The Identification Of Rice Lieht-trap Pests Based On Images

It is a common and important pests forecasting method that the rice pests are trapped and killed by lights in paddy fields, then gotten back the next day, and identified and counted by plantprotection technicians in China. This manual method is time-consuming and fatiguing, and it hasthe low accuracy rate. In fact, most of the light-trap insects are unnecessary to be monitored.Those non-forecasted light-trap insects must be excluded. So the technicians need to have a solidknowledge for classifying many species of agricultural insects. In order to reduce the burden oftechnicians, we developed an automatic method for identifying rice light-trap pests based on thepest images. The main results are as follows:(1) Image preprocessing of rice light-trap insect. Firstly, removing the background andperforming the morphological processing. Then, labeling the rice light-trap insects for eachconnected region. Finally, the rice light-trap inests are divided into large rice light-trapinsects and small rice light-trap insects by the area of each insect and the ratio of length andwidth of the minimum exterior rectangle.(2) The automatic identification of large rice light-trap forecasted pests. Firstly, we divided thelarge insects into three groups by their morphological features. Each group has one species ofthe forecasted pests and its similar-size non-forecasted pests. Then, thirty-one color, shapeand texture features are extracted from each insect image. Finally, three support vectormachine classifiers base on cross validation and parameter optimization method are used totrain and test the three groups of insects respectively. The back-side and the abdomen-side ofa pest are seen as the same species. We achieved a93.9%accuracy rate in the three species ofrice light-trap forecasted pests.(3) The automatic identification of small rice light-trap forecasting pests. Firstly, extracting HOGfeatures, LBP features, Global features of all rice light-trap insects. Then, combining thelocal features and global features into a feature vector. Finally, one support vector machineclassifier base on cross validation and parameter optimization method is used to train and testall small rice light-trap insects. We achieved90.2%accuracy rate in the small rice light-trapforecasted pests.Because most of the rice light-trap insects are non-forecasted pests, all the rice light-trapinsects are invided into two classes, large rice light-trap insects and small rice light-trap insects.many classifiers are trained for automatically identifying and counting rice light-trap forecastedpests, and rejecting a large number of non-forecasting pests.