Research On The Classification Of Rice Planthoppers In Paddy Based On Image Processing

Rice planthopper is a major migratory rice pests that brings great economic losses for rice production annually. It is important to predict the density of rice planthoppers timely and accurately. Traditional census methods often caused heavy workload, low efficiency and poor precision. We used image processing technology to count rice planthoppers automatically not only save labor, but also improve the efficiency and the accuracy of insect forecast. Yao et al, qing Jie liu, hu zhao and Cheng Chen used image processing technology to extract different planthoppers image features in order to count rice planthoppers automatically, and achieved better result. For further analysis showed that rice planthoppers detection rate and false detection rate still could be improved. We also need to classify planthoppers species, long and short-winged type and instar. This article continued to study the automatically counting of rice planthoppers based on image processing, and then studied the classification of planthoppers species, long-winged and short-winged planthoppers and instar.We developed an Adaboost classifier based on HOG features to count rice planthoppers and optimized the classifier from Adaboost preferences, size of training samples, number of samples and samples diversity angle. The results showed that the Adaboost classifier with 21 layers trained by 8000 positive samples with 32 * 64 pixels and 32,000 negative samples which all from different rice background images, and the maximum false alarm rate of 0.48 obtained better detection result. We used rich planthoppers which can be recognized by the human eye as evaluation criteria and got detection rate was 92.3% and false rate was 53.5%. This article used convolution neural network to classify rice and non-rice planthoppers. It showed that convolution neural network could reach a good result and rice planthoppers recognition rate was 99.1% and false rate was 8.7%.We combined the Adaboost classifier with convolution neural network and obtained the detection rate was 91.5% and the false detection rate was 4.7%.We used Gabor features and SVM classifier to classify long-winged and short-winged planthoppers and planthoppers nymphs, and used HOG features and SVM classifier to classify long-winged white and brown rice planthoppers.The result showed that the classification accuracy rate of planthoppers species, long-winged and short-winged, elderly and young nymphs not less than 95.0%.The result of this artical provided a theoretical basis for paddy planthoppers forecast by detecting, classifying and counting rice planthoppers based on image processing automatically.