| With the intensification of food security problems caused by the huge increase of global population and the decrease in agricultural arable land,it is especially important to guarantee the quality and yield of food.As one of the oil crops with the highest oil production in the world,rapeseed has an important role in the production of edible vegetable oil in China,therefore,seed testing of rapeseed is an indispensable part of genetic analysis and selection of good varieties of rapeseed.At present,the seed testing method of rapeseed relies on manual counting by measuring the straight-line distance between the head and tail of the horn fruit as the length of the horn fruit,and breaking the hull of rapeseed horn fruit,and manually counting the seeds through the counting board.To improve the efficiency of rape seed testing,this study proposes an AI edge computing-based intelligent seed testing method for rape kernels and seeds to quickly measure the number of rape seeds,the number of kernels per kernel,the length of rape kernels and the surface area,mainly from the following aspects:(1)Rapid counting and localization of rapeseed seeds based on P2 PNet.The P2PNet-based rapeseed grain counting and localization method was determined by comparing the P2 PNet counting network with the conventional density regressionbased CANet.The average counting error MAE on the whole test set was 3.02,the MAPE was 0.139%,and the n AP was 99.74%.High accuracy of counting and localization was achieved in both the scenarios of adding impurities as well as high density distribution of seeds.(2)Extraction of phenotypic traits of rape horn fruit under LED transmission.The image acquisition system was built by selecting specific wavelength LED light source and image acquisition equipment,and the acquired images could clearly observe the distribution of seeds inside the rapeseed horn fruit.The P2 PNet model was used to obtain the nondestructive number of kernels per corner of rape,and the average counting error MAE per corner was 1.27 and R2 was 0.8879.The average segmentation accuracy m APcoco was 82.5% by using SOLOv2 model to segment the rape kernels acquired with LED transmission.The length as well as surface area of each hornbeam was calculated by converting the Mask of segmentation result into binary image,and the method has higher accuracy compared with manual measurement.(3)Application of edge computing-based seed testing method for rapeseed corms and seeds.In order to realize the real-time application of rape seed testing and reduce the pressure of data transmission in the cloud,the rape corner fruit and seed testing model is deployed in the edge computing device Nvidia Jetson AGX Xavier.To improve the computational efficiency of the model,the input image size was reduced for training on the PC side,and the trained model was converted to ONNX format and optimized for the inference process through ONNX Runtime.Finally,the MAPE of the rape seed counting model is 2.85% and the detection speed reaches 4.13 fps on the edge device,and the MAE of the LED transmission rape horn seed counting model is 4.80%and the detection speed reaches 3.13 fps.In summary,this study proposes a new intelligent rapeseed breeding method,which provides a rapid measurement and analysis tool for rapeseed pod and seed phenotypic traits,and assists in rapeseed breeding and yield prediction. |
PDF Full Text Request