Research On Cotton Main Stem Growth Point Recognition Based On Machine Vision

Cotton planting industry is one of the important economic industries in China,which occupies a high economic status.When the cotton grows to a certain period,it needs topping operation,that is,cotton main stem growth point removal.At present,there are three kinds of topping methods used in China,namely,artificial topping,mechanical topping and chemical topping.People have gradually explored and experimented in the way of mechanical topping and chemical topping to achieve topping work in a more effective and accurate way.However,due to some problems in the operation and control of the cutter,the topping machine technology has not been fully implemented.With the development of the industry towards mechanization and intelligence,machine vision technology has developed rapidly and has been applied in many fields.In this paper,the main research contents are as follows:A data set of growth points of cotton main stem was created.All the experimental data sets are self-collected data.Cotton images were collected in different periods and time periods,and the images were sorted out uniformly to increase the diversity of the data set.The collected images included different lighting conditions,different shooting angles and different shooting heights.Each image was marked with Labelimg software,and the label information was preserved,The data set contains 11,001 images and corresponding annotated information files.The basic target detection network model suitable for detecting the growth point of cotton main stem was screened out.Through the comparison and theoretical analysis of several mainstream target detection models,the YOLOv3 network with the best detection effect is selected and further improved on this basis.These target detection models include Faster RCNN,RetinaNet,CenterNet,YOLOv3 and YOLOv4.The Faster RCNN algorithm is a two-stage target detection algorithm,and the rest are one-stage target detection algorithms.The identification network of cotton main stem growth point was constructed,and the YOLOv3 network was improved.The difference between YOLOv3 network and improved network was analyzed,and different detection results were compared.The improvement of the model mainly includes three aspects:first,in order to realize the transmission of shallow semantic information to deep semantic information,dense connection blocks are added,and the residual unit are reduced in the original network;second,the dense connection block is improved,and the nonlinear operation in the block is replaced by deep separable convolution,simplify the network model and reduce the number of parameters;third,the residual unit structure is improved,the original residual unit is replaced by hierarchical multi-scale convolution module,realize multi-scale receptive field fusion and enhance multi-granularity characteristic learning.Finally,the Map of improved network model is 90.93%,which is 1.64 percentage points higher than the original network model,and the training parameters of the model are reduced by 48.90%.The model of cotton main stem growth point recognition and location was established.The target spatial positioning was realized by depth camera and three-dimensional coordinates of the target were obtained.The imaging principle and the relationship between the coordinate systems in space were analyzed.The three-dimensional coordinate of detection target is analyzed and the results show that the maximum errors of X,Y,Z are within 1cm,the average errors are 3.59mm,3.52mm and 2.79mm respectively.The results of this paper can provide technical support for precise cotton topping operation based on machine vision.