Detection Of Soybean Plant Phenotypic Characteristics Based On Deep Learning And 3D Reconstruction

Soybean is an important food and cash crop in China.However,China’s soybean industry has been highly dependent on imports for many years,and the import sources are highly concentrated,which poses a serious threat to national grain and oil security.The low efficiency of soybean planting and high production cost are the fundamental reasons for the decline of soybean industry in China.Soybean seed test is a key step in soybean breeding by obtaining repeatable soybean phenotypic data and quantifying its relationship with soybean yield and quality.However,the traditional manual test method has high cost of manpower and material resources,low measurement accuracy,limited data collection and poor data sharing ability.Therefore this paper proposes a 3D reconstruction based on deep learning and soybean plant phenotypic feature detection method,through the use of deep learning and 3D reconstruction technology process occurring in the process of soybean Kao Zhong image data,from the image data mining multidimensional phenotypic traits of soybean,for soybean breeding to provide more dimensions data reference.The specific contents of the study are as follows:(1)Preparation of soybean plant data set.Soybean plant images were collected and processed,and a new soybean plant data set was finally established to provide data reference basis for further breeding research.(2)Recognition and counting of soybean leaf shape and color based on improved YOLOv5.In order to improve the robustness and generalization ability of the model,an improved YOLOv5 algorithm was proposed to replace the trunk network as Mobile Netv2,introduce ECANet attention mechanism and improve the loss function as CIOU＿Loss+DIOU＿NMS.The experimental results show that the m AP and FPS of improved YOLOv5 are 96.13% and 79 respectively.Compared with YOLOv5,the IMPROVED YOLOv5 algorithm improves the FPS by 17.91%,reduces the number of model parameters by 39%,and reduces the weight by 55.56% when the m AP decreases by only 0.34%.The algorithm error detection and missed detection anomaly are reduced.Finally,the recognition and statistics of soybean leaf shape and color based on improved YOLOv5 were tested and analyzed.The correlation coefficient between the measured value and manual value was 0.98384,which showed high accuracy and good consistency.(3)Soybean leaf segmentation and leaf area measurement based on 3D reconstruction.Image data set of soybean plants,respectively are estimated based on the depth of the MiDaS and YOLACT segmentation,the instance of the division of the instance mask extract leaf depth information through coordinate system transformation into a 3D point cloud,the algorithm of soybean plant leaf area measurement and accuracy analysis,algorithm measured value and artificial measurement correlation coefficient is 0.83918,compared with the traditional method,it has a great improvement.(4)Soybean plant phenotypic characteristics detection and data storage platform system.Based on the completed target detection and 3D reconstruction model,a soybean plant phenotypic characteristics detection and data storage platform system with the above functions was designed and developed for agricultural technologists and researchers,providing an effective way for breeding.