Segmentation Of Maize Seedlings And Counting Of Leaves In Field Based On Deep Learning

Maize(Zea mays L.)is one of the main field crops in China,and leaf is an important trait in the process of maize growth.It is significant to estimate the seed activity and yield of maize seeding by counting of leaves.Detection and counting of the maize leaves in the field are very difficult due to the complexity of the field scenes and the cross-covering of adjacent seedling leaves.It is impossible to count leaves of a single maize seedling.In order to solve the above problems,this paper takes maize seedling as the research object,and proposed a two-stage method to segment maize seedlings and count leaves in the field based on deep learning and unmanned aerial vehicles(UAVs).In the first stage,Mask R-CNN was used to segment maize seedlings in the field background.In the second stage,YOLOv5 was used to detect and count the leaves of maize seedlings.Because the deep learning method relies on a lot of manual labeling,it is time-consuming and laborious,semi-supervised learning method was used to train the model to reduce the workload of data labeling.The specific research contents are as follows:(1)Construction of maize seedling data set.In this paper,UAV with RGB camera was used to collect 2-6-leaf field maize seedling images,and 1005 images were cut out to construct data sets.In order to improve the generalization ability of the model,various data enhancement methods were used to expand the data.(2)Segmentation of maize seedlings based on improved Mask R-CNN.A new loss function Smooth LR was proposed to improve the segmentation performance of Mask R-CNN to separate the whole maize seedlings from the complex background.The effects of Resnet50 and Resnet101 on the segmentation performance of the model were compared.The results showed that the improved Mask R-CNN had the best segmentation performance when the backbone network was Resnet50,with Bbox m AP of 96.9% and Mask m AP of 95.2%.The inference time of single image detection and segmentation was 0.05 s and 0.07 s(3)Counting of maize seedling leaves based on YOLOv5.Based on the foreground map of maize seedlings segmented by the improved Mask R-CNN,five YOLOv5 models with different parameters were applied to detect and count the leaves of maize seedlings,and the model with the strongest counting ability was selected and compared with the mainstream models Faster R-CNN and SSD.The results showed that YOLOv5 x had better performance in detecting leaves.The average precision(AP)rate of fully unfolded leaves and newly appeared leaves of maize seedlings were 89.6% and 54.0%,respectively,the counting accuracy rates were 72.9% and75.3%.(4)Segmentation of maize seedlings and counting of leaves based on semi-supervised learning.Semi-supervised learning method Noisy Student was used to train SOLOv2,the model of segmenting maize seedlings,and YOLOv5 x,the model of counting leaves.The results showed that when the label ratio was 30%,the segmentation performance of the student model SOLOv2 was close to that of the fully supervised model,with the m AP of 93.6%.When the label ratio was 40%,the leaf detection performance of the student model YOLOv5 x was similar to that of the fully supervised model.The average precision rates of fully unfolded leaves and newly appeared leaves were 89.6% and 57.4%,and the counting accuracy rates were 69.4% and72.9%,respectively.The above results show that the two-stage deep learning strategy proposed in this study can effectively segment maize seedlings based on UAV digital images and count leaves Semi-supervised learning method can not only accurately obtain the number of leaves of maize seedlings,but also reduce the data labeling cost by 65%,which can provide new ideas for exploring the the method of count leaves of of field-grown crops based on UAV.