Algorithm Research On Rice Planthopper Recognition And Classification Based On Image Processing

China is the largest rice producer and consumer in the world,so stable and superior rice production is closely related to people’s livelihood.Currently,the most serious pest that harms rice in China is rice planthopper.In order to effectively reduce the damage of rice planthopper and solve the problem of pesticide abuse within rice farmers,it is necessary to know the density of rice planthopper in advance and develop related control strategies.Therefore,recognition and classification algorithm of rice planthopper images was presented,which can promote real-time acquisition of rice planthopper density.(1)Acquisition of rice planthopper imagesIn order to obtain complete and clear images of rice pests in different time and space,a wild insect image acquisition device designed by our team was applied.Based on this image acquisition device,the image acquisition experiments were conducted in Lishui District,and Gaochun District of Nanjing,Jiangsu province in 2017,and 2019,respectively,and a sample set of rice pest images for subsequent image processing was built.(2)Recognition and classification of rice planthopper images based on shape descriptorsBased on this rice pest image sample set,a rice planthopper image classification algorithm based on shape descriptor was proposed.The shape features of rice planthopper were described by Fourier descriptors and Hu moments,and the random forest was used as an image classifier.The numbers of decision trees and the prediction variables of the random forest were determined through out-of-bag errors,and finally,based on this random forest classifier,rice planthopper images could be effectively identified and the recognition rate reached 0.939.(3)Recognition and classification of rice planthopper images based on Mask R-CNNIn order to further improve the automation level of rice planthopper image classification,a rice planthopper image classification algorithm based on Mask R-CNN was proposed.The Microsoft COCO dataset was used as a source image data domain by transfer learning to solve the problem of too small sample size in deep learning.Then,the performance of Mask R-CNN was verified in three situations,including only rice planthoppers,only non-rice planthoppers and these two types of insects both in the images.The experimental results showed that Mask R-CNN could effectively recognize and classify rice planthopper images with an average recognition accuracy of 0.975.Besides,based on the rice pest image sample set,when the inputs are adhesions in rice pest images,the performance of Mask R-CNN was discussed;and experimental results showed that when the degree of adhesion of rice pest images is relatively light,rice planthopper images could be effectively recognized,while rice planthopper images could not be precisely recognized,when the degree of adhesion of rice pest images is severe.(4)Study on segmentation algorithm of adhesion images of rice pestIn order to effectively solve the problem that the adhesion images of rice pest cannot be accurately identified,an image segmentation algorithm for adhesive rice pests was presented.A sliding window detection method was applied to search the concave points of the adhesive image area,and the adhesive area segmentation lines were determined by the matchable concave points and isolated concave points to obtain single rice pest images.The Mask RCNN,mentioned above,was used as a rice pest image classifier again,and those single rice pest images were utilized as input data to verify whether segmentation algorithm for adhesive rice pests can be applied in rice planthopper image recognition.The experimental results showed that single rice pest images extracted by the adhesion image of rice pest segmentation algorithm could be effectively identified,and the average recognition accuracy could reach 0.987.