Study On Rice Pests Counting And Identification Based On Image

The rice is one of the most important agricultural crops from ancient in our country,and ocuppies 25% acreage in our country.But rice pests’ frequent occurrence has brought huge economic losses to the agriculture,and increasing the yield and quality of rice becomes a important goal of the agriculture in our country recently.Thus,the rice pests’ prevention is significant in the development of agricultural economy in our country.But in the previous work of rice pests prevention,due to the inadequate monitoring of the insect situation,people can not find the changes of the pest populations in time,and losing the perfect time of pest management.Accordingly,the monitoring of rice pest situation is the key of pests control.In the monitoring of rice insects situation of our country,we usually catch the insects by an trap lamp,bring them back in the next day,then complete the counting and identificaiton of the pests by people,and finally judge the situation of the rice pest.But there are lots of troubles about this method,like low efficiency,labour-intensive and subjective,and in addition,it can not feedback the information real-timely.For the agriculture is developing toward modernization and automation gradually,manual monitoring of the pest situation can not meet the modern agriculture.New method of rice pests counting and identification need to be find out to replace the manual monitoring urgently.For image recognition skill’s application in many fields,we researched rice pests counting and identification with image processing and neural network.And a system of rice pests counting and identifying was established by GUI toolbox of Matlab.The main contents of the research in this paper are as follows:1)Establishing the pests trap and image acquisition equipments.A pests trapping lamp which contains the function of taking pictures was designed according to actual rice field environment.The lamp consists of image acquisition module,FPGA control module and 4G wireless transmission module.The image acquisition processes: The lamp lures pests into the box by a special light,and kills them by infrared device,the dead pests fall on a baffle and a camera snaps the image of the pests on the baffle,the image will be sended to cloud server.This lamp provides an image acquisition system for the study of pests counting and identification.2)Image preprocessing.This paper researched the gray threshold segmentation,color conversion and adhesion image separation.The sample image was collected from the actual rice fields.There are lots of lines in the background after analysing the sample image.And this paper put forward a method which compromises image gridding segmentation and OTSU to segement the background image.Compared with traditional methods,the results of this method show that the background image is segmented perfectly and has a very good anti-noise and robustness ability.3)Segmentation of adhesive image.For the adhesive images,we proposed a marker-controlled watershed segmentation method to divide the pests areas.Compared with traditional watershed segmentation algorithms,this method can avoid over-segmentation or under-segmentation.Combinied the background segmentation and adhesive pest division,the pests counting average accuracy of the method reaches to 91.8%.4)Classification and detection of rice pests based on convolutional neural network model.Compared with traditional recognition pattern,CNN has the advantages of strong robustness,outstanding classification ability and strong adaptability.Thus,according to the characteristics of the sample images,analysised some mainstream CNN models like Alex Net,VGGNet,Le Net-5 etc and the influence of each module of CNN on recognizing accuracy,we constructed a 12-layer convolution neural network model,taked the ELU function as excitation function,led Dropout into the model,and adjusted other functional layers accordingly.We trained and tested the model with 1200 images,and then adjusted parameters of the function layer constantly.And we completed the classification and recognition of the rice pests such as mole cricket,planthoppers and yellow rice bore.The result shows that identification of rice pests by CNN is a very effective method,the accuracy rate reaches to 95.8%.5)A system of rice pests counting and recognition was established.Based on image preprocessing,and pest counting and pest recognition by CNN model,we established a system of rice pests counting and identifying by GUI toolbox of Matlab,and this system mainly contains: Image viewing,image preprocessing,counting,model training and classification.Users can obtain pests’ number and species of the sample image.