Research On The Key Technology Of Longan Flower And Fruit Information Detection Based On Machine Vision

China’s longan industry has lagged behind in automation development,insufficient infrastructure,excessive dependence on low-end labor,and lack of a mechanized production scale system.Many shortcomings in the production of longan have led to a decline in revenue,and market profits have been increasingly squeezed.Since its inception and birth,machine vision has rapidly developed into a hot research field.With the development and maturity of machine vision technology,more and more researches have applied it to agriculture.But so far,there are many gaps in the research of machine vision related to longan.The integration of machine vision into the longan planting industry can improve the production environment of longan and increase production efficiency.It not only improves the orchard informatization level,but also promotes the development of precision agriculture.Using machine vision to obtain information on the density of longan flowers is conducive to judging their growth stage,mastering the best timing for thinning flowers,fertilizing and spraying,and providing a basis for longan yield prediction;While obtaining the location information of the fruit will promote the longan picking robot’s development,further reduce the demand for labor and improve the level of orchard machinery automation.Therefore,this paper will focus on how to use machine vision technology to realize the detection of longan flower and fruit information,estimate the density of flowers and detect and locate the spatial position of fruits.The main research contents are as follows:(1)Two flower density estimation methods based on color threshold segmentation and Deeplab V3 + semantic segmentation model are proposed,and a method for quickly establishing labels is designed for model training in combination with the results of color threshold segmentation.Qualitatively and quantitatively analyzed the effect of the two methods on the detection of flower density,and compared the goodness of fit between the density obtained by the two methods and the artificially calculated density.The experimental results show that the density estimation method based on the Deeplab V3 + model has a goodness of fit of 0.9589,which can reflect the real flower density changes.(2)Designed a longan visual inspection experiment.First,an image acquisition system was constructed using Kinect V2.Color and depth images of longan were obtained under different light intensity and different fruit quantity conditions.The pre-processed images were tested.The experimental results show that YOLO V3 can detect Longan more effectively than Faster-RCNN,with an average detection time of 0.024 s,a model accuracy of 96.55%,a recall rate of 92.02%,F1 The score is 0.942,and it can resist the change of fruit quantity and light intensity to a certain extent.(3)In addition,by designing the visual space positioning experiment of longan,this paper performed Kinect V2 calibration and the calculation of the three-dimensional coordinates of the longan fruit,and constructed the median error curve of x,y and z coordinates.The calculation results of the median errors in the three coordinate directions are 5mm,6mm and10 mm,the median absolute deviations are 3.7mm,3.3mm and 5.1mm,respectively,and the average positioning time of the image is 0.071 s.Experimental results show that the method is feasible and can provide technical support for longan picking robots.This paper completes the key technology research of longan flower and fruit information detection based on machine vision.This work can help the automation of longan orchard management,provide a technical basis for intelligent flower thinning robots and fruit picking robots,reduce labor costs during longan production,and improve Longan production efficiency,thereby enhancing the benefits of planting.