Automated Recognition And Classification For Rice Planthopper Images Based On Deep Learning

The survey of the population of planthopper in paddy field is designed to forecast the occurrence of rice planthopper and better implement integrated pest management.The plant-flapping method is the most used.It require agricultural technicians to flap the plate in the field,identify the planthopper and count up the numbers,with large labor intensity and low efficiency.In order to reduce the labor intensity and improve efficiency of the survey of the rice planthopper,reference 1-2 research on the traditional pattern recognition algorithm of planthopper based on multi-feature and classifier,the recognition effect is pretty good.On the basis,this thesis proposed a four-layer automated detection and recognition algorithm based on deep learning.With establishing the convolution neural network structure which is applicable to the recognition of rice planthopper.The thesis researched on neural network learning and optimization algorithm to improve the recognition rate of rice planthoppers,used bootstrap aggregation parallel strategy to reduce the false detection rate and study multi-classification problems of different planthoppers.The main contributions of this thesis include:(1)In the first layer detection,Extracted HOG features of planthopper and non-planthopper training sample,these features were used to train different cascade Adaboost classifier which influence the detection effect.And the first layer detection algorithm was established based on sliding window technology.Selecting the appropriate Adaboost cascade classifier to test planthopper in rice base image,obtained 89.6% detection rate a nd 81.4% false detection rate.(2)In the second layer detection,the convolution neural network was built on the normalized size of the target.The thesis researched on the effects of different neural layer,weight initialization,regularization and training algorithm on classification accuracy,combined the curve and feature visualization to adjust the network,the optimal network has 98.9% recognition rate for the planthopper test data set.By testing the detection target from the first layer on test images,the second layer comprehensive obtained 89.0% detection rate and 32.9% false detection rate.(3)In the third layer detection,the thesis continued to remove some noises with output four class: long-winged adults,high-instar nymphs,low-instar nymphs and non-planthopper;the thesis researched the effects of training hyper-parameter batch and learning rate on network learning and obtained 95.0% recognition rate for test data set.By testing the planthopper target from the second layer on test images,the third layer comprehensive obtained 81.4% detection rate and 20.6% false detection rate.(4)In the fourth layer detection,in order to eliminate the false detection targets from non-planthopper and the wrong identified planthopper from long-winged adults,high-instar nymphs and low-instar nymphs,the thesis selected false detection targets by testing more than 20,000 rice planthopper images as training data set.And respectively classificated with three kinds of planthopper to make the classifier more robust to non-planthopper.By testing the third layer targets and obtaining 98.4%,100.0% and 100.0% generalization performance.The comprehensive detection rate of the four-layer automatic detection and recognition algorithm based on deep learning was 81.1%,and the false detection rate was 15.4%.Compared to pattern recognition methods,the four-layer detection and recognition algorithm based on deep learning has higher detection rate and robustness,and shows better generalization ability.