Research On Two-stage Network Models For Detection And Classification Of Forestry Pests

Forestry pests are one of the main obstacles to the development of forestry.If forestry pests are not prevented and controlled in time,they cause certain damage to trees,which will seriously affect the growth of trees,destroy the ecological environment,and bring losses to the national economy.Therefore,the detection and classification of forestry pests is an important step during forestry pest management.The goal of pest management is to detect pests as early as possible and treat them in time,so as to greatly reduce the losses caused by pests.For the problem of pest detection and classification,researchers have collected some pest image datasets and proposed effective single-stage classification methods.However,there is currently no image dataset specifically for forestry pests,and the recognition accuracy of existing single-stage methods often decreases significantly as the pest images become more complex and the number of pest categories increases.To address the existing problems of current methods,this thesis proposes a two-stage network framework based on the biological hierarchical structure for detection and classification of forestry pests.Under this framework,this thesis improves the classification optimization of the second stage by introducing the edge features of pests,and designs a forestry pest detection system supported by multi-platform.The main work of this paper is as follows:(1)To our knowledge,no a publicly pest image dataset is specially for forestry detection and classification.Therefore,this thesis collects and organizes a new forestry pest image dataset according to the distribution of forestry pests in Jiangsu Province of China.This dataset contains 7253 high-quality images of 60 common and easily obtainable forestry pests from 15 biological families.We annotate pests in this dataset with both biological families and species,so that this dataset can be applied to related research work,say detection and classification of forestry pests.(2)To improve the performance of single-stage classification methods,this thesis proposes a two-stage framework based on biological hierarchy(BH-TSF)for forestry pest detection and classification.Generally,pests belonging to different biological families may have greatly different features,and pests in the same biological family but different species have similar features.Thus,it is hard for the existing single-stage models to learn both the global feature of different biological families and the local characteristics of different species under the same biological family at the same time.In BH-TSF,the first stage uses an object detection model to locate pests in images and judges their biological families,and the second stage uses multiple classifiers,each of which is for distinguishing specific species under the same biological family,to train extracted batches with pests.To implement BH-TSF,this thesis uses a YOLOv5(You Only Look Once version 5)model in the first stage and multiple DenseNet(Dense Convolutional Network)models in the second stage.Experimental results on our collected dataset show that the average accuracy of our model is up to 95.52%that is higher than that of the current mainstream algorithms.In addition,BH-TSF is relatively less affected by the increased number of pest categories.(3)In view of the problem that the classification performance in the second stage is unsatisfactory for some biological families,this thesis constructs new features for the classification in the second stage.The classification objects in the second stage are species in the same biological family that have similar features each other.To further explore the specific information in pest images within the same biological category,this thesis designs a forestry pest edge features extraction blocks and integrates it into the Swin Transformer(Shifted Windows Transformer)algorithm,proposing the PEST(Pest Edge Swin Transformer)classification algorithm.Under BH-TSF,multiple PEST classification models are used in the second stage to classify different species within different biological categories.Through experiments,it is demonstrated that replacing the second stage classifier with PEST is reasonable and feasible because PEST outperforms DenseNet and Swin Transformer in classification performance.On our collected dataset,the new two-stage model has better classification performance that the average accuracy is up to 96.83%.(4)This thesis designs and implements a multi-platform forestry pest detection and classification system based on BH-TSF.The main functions of this system include pest image detection,pest information encyclopedia,and pest warning and analysis.This system is practical significance because it can provide industry analysis,regional analysis,trend prediction and other analysis report data based on plant protection big data for forestry experts and regulatory departments,contributing to the development of the national forestry economy.