Research On Identification And Diagnosis Of Apple Leaf Diseases With Small Samples Based On Deep Learning

Due to the complex background of apple leaf disease images in the natural environment of the field,and the wide variation of illumination and shooting angle,a large amount of manpower,material resources,and financial resources are needed to collect and label a large number of disease data.In addition,due to plant diseases themselves,the incidence of some apple diseases is low,resulting in less,uneven,or incomplete data collected on apple leaf diseases.Therefore,it is of great significance to study how to solve the problem of low recognition accuracy of deep learning models caused by the small amount of apple leaf disease image data or the imbalance among disease categories.This paper takes apple-infected leaves as the research object,focusing on apple leaf disease recognition based on deep learning and apple leaf disease image semantic segmentation based on deep learning.To provide technical support for automatic,efficient,and accurate diagnosis of apple leaf disease.The main research contents and conclusions are as follows:(1)For complex low recognition rate under the background of the apple leaf disease needs a large amount of calculation,and the model was not suitable for deployment in mobile and embedded devices,deep learning network model was proposed for a lightweight Reg Net,and used in apple leaf single disease identification,vector analysis discusses the different training mechanism,and the optimizer will affect the performance of the model.The results show that,compared with only using neural network structure for feature extraction,fine-tuning all parameters of the pre-trained neural network structure can obtain higher classification accuracy.Compare four different lightweight convolutional neural network models(Reg Net,Mobile Net V3,Shuffle Net,and Efficient Net-B0)using different optimizers(SGD,Adam,RAdam,Ranger).It was concluded that the model Reg Net equipped with Adam optimizer can obtain better classification performance,and the average accuracy of the model on the validation set was up to99.8%.The comparison model Reg Net-Adam was equipped with different learning rates.When the initial learning rate was set to 0.0005,its accuracy on the test set reaches 99.23%.(2)Aiming at the problem that multiple diseases occurring simultaneously in a single apple leaf were similar and difficult to identify,an improved deep learning network model Reg Net was proposed to identify apple leaf diseases.Based on the baseline network model Res Net50,the influence of three different data augmentation methods on its generalization performance is studied.The experimental results show that compared with the other two expansion methods(online expansion and offline+online expansion),the network model with offline expansion can obtain better classification performance.(3)The influence of image background,training mode,and optimizer selection on the performance of four network models(Res Net,Res Ne Xt,Reg Net,and Res Ne St)was compared and analyzed,and the network model for apple leaf multi-disease recognition was optimized.The background clipping and transfer learning and optimizer selection strategies were used to test the four network models under different parameter combinations to get better parameter combinations.The experimental results show that the Reg Net network model trained by the Ranger optimizer has the best performance and achieves the highest recognition accuracy(99.6%)on the validation set.Finally,the analysis shows that the Reg Net network model trained on the data set obtained by background clipping has a stronger generalization ability,and the recognition accuracy of the original data set 1 and the background clipping data set 2 was 93.85% and 99.23%,respectively.(4)Compared the performance of three classical semantic segmentation network models Deep Lab V3+,PSPNet,and GCNet in the segmentation of apple leaves and diseased spots,and carried out optimization experiments on the combination of optimal hyperparameters and backbone network for the network models with good performance.The experimental results show that the Deep Lab V3+ network model has better segmentation performance than the other two segmentation network models(PSPNet and GCNet),with an average intersection ratio of 80.42%.The learning rate of the optimal hyperparameter combination is 0.001.When SGD was selected by the optimizer,Res Net50 pre-trained on the Image Net dataset was used as the backbone network of Deep Lab V3+ to obtain the best segmentation effect.The MPA and MIo U of this model reach97.26% and 83.85%,respectively.(5)Design of apple leaf disease degree autonomous diagnosis APP based on Android development environment.The Deep Lab V3+ network model trained on the self-built apple leaf disease dataset was deployed to the mobile phone,and the apple leaf disease recognition APP based on the Android system was developed.The results showed that the APP could not only realize the identification of apple rust and concentric ring marks but also segment the disease spots.Based on the segmentation results of the disease spots,the disease degree was diagnosed according to the national standards,which could provide an auxiliary diagnostic tool for agricultural technicians.