Researches On The Recognition Methods Of Weeds And Pests In Vegetable Greenhouses

The vegetable industry is an important pillar of Chinese agricultural economic development.How to produce green and healthy vegetables while increasing income is an problem that needs to be solved at present.As a key facility in the vegetable industry,the manual weeding and insecticide methods in vegetable greenhouses are high in cost and intensity;chemical weeding and insecticide methods will improve the harmful components in crops,both of which do not meet the development requirements of modern agriculture.So we need to develop autonomous robots to carry out the weeding and insecticide work in the vegetable greenhouse environment.Intelligent recognition of weeds and pests is the key technology for the development of the robots.In this paper,combined with the needs of the project,two key technologies of weed recognition and pest recognition in vegetable greenhouse environment are studied.The main research contents are as follows:(1)Research on the recognition method of weeds based on deep learning.Three classic network models Faster-RCNN,SSD and YOLOv3 are used for training and testing.The boxes＿cell matrix is used to remove the crop images in the prediction frames,combined with the super-green feature and the area filtering method based on the connected region label to remove the residual noise,and the weed images are effectively retained.The recognition results under the influences of different light conditions,different weed types and different growth stages of crops are compared and analyzed,and the most suitable network model is selected through comprehensive consideration.Experimental research shows that the YOLOv3 network model is most suitable for weed recognition in the vegetable greenhouse environment.The detection time of one single image is 0.032 s and the AP value is 0.95.The recognition rate of the method proposed in this paper is 95.71% and the misjudgment rate is 12.27%.This method can adapt to changes in light within a certain range,which is less affected by weed species and has a good recognition effect for crops in different growth stages.(2)Research on the recognition method of weeds based on three-dimensional information.Combining three methods: through filtering,neighborhood operator filtering and voxel filtering,we remove noise points in point cloud data and improve the efficiency of subsequent algorithms;super green feature and Euclidean clustering segmentation algorithm are used to segment soil background,single crop and single weed.The Z value of the highest point is extracted from the segmented crop and weed point cloud cluster,and the weed recognition is completed according to the height information.The recognition rate of the method based on three-dimensional information proposed in this paper is 86.48%.(3)Research on the pest recognition method based on spectral analysis.In this paper,the snails are taken as the research objects.The maximum likelihood classification,minimum distance classification and spectral angle classification in supervised classification and recognition are used for comparative analysis,and the most suitable method is selected,finally the research is compared based on the full-band and single-band.In the pest recognition research based on spectral analysis,the maximum distance classification has the best recognition effect.The overall accuracy is 97.6492%,the Kappa coefficient is 0.9543,the snail pixel recognition rate is 97.10% and the single-band processing effect after applying the principal component analysis method is better than the full band.The overall accuracy is increased by 0.8340%,the Kappa coefficient is increased by 0.0198 and the snail pixel recognition rate is increased by 1.46%.This paper provides a general and innovative method for weed recognition and pest recognition in vegetable greenhouses.