Research On Automatic Tiny Pest Recognition And Counting

Pest infestation is one of the major factors causing crop yield reduction.In our country,there are many pest species that might occur in the whole growth cycle of field crops.Thus,how to effectively control field pests to improve agricultural production and reduce blind use of pesticides is becoming an increasing important problem.At present,pest monitoring in our country mainly relies on the in-field manual survey of by agricultural experts,which obviously leads to limitations such as time-consuming,laborious,subjective and error-prone.Currently,deep learning and computer vision based techniques aiming at addressing these issues might not achieve satisfied performance with low accuracy and poor robustness.Therefore,till now,it is highly demanded to develop novel approaches in agricultural pest monitoring applications.In this case,this dissertation focuses on systematic research on pest monitoring tasks in the wild field,which contains four major aspects to deal with:(1)tiny pest data augmentation;(2)pest recognition and counting;(3)pest occurrence severity estimation.This thesis was supported in part by STS project "Smart Agriculture Core Technology Breakthrough and Integration" from Institute of Intelligent Machines,Hefei Institutes of Physical Science,Chinese Academy of Sciences(Grant No.KFJ-STS-ZDTP-057),and supported in part by "Test Plan of Intelligent Collection Equipment for Crop Diseases and Pests" from National Agri-Tech Extension and Service Center.The works in this dissertation can be summarized into follows:1.To address the problem of difficulty in collecting tiny pest images,we propose an effective data augmentation strategy to solve the problem of feature representation when the tiny pests are rotated,shifted and scaled under the limited condition of image training data.In this method,we first extract the visual features of annotated pest location regions and automatically generate a pest template dataset that cleans the noise background from pest objects.Then,the pest templates are randomly sampled and fused into their potential occurrence locations of real-captured field crop images to improve the rationality of composited image.Experiments show that with our data augmentation strategy,current pest detection approaches could obtain a remarkable improvement on detection accuracy and robustness.2.To solve the problem of low recognition accuracy and difficulty in extracting typical features of tiny pests in the field,we propose a novel Coarse-to-Fine Network for in-field pest detection task.This method proposes a new framework for pest detection which is the first one for this task in the world.In this method,we divide the total pest detection task into two individual phases.In the first stage,we design a Coarse Network to localize the pest occurrence regions that contain pest cliques in image.Secondly,a Fine Network is developed to finely locate each pest instance separately.The final result is integrated with detection results from Coarse Network as well as Fine Network.Experiments prove the significant improvement on robustness and accuracy of pest detection task.3.To reduce the high missed detection rate and improve the detection accuracy of tiny pests with high density distribution,we propose a novel Detect-Density Network for pest counting task.In this work,we design a simple pest density classifier to divide pest densities into low-density and high-density groups,and adopt different ways to process the two groups.For low-density images,we use a popular pest detection apporach to find the locations and recognize species of pests.In terms of high-density images,a novel multi-scale density map estimation method is designed to count the pest population by predicting Gaussian density map.Our Detect-Density Network effectively exploit pest detection and pest density map estimation to further boost pest monitoring performance,especially on high-aggerated tiny pest detection and counting task.4.Introduce a novel pathway for transferring deep learning models into practical pest monitoring applications.To this end,this work aims to design an expert-level pest severity estimation and prediction model.Specifically,we develop an industrial standard for image capturing and manual pest severity assessment.To correctly estimate pest occurrence severity of captured pest image,we adopt Detect-Density Network to extract high-level vectors containing pest species,localizations and population information as input to fit our expert-level severity estimation model.Extensive experiments demonstrate that the predicted pest severity is highly consistent with manual severity judgement.At present,this technology has been promoted and applied in China.