Research On The Methods Of Wheat Ears Counting And Spatial Distribution Visualization

Wheat is an important food crop in the world,and accurate estimation of wheat production is of great significance to ensure world food security.The number of wheat ears is an important parameter for wheat yield estimation.Therefore,this paper uses five deep learning models of Faster R-CNN,SSD,Retina Net,YOLOv3 SPP and YOLOv5 s to train,verify and test different wheat ears data sets.Among them,the YOLOv5 s model wheat ear recognition accuracy rate of wheat ear dataset A based on UAV data is the highest,reaching 97.39%.In order to further improve the recognition accuracy of the model for wheat ears,the Faster R-CNN and YOLOv5 s models have been improved.The improved Faster R-CNN model and the YOLOv5 s model have higher recognition accuracy of wheat ears for the three wheat ears data sets respectively.96% and 98%.In addition,this paper also uses image processing,U-net semantic segmentation and YOLOv5s+Deep Sort method to process wheat ear images and videos,and generates a corresponding visual image of the distribution of wheat ears in a certain space,and develops wheat ears identification and counting system.The main work of this paper is as follows:（1）Different wheat ear datasets were constructed.According to the different models and detection targets,this paper uses different methods to preprocess the wheat ear video data collected by the drone,and then uses different data labeling tools to construct two types of wheat ears from the preprocessed wheat ear images.Data sets,one for wheat ear target detection and the other for wheat ear semantic segmentation.（2）Counting method of wheat ears based on deep learning.In order to compare the detection effect of the model on the wheat ear data set,the wheat ear data set based on UAV data,the wheat ear data set based on GWHD and the mixed wheat ear data set based on UAV data and GWHD were imported into Faster R-CNN,SSD,Retina Net,YOLOv3 SPP and YOLOv5 s models are trained,validated and tested,and Precision,Recall,Accuracy and F1-Score are calculated based on the detection results.The Faster R-CNN and YOLOv5 s models are improved to further improve the recognition accuracy of the model for wheat ears.（3）A visualization method for the spatial distribution of wheat ears.In order to facilitate the observation of the density and size of wheat ears in a certain space and the difference in the number of wheat ears in different sections at the same interval,the important problems in wheat agricultural production were solved by in-depth analysis of the reasons for the differences.This paper uses image processing,U-net semantic segmentation and YOLOv5s+Deep Sort method to process wheat ear images and videos,and generate corresponding visualization images.（4）Design and implementation of wheat ear recognition and counting system.In order to make the improved Faster R-CNN and YOLOv5 s model widely used in wheat ear counting research,this paper designs and develops a system with wheat ear recognition counting and semantic segmentation functions.