Detection And Location Of Honeydew Melon Canopy Based On Machine Vision

China’s greenhouse area ranks the first in the world,but the overall greenhouse cultivation automation degree is low,many areas of the country still use artificial spray,liquid spraying efficiency is low,while causing waste,workers poisoning incidents also occur.Therefore,it is of great significance to develop spraying robot for greenhouse to realize precise application of pesticide,improve the efficiency of pesticide use,reduce environmental damage and pesticide residue.Hetao honeydew melon has a long history of greenhouse cultivation in Hetao region of Inner Mongolia and is the main variety of thick skin melon in Hetao region of Bamun.In this paper,the weed identification and location technology of greenhouse robot is studied.Since it is difficult to distinguish honeydew melon plants and weeds quantitatively by visual detection algorithm alone,this study intends to adopt the detection method combined with spectrum.The ground object spectral data of hetao honeydew melon plants and common weeds such as Purslane,setaria setaria and Leucocephera were collected,spectral characteristics were analyzed,and the discriminant model was established to realize the identification and differentiation of Hetao honeydew melon plants and common weeds.Binocular vision system was used to locate honeydew melon plants.The specific work content is as follows:(1)The spectral data of honeydew melon,Portulaca oleracea and Setaria setaria were collected by American SVC spectrograph.After screening,the data were preprocessed by S-G convolution smoothing,standard normal change,first derivative and compression,and321 groups of samples were reserved,each group containing 317 data points.Stepstep discriminant analysis method was used to extract four characteristic wavelengths of 529 nm,742nm,868 nm and 1027 nm based on the principle of wilcrams minimum.SPSS software was used to establish a typical discriminant model to predict the prediction set,and the recognition accuracy of the typical discriminant model reached 96.93%.(2)Using the continuous projection method and based on the root mean square error curve,5 feature wavelengths of 421 nm,670nm,910 nm,1321nm and 1348 nm were selected to establish the Bayesian discrimination model,and the prediction set was tested.Under different prior probabilities,the model recognition accuracy was 93.87% and95.92%,respectively.(3)The binocular vision detection system is designed and calibrated.The calibration results show that the system has good accuracy and stability.(4)Binocular camera was used to collect images of hetao honeydew melon in the growing period,and the collected image data was denoised by bilateral filtering;Then the histogram threshold segmentation method is used to separate the target from the background.Then,Canny edge detection was used to detect the target edge feature,and the target feature extraction of honeydew melon canopy was realized.(5)Stereo correction is carried out on the left and right eye images after distortion removal to ensure the corresponding relationship of pixels during matching.Local stereo matching algorithm and global stereo matching algorithm are used for stereo matching,respectively,to achieve the location and detection of target crops.The experimental detection shows that the target error of BM algorithm is 3%-9%,and the target error of SGBM algorithm is 2%-3.5%.The parallax map obtained by stereo matching was combined with the binary map extracted from edge features to calculate the average depth of the target area combined with the number of pixels of the target area,and then the actual canopy area was calculated by the projection area method.