Research On Phenotypic Parameters Measurement Of Rapeseed Based On Image

Rapeseed(Brassica napus L.)is one of the most important oil crops in China,with high economic value and great development potential.The phenotypic study of rapeseed is the key to the above technical means,but the non-destructive automatic measurement of the phenotypic parameters of rapeseed plants is still an urgent problem to be solved.At present,the acquisition of rapeseed phenotypic parameters mainly stays at the stage of manual measurement.However,this method is featured by many shortcomings,such as excessively relying on subjective judgement of operators,inefficiency and lacking of real-time performance.And it is difficult to meet the demand of high-throughput measurement.With the development of machine vision technology,it has been widely used in agriculture.In this study,the main phenotypic parameters of rapeseed were obtained by image processing technology aiming at the automatic measurement of phenotypic parameters.The main research contents are as follows:(1)Measurement of phenotypic parameters of canopy in rapeseed.Firstly,the measurement of the leaf area of rape was completed through image geometric correction,connected area labeling and other processing.The experiment showed that the relative error of the measurement of the leaf area of rape by this method was less than 3%.After that,aiming at the problem of shape reduction in the case of edge damage in the leaves,this paper proposed the Active Shape Model(ASM)to conduct model training and shape fitting.Some progress had been made.The IoU of the reduced leaf area was 0.923,and the IoU of the damaged area was 0.873.In the bud part of the rape we solved the counting problem.Using the three-dimensional shape characteristics of rape buds,the image enhancement algorithm was carried out in HSI color space to obtain the number of buds,and the average relative error of counting was 4.42%.Based on this,a set of image shooting portable device was designed.(2)Measurement of plant type parameters of rapeseed.This study mainly completed the measurement of plant height,branch and pod angle.After image calibration,the rape height was calculated by the minimum external moment of rapeseed contour in binary image.In addition,the problem of fracture of the connected region of the main stem in the process of image segmentation was solved.The average relative error of plant height measurement was 3.17%.In the process of angle measurement,the bifurcation point and the pod birth point were obtained by the corner point finding algorithm,then according to the feature points the image was divided into two parts of the trunk and branches.After extracting the contour spindle,the liner equation of the spindle was obtained and the angle was calculated as well.The average relative error of the branch angle measurement was 2.64%,and the average relative error of the pod angle measurement was 1.95%.(3)Study on principal root segmentation algorithm in rape seedling stage.Aiming at the difficult problem of identification of main root in rapeseed root analysis,this study based on total convolutional neural network(FCN)to realize the automatic segmentation of main roots in rape.Firstly,aiming at the lack of existing data sets,the establishment and data enhancement of the main root segmentation image data set in rape seedling stage were completed.After that,the Pytorch framework was used to implement the FCN-8s model and conduct training and optimization to obtain the segmentation results of the principal root image.In order to evaluate the segmentation results,the results of this study were compared with those of traditional image segmentation algorithms.The experimental results showed that the segmentation algorithm based on the total convolutional neural network could obtain relatively complete shape and clear contour of the principal roots,with the segmentation pixel precision(PA)reaching 0.987 and the segmentation intersection ratio(IoU)reaching 0.835.(4)Measurement of seed parameters of rapeseed.In view of the disadvantages of existing particle counters,such as low efficiency and high noise,this study designed a thousand-grain weight measurement system based on the Android platform,and designed the corresponding hardware system and Android application software.The weighing link used serial communication technology to connect electronic balance with intelligent mobile device.In the grain counting step,k-means clustering segmentation algorithm was adopted to solve the problem of image segmentation caused by uneven illumination in L*a*b color space.Marker controlled watershed segmentation algorithm and area threshold method were used to solve the problem of counting adhesive seeds.The experimental results showed that the counting error rate can be controlled within 3% when the number of rapeseeds is less than 600.The system can also be used to measure the thousand kernel weight of different kinds of crops such as indica rice and mung bean.The ideal results can be obtained by running the system on different types of android devices.In summary,this study used image processing technology to measure the phenotypic parameters of rapeseed,which provided a feasible approach for phenotypic grouping and genetic breeding research of rapeseed.