Research On Apple Leaf Disease Classification Based On Deep Learning

Shaanxi,as the province with the largest apple cultivation area and the largest yield,ranks first in the country in terms of the quality of apples grown.To ensure the quality of the fruit,the prevention of apple diseases during planting is a top priority.The traditional method of decontamination is to rely on artificial production experience to determine the disease type of apple leaves,and carry out corresponding decontamination operations.This method not only requires agricultural experts to have rich experience in disease identification and professional decontamination technology,but also has low practical operation efficiency and high labor costs,and cannot be used as an objective basis for judging large-scale orchard planting.To solve the above pain points,this thesis uses high-efficiency and low-cost deep learning methods to construct an apple leaf detection model and an apple leaf disease identification model according to the two-stage research idea of first detecting the blade and then identifying the disease.The main work of this thesis is as follows:(1)Research on the detection method of apple leaves.In this thesis,the Faster R-CNN detection model is used as the basic network for apple blade inspection tasks.First,more than6,000 apple leaf images were manually labeled for training the YOLOv5 s automatic labeling network,and then more than 20,000 unlabeled images were sent to the network to complete the automatic labeling of blades,and the apple leaf sample dataset was generated after manual verification.Then,aiming at the problem of insufficient expression ability of feature extraction network,it is proposed to use the Feature Pyramid Networks for the fusion of multi-scale blade feature information.Then,aiming at the problem that the fixed anchor frame in the Faster R-CNN cannot fully frame the blades of different sizes,it is proposed to use the K-means++ clustering algorithm to re-select the blade anchor frames to adapt to the different sizes in the samples.Finally,aiming at the problem that the proportion of healthy leaf area in the sample is small,resulting in poor detection effect of such blades,it is proposed to use the Efficient Channel Attention module to interact with the blade characteristics across channels.Experimental results show that the m AP of extracted leaves of the apple leaf detection model proposed in this thesis is 90.9%,of which the AP of diseased leaves is 95.6%,and the AP of healthy leaves is 86.1%,which is 7.1% higher than that of the original Faster R-CNN model.(2)Research on disease identification method of apple leaves.This thesis uses an improved Mobile Net V3-s network for disease identification of leaf images extracted from Phase I detection.In the apple leaf disease identification model,a lightweight multi-channel feature fusion module with integrated residual structure,deepwise separable convolution,and channel shuffle are first innovatively proposed,which improves the feature extraction ability of the model for diseases.Then,aiming at the identification difficulties of apple leaf disease area with small disease area and high interclass similarity,it is proposed to use the Shuffle Attention module to enhance the model’s attention to the characteristics of the disease spots.Finally,the combined Center Loss+Softmax Loss function is used to improve the classification effect of the network on different lesion types.Experimental results show that the accuracy of the proposed apple leaf disease identification model is 98.17%,which is 2.75%higher than before the improvement,and the model has a good identification and classification effect on different diseases of apple leaves.This study has certain guiding significance for experts to guide growers to carry out agricultural prevention and control in a timely manner,and has important engineering value for improving fruit quality,increasing the income of fruit farmers,and ensuring the sustainable and healthy development of the apple industry.