Research On Apple Leaf Disease Detection Method Based On Deep Learning In Complex Environment

Apple is an important economic crop of our country,its output in the world occupied more than 50%of the share.The backward disease prevention and control level has become the main bottleneck of the development of apple in China.Disease prevention and control is an urgent problem for the fruit farming industry.Effective disease detection and control can improve the quality and yield of apples.The background of naturally growing apples is more complex,which will affect the accuracy and speed of apple leaf spot recognition.Disease detection models in complex scenarios often need more modules to adapt to the extraction of different features,and the existing detection models still cannot meet the requirements of real-time detection.To solve the above problems,this thesis takes five types of apple leaf disease as the research object,and proposes a deep learning-based apple leaf disease detection method for complex environment.The main research contributions are as follows:（1）A new Incremental RPN（Region Proposal Network）based on hierarchical detection is proposed to detect apple leaf disease in a complex environment.Firstly,a low-level feature aggregation module is designed to make better use of the features generated by RPN in the first stage.Secondly,multistage ROI Align and attention module are added to the progressive module,so that the converged features are more focused on the difficult to detect spots.Finally,a positional anchor box generator is designed to match the asymptotic module,so that the features and the generated anchors match each other.According to the data set FALD＿CED,the average recall rate of various diseases reached 49.0%in₁₀₀,which was 11.4%,1.4%,2.3%higher than traditional RPN,GA RPN and Cascade RPN,respectively.The experimental results show that the proposed Inc-RPN has better performance than other detection methods in terms of disease recall rate,and the detection accuracy has been improved to some extent.This method provides a new idea for apple leaf disease detection in complex environment and can be better applied to the intelligent management of orchard in natural environment.（2）Light Apple Disease Detector（LADD）is designed.Although the hierarchical disease detection model in complex environment has good detection performance,the structure of the hierarchical model is relatively complex and the number of parameters is large,so it is difficult to achieve real-time or quasi-real-time detection when deployed to terminal for disease image reasoning.Therefore,based on the reciprocal residual module,a double-branch reciprocal residual module is designed in this thesis to lightweight the backbone network.Meanwhile,layer pruning is used to improve the lightweight of FPN and RPN modules.Then,a multi-point distillation strategy was designed,and Incremental RPN was used as the teacher network for channel dimension distillation of LADD.The experimental results show that the floating point operands of the LADD model with the number of channels set at 128 and 96 are only 6.13%and 4.88%of Incremental RPN.Through multi-point knowledge distillation,the recall rate and accuracy of LADD detection are also greatly improved.In the case of single GPU reasoning,when the input image dimension is[3,1333 3,800],the detection speed can reach 23.5FPS.（3）In order to make the detection model more convenient to be applied in the agricul-tural field,this thesis designs and implements a detection APP based on the Android platform.Based on the analysis of APP requirements,the Android environment is configured in three aspects,the lightweight model obtained by training is integrated into the detection APP,and the Android APP with model selection and detection interface is developed.Finally,the APP was deployed to the Android mobile platform,and multi-dimensional testing was carried out on the detection APP.The detection model could reach 8FPS only by CPU reasoning.The detection APP basically meets the requirements of agricultural production detection,and can provide reference for intelligent management of apple planting industry.