Font Size: a A A

Research On Detection Algorithm Of Rice Field Planthopper Based On Image Local Features

Posted on:2015-01-23Degree:MasterType:Thesis
Country:ChinaCandidate:H ZhaoFull Text:PDF
GTID:2253330428463201Subject:Signal and Information Processing
Abstract/Summary:PDF Full Text Request
Rice planthopper is an important class of migratory pests, generally infestation isconcentrated in the lower part of rice. Mastering planthoppers’ field population density dynamicchange is critical for accurate forecast and planthopper’s reasonable prevention. Currently,planthoppers’ field survey methods of forecasting has mainly taken the visual method, sweep netmethod and the disc shot method. These methods have labor-intensive, time-consuming,poorcounting accuracy and long-standing problem, in order to reduce the labor intensity of plantprotection officers survey the field planthoppers, improve efficiency and accuracy, Liu Qing-jieet al [1] use image processing techniques to detect rice planthopper and counting. On the basis ofLiu Qing-jie, etc., this thesis continues in-depth study of algorithm of rice planthoppers based onimage local features. The main contents and results include:(1)In the study of first layer detection algorithm of rice base’s white-backed planthoppers,this paper studied under different combinations of size, number of samples and the negativeselection of training samples, training AdaBoost classifier based on image haar feature, and testimages with field planthopper. select the optimal classifier as the first layer detector, thedetection rate is85.25%, the false positive detection rate is99.1%.(2)In the study of second layer detection algorithm of rice base’s white-blackedpalanthoppers,to classify sub-images of first layer detection,utilizing four image classificationalgorihms,SPM (Spatial Pyramid Matching), ScSPM (Linear Spatial Matching Using SparseCoding for image classification), LLC (Linear Locality Coding for image classification) andopponentcolorsift fusion HOG SVM.To statistic classification rates and false-positive rates andanalysis and evaluate results.the result shows that opponentcolorsift fusion HOG SVM gets thebest result,on the basis of sub-windows of first detection layer,using pyramiddense samplemethods to sample sub-image and use opponentcolorsift to describe it.then using bag of wordsmethods to extract sub-window bof feature combining hog(Histogram of Gradients) to describethem,finally use SVM to classify sub-images.The detection rates and false-positive rates are67.3%and37.7%,comparied with the first layer detection,the second layer’s detection ratesdecreases18%,but false-positive rates decreases61.4%. To analysis detection rates and false-positive rates,we find that above methods can notsatisfy the field survey’s real-time requirements.However,these methods can offer an referencefor the latter researcher.
Keywords/Search Tags:Haar feature, AdaBoost classifier, bag of feature, SIFT feature, image localfeature, Support Vector Machine
PDF Full Text Request
Related items