Research On Recognition And Localization Method Of Crops In Seedling Stage For Weeding Between Plants

With the rapid development of science and technology,the applications of visual technology in agricultural operations are more and more extensive.The ability of weeding machinery can be significantly improved by using visual technology to identify crops and weeds.At present,the weeding machinery widely used can only remove the weeds between rows,which are powerless to the weeds between plants.Technologies of weeding between plants based on vision can clear weeds around crops and effectively solve the problem of weeding between rows and plants.Aiming at the segmentation of crops and weeds in seedling stage,this paper takes corn seedlings in the field as the research object.The crop images covered two kinds of light environment,sunny and cloudy days,and there are more,less and few weeds.Three methods are used:The general contour skeleton feature method、SVM classifier based on SIFT feature and Harris corner feature and the deep learning method based on Faster-R-CNN model.In order to find out the methods with high recognition accuracy and good real-time performance to recognition of maize seedling in image.The main work of this paper is as follows:(1)The method of crop recognition and location based on skeleton feature is studied.The accuracy of this method is 85.9% when the effective distance is less than 10 mm.The average processing time of the image is about 516 ms.(2)The SVM trainer which combines SIFT feature and Harris corner feature is studied,and the training model can be classified.The trained model is used to recognize and classify the image.In the contrast experiment,it is found that the SVM classifier with RBF kernel function is the best,and the recognition accuracy of corn seedling is 88.7%.The average processing time of each image is about 1524 ms.(3)In order to solve the problem of limited feature selection and poor real-time recognition,a maize seedling recognition method based on Faster R-CNN model was studied.Firstly,the image of corn seedling is collected and the data set is established,then the sample image is labeled,and the recognition efficiency and accuracy are improved by improving the RPN network candidate frame method.The recognition accuracy of maize seedling was 96.1%.The average processing time of each image is about 1242 ms.The experimental results show that although the image recognition processing speed based on skeleton extraction method is fast,the recognition accuracy is not high,and there are some problems in the complex situation of noisy and leaves block;SVM classifier training can find better classification effect,classifier parameters and kernel functions through experiment comparison.The classification accuracy is slightly improved,but the image processing speed is slow;Based on the recognition of deep learning method,the model and parameters of Faster R-CNN network are adjusted and improved through category tagging,and the recognition efficiency of corn seedling is very high with the data set obtained through training,and the recognition speed has been greatly improved compared with SVM.