Research On Apple Leaf Disease Identification And Segmentation Based On Convolutional Neural Network

Accurate identification of apple diseases is the premise of accurate control of apple diseases.The traditional diagnosis method has poor accuracy and low efficiency,which affects the control effect.The deep learning technology based on computer vision can realize the feature extraction,type recognition and disease region segmentation of apple diseases,which can provide a theoretical basis for rapid and accurate control.Aiming at the problem of apple leaf disease identification and segmentation,this paper took five kinds of common apple leaf diseases（scab,black rot,cedar rust,gray spot and powdery mildew）as the main research object,studied and improved the existing classification and segmentation network,put forward a new method of apple leaf disease identification,and developed an apple leaf disease identification and segmentation system.Experiments showed that the improved classification and segmentation network proposed in this paper can accurately and effectively identify the types and regions of apple leaf diseases,and provided a new technical means for disease control in smart orchard.The main research results are as follows:（1）Study on classification algorithm of apple leaf disease.According to the characteristics of apple leaf disease,the disease classification method based on improved ResNet18 network was proposed.By adding channel and spatial attention mechanism branches on the basis of the original ResNet18,the attention of the network in the main disease areas was strengthened;In order to further optimize the network structure and improve the training speed,the random clipping branch of characteristic graph was introduced.Two improved networks ResNet18-CBAM-RC1 and ResNet18-CBAM-RC2 were constructed according to the different locations of channel and spatial attention mechanism branches added in ResNet18-RC network.The classification accuracy of the two networks in five categories of apple leaf disease were 98.25%and 97.69% respectively,which were higher than 93.19% of ResNet18.（2）Research on region segmentation algorithm of apple leaf disease.Aiming at the problem of insufficient feature extraction of the receptive field of ResNet18 in the backbone network of U-Net segmentation network,this paper improved: the high-dimensional features extracted by the backbone network were directly up sampled by 16 times,summed with the features output by the U-Net network,and the final segmentation results were predicted through the convolution layer and trained in parallel with the classification task.In the segmentation task of five kinds of diseases,the mean pixel accuracy of the improved U-Net-Up Cls network was 95.47%,and the mean Intersection over Union was 89.61%,which were better than 93.03%and 86.71% of U-Net,and improved the extraction ability of high-dimensional features.（3）Research and development of apple leaf disease identification system.In order to visualize the classification and segmentation results of apple leaf diseases,an apple leaf disease identification system based on Python is designed.The system had the functions of disease type introduction,disease identification,prevention and control suggestions and disease tracking.The disease recognition function can identify the disease images of apple leaves in batches,and input the disease images into ResNet18-CBAM-RC1 and U-Net-Up Cls network to realize the visual display of classification and segmentation.