Research On Detection And Counting Methods Of Maize Plant Population During The Seedling Stage

The accurate counting of maize seedling populations at the seedling stage not only enables effective monitoring and management of the seedlings,but also allows us to determine the seedling emergence rate based on the counting results.This information is crucial for making breeding plans for the following year.With the rapid development of computer vision technology and the concept of smart agriculture,the traditional manual counting method is no longer sufficient.Therefore,this thesis combines deep learning techniques with target detection algorithms to conduct research on the growth state of maize plants at the seedling stage.The main research work is outlined as follows:(1)The thesis explores the application of box-based supervised algorithm to detect and count maize plants at the seedling stage.To facilitate the model’s transferability to mobile platforms for seedling maize plant detection,the YOLOv5 network model’s backbone network is lightened.The nearest neighbor upsampling,which fails to capture the rich semantic information of maize seedling features,is replaced with the content-aware reorganization module.Additionally,the accuracy of the lightened backbone network is enhanced by introducing the Sim AM attention mechanism without increasing the number of parameters.Experiments were carried out on seedling maize cluster data set to verify the effectiveness of the improved network structure,and the detection results were analyzed.The YOLO-MSC algorithm demonstrates a 3.04% improvement in P value,1.69%improvement in R value,2.39% improvement in F1 value,and 2.60% improvement in m AP compared to YOLOv5 on the maize seedling dataset.and its detection and counting effects were improved,indicating that YOLO-MSC algorithm has certain application value for maize seedling breeding research.(2)The thesis investigates the application of the point-based frame algorithm for detecting and counting maize populations at the seedling stage.To better capture global target features,the feature layer in P2 PNet is connected with residuals and fused with features using convolution to increase the number of channels and bilinear interpolation to reduce the width and height.The acquired new features are then inputted into the Neck layer for feature enhancement.Compared with P2 PNet,the P2P-IM algorithm reduces the MAE by 4.00 and MSE by 10.85.(3)Mobile System Interface Design for Seedling Maize Plant Detection and Counting.This thesis focuses on the development of a mobile system interface for the purpose of detecting and counting seedling maize plants.In order to assist researchers in accurately identifying and quantifying maize seedlings in field conditions,we have adapted the YOLOMSC and P2P-IM network models for mobile deployment.Additionally,the mobile system interface showcases the detection outcomes and counting results produced by both algorithms.To mitigate counting errors,the counting results obtained from the two algorithms are averaged.The improved algorithms(YOLO-MSC and P2P-IM)presented in this study demonstrate superior performance in detecting and counting seedling maize plants compared to the original algorithm when evaluated using the seedling maize strain dataset.This research provides a novel approach for detecting seedling maize clusters in a complex environment,exhibiting both theoretical research significance and practical applicability.Furthermore,the enhanced algorithms hold significant potential for broad application in the field of agriculture.By offering scientifically accurate technical support for agricultural production,they contribute to the advancement of smart agriculture.