Research On Soybean Plant Phenotypic Feature Detection Method Based On Machine Vision

Soybean is one of major source of protein in people's daily diet,which contains high-quality plant protein.The research of soybean germplasm resources is the basic guarantee for sustainable development of soybean industry.Crop phenomics research is not only to obtain high-quality,highprecision and repeatable plant phenotypic data,but also an important part of germplasm resources research.It is the investigation of plant and seed characters of mature soybean to understand the effects of different varieties and different treatment elements.Breeding experts constantly carry out the cycle of hybridization,seed selection and re-hybridization to identify and screen germplasm resources,and screen out the seeds of excellent soybean plants in each stage,and continuously hybridize with other excellent seeds,to cultivate high-yield and high-quality soybean varieties.However,the traditional germplasm resources inspection mainly relied on manual operation,observing,measuring and recording data one by one for crop plants or seeds,and there were some problems,such as poor repeatability,limited amount of survey data,large human factors,low efficiency and high cost.Soybean breeding could be made the design of breeding program more efficient by the vision technology.This paper used machine vision as the theoretical basis to construct soybean plant phenotype platform,through image processing,image segmentation,convolutional network,artificial intelligence and other algorithms to process soybean plant image.To realize the informatization on soybean seeds inspection and provide technical support and theoretical basis for the realization of intelligent breeding.Specific studies were as follows:(1)Research and design of soybean plant image acquisition platform based on deep convolutional neural network.In order to recognize and locate the pods and stems of soybean plants more efficiently,an improved SSD convolutional network(IM-SSD)was proposed.At the same time,a labeling method for pods and stems in soybean plant images was proposed.The purpose was to reduce the misjudgment caused by occlusion.For the labeling of stems,only the parts of stems that were not occluded were labeled.In addition,in order to increase the number of training samples,the labeled training samples were randomly amplified to improve the overall performance of the network.IM-SSD added two residual network layers based on traditional SSD structure.Through the fusion of low-level features and high-level features,the ability of small target detection was enhanced,and the recognition rate of the new network was higher.Input images using 600 × 300 pixels to reduce the impact on compression deformation.Through the experiment,the pod recognition precision of SSD 300,SSD 512,and IM-SSD manual detection was 83.49%,87.52%,and 89.68% respectively.Among three deep convolutional networks,the IM-SSD model was more stable and effective in the detection of soybean pods and stems.(2)The method of extracting to stem model based on ant colony algorithm was used to compensate for the deficiency in IM-SSD convolution neural network of segmented stem location.Through IM-SSD convolution neural network to get the stem location,the ant colony algorithm is used to connect the segmented stems and extract the stem model.According to the complete stem model,the phenotypic characteristics of soybean stem were extracted,including the number of effective branches,main stem,plant type,plant height and bottom pod height.The precision rates were 95.56%,96.12%,90.28%,94.43% and 92.51% respectively.Through comparative experiments,the precision of extracting soybean stemmed phenotypic characteristics based on ant colony algorithm was slightly lower than that of manual measurement method,but it reduced human labor.From the overall precision of stem phenotypic feature extraction,the method could basically meet the requirements of soybean stem phenotypic feature extraction.(3)Pod phenotype features extraction based on watershed image segmentation.An improved watershed algorithm was proposed,which divides the image segmentation into three stages: image preprocessing with MC-Watershed algorithm;subdivision of under segmented regions;merging of over segmented regions.Adaptive size structures unit region and fast robust fuzzy c-means clustering algorithm were used to re-segment the under-segmented region.Because the preprocessing stage and the subdivision stage were produced over-segmented regions,it was necessary to merge over-segmented regions of the segmented image.In the merging process,the skeleton of the region image to be merged was extracted based on the thinning algorithm,and the least square method is used for line fitting.If the fitting error was less than the maximum error value,the pre fitted regions were merged,otherwise they were recombined,until all regions did not meet the merging conditions,the over segmentation merging was ended.Through comparative experiments,the VAC value of the improved watershed algorithm was 25.09% higher than that of MC-Watershed algorithm.It showed that the improved watershed algorithm was more effective in pod segmentation of soybean plant image.In pod phenotypic feature extraction,image binarization,skeleton extraction,image rotation,adaptive regression curve extraction and local maximum of regression curve were used to obtain pod length,pod width,pod size,pod type and number of seeds per pod.The comparison of this method with Vgg16,Alex Net and Google Net,showed that the accuracy of this method was slightly lower,but the preprocessing time was short,no network training was needed,no manual annotation was needed,and it was more suitable for the discrimination of small sample set.(4)Construction of soybean plants rotating device based on multi view matching.The platform includes: computer processor,camera and lens,2.4G wireless timing remote controller,acquisition platform box,tricolor fluorescent lamp,fixing clip,electric motor and blue background.Through the rotating parts to make the plant rotate continuously,the 360 degree multi angle views image of each plant is collected.The IM-SSD network was used to identify and locate the pods of soybean plants.The feature points of three adjacent images were extracted and matched by the accelerated up robust features algorithm.The overlapped relationship between views on image matching was used to complement the information of images from different angles.The results showed that the precision of pods per plant detected by multi-angle was 4.68% higher than that detected by single angle.(5)The soybean plant images were obtained by the constructing soybean phenotypic detection platform and the soybean plant rotating device respectively.Experiments showed that plant height,effective branch number,main stem,plant type,bottom pod height,the number of pods,pod width,pod length and pod type could be obtained through deep convolutional network,ant colony algorithm,skeleton thinning,Hough detection and SURF algorithm,et al.The data of soybean plant phenotypic characteristics were stored in the database,and the subsystem of soybean plant phenotypic characteristics detection was established.