy rate reached 97.83%.Subsequently,comparison experiments were conducted with Alex Net,VGG,Res Net34,and Res Net50 network models,and the experimental results showed that the improved deep residual network model based on the attention mechanism can improve the discriminative ability of the model while deepening the network depth,and has better accuracy and applicability in apple leaf disease recognition.