Researches On Soybean Phenotype Measurement Method Based On Deep Learning

Soybean is one of the main sources of protein in human food and an important crop.Soybean planting has a long history in China.However,with the growth of soybean demand,due to the low yield per mu,the self-yield of soybean in China has been seriously insufficient.At present,most of the supply depends on imports,which has threatened the national grain and oil security.In order to solve this problem,it is imperative to cultivate high-quality and High-yield Soybean Varieties.The mature dry soybean trees are generally used in soybean seed test.After observing the mature state of soybean plants,the phenotypic data of soybean plants are collected.Among them,pod number index is the key data which directly shows the yield of soybean per plant.At present,the method of soybean seed test is mainly manual observation and count,which is inefficient and difficult to ensure the accuracy of dataset.Therefore,this thesis proposes a method of soybean phenotypic measurement based on deep learning,in order to realize the intelligent seed test of soybean.The main researches are as follows1.A kind of soybean plant phenotype data measuring device was designed,which realized the semi-automatic acquisition of soybean plant image by combining PLC controller with mechanical equipment.Based on computer vision processing technology,a program was written to make soybean plant phenotype data set from the collected images;2.A method of soybean plant data enhancement based on DCGAN is proposed.DCGAN and conventional methods are used to effectively expand the initial soybean plant data set and reduce the over fitting phenomenon in deep learning model training.After training,the detection effect of the detection model is greatly improved;3.An embedded measurement method based on Jetson nano is proposed,which combines Jetson nano embedded motherboard with software,uses lightweight network YOLOv4-tiny to train soybean phenotype detection model,uses tensorRt technology to accelerate and deploy the model,and completes rapid detection of soybean phenotype on the basis of equipment volume less than 1LIn this thesis,a kind of soybean plant phenotypic data collection device was developed and put into operation.The soybean phenotype measurement method based on deep learning proposed in this thesis can greatly reduce the workload of manual image acquisition in the early stage,effectively improve the quality of the collected image,and the data set after data enhancement can effectively avoid over fitting problem.The embedded measurement method combines software and hardware,greatly improves the practicability of the method,and provides big data support for breeding experts to design accurate breeding programs.