Research On Detection And Counting Method Of Wheat Spikelets Based On Deep Learning

Wheat is one of the important food crops in my country and plays an important role in ensuring national food security.In wheat breeding and yield prediction,the number of wheat spikelets is an important ear characteristic parameter,which can reflect the quality,growth and yield of wheat varieties.Breeding and yield estimation are of great importance.This paper takes the images of wheat ears of multiple varieties and growth periods obtained in the field environment as the research object,and realizes the counting of wheat spikelets based on the deep learning method.At the same time,the related application system is designed and implemented.The main work and results of this paper are as follows:(1)Wheat ear image collection and spikelet data set production.In order to improve the diversity and complexity of experimental data,the images of wheat ears of four wheat varieties at three different growth stages of flowering,grain filling and maturity were captured by smartphone,and then a network model for target detection was created.Trained wheat spikelet dataset.(2)Spikelet detection and counting method based on deep learning.Three object detection network models,SSD,Faster R-CNN and YOLOv5,are constructed,and the spikelet dataset is used for training and testing.In the test set images,the three models were tested and evaluated according to the variety and growth period.The results show that the YOLOv5 model has the best performance,the m AP value on the test set is 0.997,the coefficient of determination R2 between the model predicted spikelet number and the artificially counted spikelet number is 0.89,the root mean square error RMSE is 0.60,and the average count The accuracy Acc is 98.88%,which has high counting accuracy and strong robustness,and can effectively count wheat spikelets.(3)Lightweight improvement based on the YOLOv5 s model.The YOLOv5s-T model is obtained by lightweighting and improving the backbone network Backbone of the YOLOv5 s model,and the YOLOv5s-T+ model is obtained by reducing partial upsampling and feature fusion modules.The experimental results show that the YOLOv5s-T+ model further reduces the dependence on hardware computing resources.The GPU memory usage during the training process is 2.92 G,which is 0.5G lower than that of YOLOv5s-T,and 1.06 G lower than that of YOLOv5s;the model inference time is 1.8ms,which is 0.2ms lower than YOLOv5s-T and 0.7ms lower than YOLOv5s;the model size is 9.1M,which is 35.5% smaller than the original model YOLOv5 s.The average accuracy m AP on the test set reaches 99.3%,which is only 0.4% lower than that of YOLOv5s;the R2 for wheat spikelet counting is 0.88,which is only 0.01 lower than that of YOLOv5s;the average counting accuracy Acc is 98.66%,which is only 0.22% lower than that of YOLOv5 s.(4)Design and realization of wheat spikelet detection and counting system.With YOLOv5 as the core algorithm of the system,a wheat spikelet detection and counting system is implemented based on the Vue and Flask framework,which can quickly obtain information such as the location,category and number of spikelets in the image according to the wheat spike picture uploaded by the user,which can be used for wheat breeding and Spikelet counting in the process of field yield prediction can provide some help.