Research On Soybean Terminal Test Phenotype Acquisition Method Based On Machine Vision And Deep Learning Algorithm

Soybeans are important grain and oil crops which originated in China.However,we still have tens of millions of tons of gaps that need to be imported to maintain food supply each year.The reason is that the lack of high-yield and high-quality varieties has caused farmers to lack enthusiasm for planting,and the planting area is constantly decreasing,too.In response to this situation,the breeding of high-yield and high-quality soybean varieties has become an urgent task for soybean breeding workers.How can new soybean varieties be cultivated stably and efficiently? It will play a key role in optimizing the traditional breeding process.The most important current optimization breeding strategies are molecular assisted breeding and molecular design breeding.However,whether these two strategies can be effectively implemented depends on the degree of analysis of the genetic basis of related traits.The precise and general quantification of phenotypic identification is an important basis for analyzing the genetic laws of traits,which shows that automatic and accurate phenotype acquisition and measurement has become one of the key nodes to accelerate the breeding process.This research aims at the phenotype survey stage of soybean harvest,that is,the soybean plant in the "seed test" stage,using various model algorithms of machine vision and deep learning technology as means to automatically identify,position,and measure the decomposed stems,pods and grains,and then obtain a series of high accuracy and high stability algorithms and models.In addition,for another important issue in the breeding process is the selection of seeds.This study explored,and also harvested two efficient and accurate neural network models.The experimental materials in this article come from the resource group collected in the early stage of the laboratory and the completed RILs and CSSLs groups.The image processing and deep learning tasks are in an operating system of Windows10,running memory of 32 G,and the processor of Intel Core i9 on a server equipped with a NVIDIA Titan Xp GPU.The specific research work of this article is as follows:(1)Automatic identification of related main soybean stem phenotypes based on target detectionThe main phenotypes of soybean main stem mainly refer to the length,number and interval of main stem.This research uses the pixel-level target detection network Mask R-CNN.First,we use the main stem node as the target,and the target detector is obtained through training and optimization of the network.Then,by comparing,classifying and framing the target,the target object position and size in pixels are calculated,and finally,the true phenotype of the main stem is obtained by reference point correction.In addition,the material selection method based on the main stem phenotype is also given in this section,and practiced in an introduction line group.(2)Soybean pod classification and automatic extraction of phenotypes based on transfer learning strategySoybean pod classification refers to the classification of pods with different pod numbers(generally can be divided into one pod,two pods,three pods and four pods).The phenotype of a single pod is mainly the length and width of the pod.First,this study trains,optimizes,and models multiple migrated deep networks.Then,a comparative analysis is performed to obtain a pod classifier with higher accuracy.Finally,the length and width of a pod of a certain type are calculated by machine vision technology.(3)Research on soybean seed testing and selection based on machine vision and deep learning algorithmThe phenotypes of soybean seeds tested mainly include grain length,grain width and seed coat color;soybean selection mainly refers to the screening of seed quality before planting.First,in this study,machine vision technology was used to extract phenotypic parameters such as shape,color,and texture of soybean seeds.Then,through a series of operations such as designing the network,training the network,and optimizing parameters,a convolutional neural network that can efficiently and accurately identify the quality of the kernel is obtained.Finally,a high-precision convolutional neural network that refines and classifies soybean grains of different qualities is obtained through transfer learning.Through the exploration of the above research work,this project obtained a series of highly accurate and highly stable algorithms and models.These results provide the most important core algorithms and models for the soybean automatic seed testing platform to be built and the automatic seed selection before sowing support.