Research On The Method Of Maize Tassel Counting And Heading State Recognition Based On Semi-supervised Learning

As one of the main grain crops in China,maize’s tassel quantity during the tasseling stage is closely related to its yield.The successful tasseling of maize is the key to its pollination.Therefore,identifying the tasseling state and quantity of maize tassels during this stage is of great significance for estimating yield and conducting scientific artificial pollination.In the process of using machine learning technology to identify maize tassels,a large amount of labeled data is required to train accurate machine learning models.However,labeling maize tassel data is extremely tedious and requires a lot of manpower,material resources,and financial resources.Furthermore,the complex field environment and variable weather conditions can lead to noise,background interference,and other issues in the images,which can affect the accuracy of identification.In response to the above problems,this paper takes maize tassels as the research object,uses deep learning technology and semi-supervised learning technology to monitor the tassel heading state at the heading stage,designs and implements a maize tassel heading state monitoring system,and intuitively shows Differences in maize tassel heading status and distribution.The main work done in this paper is as follows:(1)Construction of Maize Tassel Dataset.The research data in this paper is the image data of maize tassels in the maize experimental base of Shandong Agricultural University,and the collection methods are high-definition digital camera shooting and UAV collection.First,use image enhancement and image noise reduction to expand the data set,and then use Label Img software to label two types of maize tassel image data,and make 4896 maize tassel data sets based on high-definition digital cameras and maize tassel images based on UAV images.There are 958 tassel data sets,and the two types of data sets have marked more than 200,000 tassels.(2)Construction of maize tassel monitoring model.In this paper,according to the different characteristics of maize tassel heading,it is divided into two states: heading and non-heading.Based on the Faster R-CNN network structure,the Ftassel R-CNN model is constructed to identify and count the tassel heading state at the maize heading stage.The model uses Res Ne Xt50 as the feature extraction network to avoid the gradient disappearance and gradient explosion problems encountered in the maize tassel detection process,using the Recurrent Criss-Cross Attention mechanism module(Recurrent Criss-Cross Attention,RCCA)and the empty space pyramid module(Atrous Spatial Pyramid Pooling,ASPP)method to capture the key information of maize tassel heading state.The experimental results show that the improved Ftassel R-CNN model has achieved excellent results in the accuracy,recall,and average precision of maize tassel heading status recognition and counting on the two types of data sets,which are 95.57%,94.34%,94.36%,and 92.05%,93.42%,91.51%.(3)Design and implementation of a semi-supervised learning algorithm based on updated pseudo-labels.To address the large number of tassels in the maize tassel dataset and the difficulty in labeling,a semi-supervised learning method based on updated pseudo-labels is proposed in this paper.The method incorporates a pseudo-label update mechanism and changes the loss function to Focal loss to address the problem of imbalanced positive and negative samples in maize tassel recognition.Experimental results show that the proposed method performs better than both the fully supervised method and traditional pseudo-label semisupervised learning methods for maize tassel detection on datasets with different proportions of labeled data.The average precision of tassel detection on the two datasets with half of the annotations is 92.56% and 89.45%,respectively,which is close to the average precision achieved on the fully labeled dataset while reducing labeling costs and ensuring recognition accuracy.(4)Design and implementation of maize tassel counting and heading state recognition system.In this paper,the improved Ftassel R-CNN detection model and the common detection model in the comparative experiment are integrated into the system,and the tassel change curve diagram module is added to facilitate the user to directly observe the change of the number of maize tassels.