Research On Recognition Method Of Mangrove Disease And Insect Pest Images Based On Deep Neural Networks

Diseases and insect pests are the combination of diseases and insect pests,which seriously affect the stability of mangrove ecological environment.In order to achieve early intervention and symptomatic control,it is necessary to locate and recognize diseases and pests quickly and accurately.With the development of artificial intelligence technology,diseases and insect pest images recognition method based on deep neural networks is widely used in plant protection,agriculture,entomology and environmental science.With the support of major special projects of Guangxi science and Technology Department,this project takes Guangxi Fangchenggang and Beihai mangrove wetland reserve as application scenarios,aiming at the problems of data scarcity,complex background and long-tailed distribution of mangrove diseases and insect pests image data,researches on image dataset construction,data augmentation,object localization and long-tailed distribution data classification of mangrove diseases and insect pests are carried out.This project has important academic value,ecological significance and great engineering significance.The specific research contents of this subject are as follows:(1)In order to solve the problem that the image data of mangrove diseases and insect pests are scarce and the related open source datasets are few,this thesis collects and constructs a mangrove diseases and insect pests image dataset named MIPDGC(Mangrove Insect and Pests Dataset in Guangxi of China).The dataset includes more than80000 image data of 120 kinds of common pests and plant disease types in Guangxi Mangrove area.The dataset has a variety of different pest development patterns,lesion types,host vegetation and natural background,and has the characteristics of extremely imbalanced distribution of natural long tail type.In order to verify the validity of the constructed dataset,the baseline tasks experiments are carried out on MIPDGC.The experiment results show that the dataset can be used to realize the detection and recognition of mangrove diseases and insect pests image in Guangxi,as well as the research of fine-grained vision and imbalanced classification.In addition,in order to improve the data diversity and model generalization ability,and reduce the data imbalance,this thesis constructs a data enhancement method based on Pest-Rest Pool,which is used to enhance the background diversity and foreground diversity of pest images.The experiment results show that the data augmentaion method proposed in this paper can effectively improve the accuracy of mangrove pest image recognition.(2)In view of the difficulty of localization caused by the complex background of mangrove pest image,this thesis proposes a method of mangrove pest saliency detection based on multi-scale U-Net.Firstly,in order to reduce the phenomenon of multi-scale object missing detection in mangrove pest image,this method introduces a multi-scale context aggregation module based on dilated convolution,which makes the receptive field wider.Secondly,in order to improve the effect of mangrove pest edge extraction,this method introduces multi-scale pyramid pooling attention module,which makes the feature scale diversified and selectively fuses different levels of features.The experimental results show that the method can extract multi-scale pest objects well and has clear edges.Aiming at the high cost of data annotation in the previous method,this thesis proposes another method of mangrove pest weakly-supervised object localization based on local max pooling.Firstly,the method only uses low-cost image level data labels,and uses weakly-supervised learning method to locate the obejcts.Secondly,in order to alleviate the phenomenon of multi-object and small pests missing detection,this method uses local max pooling operation to optimize the weights initialization process.The experiment results show that this method can better locate the pest objects location only using image level supervision.(3)Aiming at the imbalanced long-tailed distribution problem of mangrove pest image dataset,this thesis proposes a dynamic feature weighting method for mangrove pests image classification with heavy-tailed distributions.Firstly,in order to filter out the interference of natural background and accelerate the convergence of network,the method uses the weakly-supervised object localization method in(2)as the preprocessing module to locate mangrove pests.Secondly,in order to overcome the long tail effect of MIPDGC dataset,the range loss is used to train the feature extraction network,which makes the intra-class distance decrease and the inter-class distance increase in the feature space.Then,in order to make up for the over fitting phenomenon caused by the lack of few-shot class feature space,this study uses the mean method to extract and store the class centroid feature vector.Finally,in order to enhance the memory between the head and tail classes,this study uses the extraction,fusion and augmentation strategy of the dynamic feature weighting module to enrich the feature space of the fewshot classes at the tail.The experiment results show that the method improves the accuracy of classification of mangrove pests and diseases images,and improves the ability of classification of rare mangrove pests and diseases.