Research On Acquisition Method Of Maize Number In Different Growth Stages Based On Improve YOLO

Plant number as an important information of maize growth stages,is the basis for obtaining other agronomic traits.Using plant number information can assist maize management to achieve maize yield increase.In order to quickly and accurately obtain the number of corn plants in different periods,this paper used unmanned aerial vehicle to collect the visible orthophoto image of corn in three periods,constructed the number detection model of corn plants based on YOLO algorithm,designed and realized the automatic acquisition system of corn plant number.In order to establish an efficient and accurate maize monitoring method and plant number acquisition method.The main research contents and results of this paper are as follows:(1)Data preprocessing in three periods of maize and production of plant number detection datasets.Aiming at the problems existing in the datasets,the of unmanned aerial vehicle(UAV)use to collect corn visible orthophoto,after the collection of image data from a corn plant were analyzed,according to the characteristics of corn plants under visible light orthophoto,the main features suitable for learning the target detection model were determined and the data were annotated.datasets of 900 images of corn plants in three periods containing about 30,000 plants was marked,which was sorted into VOC format for subsequent analysis and research.(2)Construction and optimization of maize plant detection model based on lightweight convolutional neural network.Firstly,the characteristics of the maize target bounding box of the datasets were analyzed theoretically,the target area histogram,spatial distribution and width to height ratio diagram were obtained,and the optional improved method was selected according to these data characteristics.Then,a plant number detection model based on the lightweight feature pyramid was designed,different detection models were constructed by using the alternative improved modules for ablation experiments.The optimal maize plant detection model structure YOLO-FE2 was selected.And final model FE-YOLO was obtained by further optimizing.According to the experimental test,the AP value of this model at different stages is: 86.40% in seedling stage,79.06% in growth stage,80.66 in maturity stage,and 82.04% in MAP of three stages.The total parameters of the model was only 17.31% of that of YOLOv3,the calculation amount of detection process was 10.4 GFLOPS,and the average detection time of single image was 7.8 milliseconds.Compared with YOLOv3 model,the plant recognition accuracy of the model was improved by 6.79% in the robustness test.The experimental results show that the model has high accuracy and good robustness,and basically meets the requirements of maize plant detection and plant number acquisition.(3)Design and evaluation of automatic maize plant number acquisition system based on online access.To better use this article builds the model of applied in the actual detection,according to the current number of the problems existing in the acquisition system,was designed and built a number of automated test system based on online access to the system through We Chat small program to upload visible light projective like corn,is transmitted to the remote server after cutting,finally to identify and cut the image return to the test results.The orthophoto images taken at three stages were randomly selected to test the system.The recognition rates of the three stages were as follows: 92.76% in seedling stage,88.36% in growth stage,86.28% in maturity stage,and 89.42% in average accuracy.The experimental results showed that the accuracy of the system in detecting plants basically met the practical requirements.The corn plant number acquisition process based on the constructed FE-YOLO plant detection model and the online access plant number acquisition system has certain application value because the process does not contact with corn directly,does not need to customize additional equipment and consumes less manpower and material resources,and can quickly and efficiently obtain the corn plant information and planting density.It can provide certain support for the management and testing of corn.