Research Of Rice Pest Image Segmentation And Detection Algorithm

In rice growing process, many species of pests damage rice. In China, two methods are usedto forecast rice pests. One is forecasting lamps for trapping pests and other is paddy pest survey.Pests trapped by forecasting lamps will be brought to the laboratory the next day and plantprotection technicians visually identify and count pests. Paddy survey of rice planthoppers needsone hold a white disc, enter the paddy, flaps the bottom of rice stem and count the planthoppers,other will write down the data. The two man-made forecasting methods can lead to laborintensive, low counting accuracy and time-consuming. In our study, image processing andmachine vision are used to research segmentation of touching insects images captured by ourinsect imaging system for automatic identification of rice pests and automatic counting ofplanthoppers on rice stems. The main research contents and results are described as follows:（1） Firstly, a background image is captured by our insect imaging system. Then, thelight-trap pests are spread on the glass platform by different size mesh screens. The glassplate is slightly tapped by hand. Two pest images are captured before and after tappingthe glass plate. The differential analysis is used to remove the background noises. Theglobal optical flows are calculated by the two images. For improving the accuracy ofoptical flows, the optical flows of background are assigned to zero. The angles of opticalflows are redefined by optical flow vectors. At last, NCuts is used to segment thetouching insects. The weight of NCuts algorithm is defined using the angles of opticalflows. The results showed this method could achieve good segmentations of touchinginsects.（2） In the counting research of white back planthoppers on rice stems, three layers ofdetection mechanism are used to detect the planthopper. Firstly, the first layer ofdetection is the Adaboost algorithm to detect the white back planthopper. The positiveand negative samples are developed and Haar features are extracted. A series of weakclassifiers are trained and constituted into a strong classifier to detect white backplanthoppers. A higher detection rate and a higher error detection rate are gotten byAdaboost. So HOG features and SVM classification are used as the second layer to remove the incorrectly detected noises by Adaboost. At last, double-threshold method isused to remove the background automatically and extract three conventional featuresfrom sub-images. Some images with reflection or drops are removed by comparing withthe thresholds of three features. The results showed that high detection rate and low errordetection rate of white backed planthopper are achieved.This research converts static image segmentation into dynamic image segmentation. Itovercomes the over-segmentation and under-segmentation in traditional segmentation. It exploitsa new direction for segmentation of irregular targets. The automatic detection and count of whiteback planthopper on rice stems are put forward from the view of machine vision. The detectorand classifier obtained by learning samples are used for real-time detection, which avoids thenoise interference in traditional color image segmentation. This method has good robustness.