Application Of Convolution Neural Network In Classification And Detection Of Apple Diseased Leaves

Most of the existing methods for detecting diseased leaves of plants use manual experiential judgment or laboratory spectral analysis,which consumes a lot of manpower,material resources and time.At present,the detection technology based on machine learning has become a research hotspot,but some of the proposed detection technology of diseased leaves has some problems,such as the classification is not fine enough,and the model generalization ability is weakened with the increase of data volume.Therefore,it is of great significance for precision planting to improve the detection method of apple diseased leaves,improve the accuracy of classification of diseased leaves and improve the generalization ability of detection model of diseased leaves by using the new neural network technology.In view of the fact that a single detection algorithm is not suitable for the detection of diseased Apple Leaves in a complex background,this thesis proposed an Apple Leaves Mask R-CNN(Apple Leaves Mask R-CNN)model for the segmentation of diseased Apple Leaves,and then combined with the improved Dense Net diseased leaf classification model for the detection of diseased Apple Leaves.Firstly,Feature Pyramid Network(FPN)and Residual Networks(Res Nets)were used in apple leaf segmentation model to limit the potential Feature extraction from end-to-end convolutional neural Network images due to different sizes of diseased leaf regions in the original data set.From the perspective of image features,the multiscale fusion of feature images extracted from each layer is carried out by using the multi-layer structure horizontally connected by FPN,so that the feature images fused from each layer have multi-scale feature information,and the original image is segmented with fine granularity.Secondly,in the classification model of apple diseased leaves,some of the proposed diseased leaf detection techniques focus too much on a certain kind of image features and ignore the generalization ability of the model,resulting in low accuracy of apple leaf classification and high cost of model training and testing time.On the basis of studying the structure of the diseased leaf classification model,the apple leaf classification model based on improved neural network was adopted in this paper.In the data preprocessing stage,data augmentation method is used to expand the input data to the data distribution balance,and an evenly distributed mask layer is added to the expanded training set to alleviate the over-fitting of neural network training.The hyperparameter Generalized Mean Pooling(Ge M)strategy is used to improve the original Pooling method to increase the contrast of feature maps and better extract feature information from the convolutional layer.A Label Smoothing(LS)loss function with Smoothing index was used to improve the original cross entropy loss function,alleviate the model’s overconfidence in classification due to Label ambiguity,improve the model’s generalization ability and improve the classification accuracy.Finally,in order to verify the validity of the proposed and improved algorithm,Plant Pathology 2020-FGVC7 dataset was used as experimental data,and AL Mask R-CNN was used to segment the image of apple diseased leaf dataset,and the segmentation result was taken as the test set of the detection model.Combined with the improved diseased leaf classification model,apple diseased leaf detection model was used.The simulation results show that al-Mask R-CNN model’s pixel accuracy,cross ratio and dice coefficient are 0.9633,0.9316 and 0.9615,respectively,which are better than Res-Unet model.Denes Net combines Grid Mask,Generalized Mean Pooling(Ge M),and label smoothing for an improved apple diseased leaf classification model in Plant Pathology The overall classification accuracy,Kappa coefficient and recall rate of 2020-FGVC7 data training set were 0.9835,0.9881 and 0.9665,respectively,which were better than the original model classification results.The classification evaluation index of the test set segmented by Al-Mask R-CNN was better than that of the original unsegmented apple diseased leaves.In order to further verify the generalization ability of the model,this thesis applied the improved diseased leaf detection model to the mixed data set of pepper,potato and tomato diseased leaves,and the test also showed high accuracy and recall rate.