Algorithms Research On Collecting Multiple Phenotype Parameters Of Soybean Grain

China is one of the largest soybean consumers in the world.At present,domestic soybean production cannot meet the growth of soybean demand brought about by the improvement of people’s living standards,and the situation of soybean import increasing year by year has threatened the national food and edible oil security.Therefore,it is urgent for soybean breeders to cultivate new varieties with high quality and high yield.Seed selection,also known as variety selection,is a crucial step in breeding.During the seed selection process,the phenotypic parameters of soybean grains serve as the main basis for breeding experts to select experimental materials.Traditional grain selection work requires a large amount of manpower and time,mainly relying on manual observation,sorting,weighing,and measuring,which are easily influenced by subjective factors and unable to guarantee the accuracy of selection data.Moreover,a limited number of phenotypic parameters are collected,which cannot provide comprehensive and accurate data support for computational breeding.In this thesis,the seeds of Northeast soybeans used in the breeding experiment of Northeast Institute of Geography and Agroecology,Chinese Academy of Sciences were taken as the research object,uses computer vision technology to process the soybean seed image,proposes studying on multi-phenotype parameter acquisition of soybean seed based on OpenCV and the soybean hilum phenotype detection algorithm based on the lightweight YOLOv5s model.realizes the efficient acquisition of the soybean seed multi-phenotype parameters,and has been applied in the soybean seed phenotype parameter acquisition.The method has been applied and optimized in soybean seed testing.The research work of this thesis mainly includes:(1)In this thesis.The design of a soybean grain image acquisition device,which combines shooting equipment,receiving trays,computers,etc.,to complete the image acquisition of soybean grains.The seed samples are detected on the captured images of soybean hilum,including the color change of seed umbilicus and kidney spots,and the image categories are marked with Labellmg software.By combining median filtering and Mosaic data enhancement technology to remove image noise and expand the original data set,a large soybean seed umbilicus phenotype data set is produced to provide data support for the algorithm.(2)Based on the OpenCV visual algorithm library,after image graying,binarization,scale positioning and denoising,and improving the soybean grain contour segmentation method,multiple phenotype parameters of soybean seeds can be obtained from the image at one time,including the number of seeds,the circumference and area of seeds,RGB,the ratio of long and short axes,and roundness,and the advantages and disadvantages of soybean can be recognized at the same time.The experimental results show that the recognition rate of total soybean,soybean grain,damaged soybean grain and diseased soybean grain in a single image is 98.43%,95.25%,91.25%and 88.94%,respectively.the accuracy of short axis length is 95.82%,and the accuracy of long axis length is 96.78%.At the same time,the running time is greatly shortened to meet the high-throughput automatic seed testing of soybean seeds.It provides a new method for soybean seed testing that does not need data annotation and model training,and can quickly respond to the demand of capturing new phenotype parameters.(3)This thesis proposes the soybean hilum phenotype detection algorithm based on the lightweight YOLOv5s model.The algorithm introduced K-means clustering algorithm to adjust the target prior box,replaces the last convolution module of C3 module in the trunk network of YOLOv5 with channel attention mechanism and spatial attention,and generates the CBAMYOLOv5s model to improve the accuracy and computational efficiency.Finally,the contour of seed umbilical and seed kidney is visualized and the corresponding length and width are calculated.The experiment shows that the average precision of CBAM-YOLOv5s model is 97.5%,the average precision of hilum is 84.4%,and the average precision of kidney is 86.7%,which is 5.7%higher than YOLOv5s model,mAP@0.5 is 0.895.The improved model reduces the computation by 53.0%,focuses on the target area,saves more computational power,and has stronger detection performance.In this thesis,through OpenCV computer visual library and YOLOv5 target detection algorithm and other related technologies,completes the preliminary experiment of soybean seed phenotype parameter extraction,and has been applied in seed testing work,effectively improving the efficiency and accuracy of seed testing work.It also provides reliable data support for seed examination and a new effective method for automatic seed examination of soybean.