Research On Maize Phenotype Detection Method And System Based On Computer Vision

Maize is an important food crop in China,with a wide distribution,high yield potential,and abundant nutrients.Crop growth models are the interactive relationship between crop growth and the environment.They are an important research direction for crop environmental perception and control,as well as for predicting crop growth and yield.They play an irreplaceable role in predicting crop yield and providing timely information on crop growth status.The current common method for detecting maize phenotype is manual measurement,which not only consumes a lot of manpower and resources,but also prone to human measurement errors.Designing and researching maize phenotype detection methods based on computer vision and convolutional neural networks,constructing the optimal model for maize phenotype detection,and achieving efficient non-destructive detection of maize phenotype detection are of great significance.The common convolutional neural network models are relatively large and difficult to deploy on mobile devices.This article focuses on the study of convolutional neural networks with fewer layers and constructs a neural network model with fewer maize phenotype detections,reducing the number of model parameters and computational complexity.The commonly used mobile devices for deploying convolutional neural networks currently have low performance and high power consumption in GPU deployment,which cannot meet the needs of mobile devices.Therefore,this article uses an FPGA development board to accelerate the deployment of convolutional neural networks.In order to achieve efficient and rapid detection of maize phenotype and deploy the model on the mobile end,this article conducts experiments at two levels: maize leaves and maize plants.The relevant work is as follows:(1)Collect image data of maize leaves and plants,and collect phenotypic data related to maize leaves and plants.The Deep Lab V3+network model trained on the PASCAL VOC semantic segmentation dataset was used to segment maize plant images.The maximum inter class variance method was used to segment maize leaf images.The segmented images were used to construct a maize phenotype detection dataset,which was normalized,partitioned,and enhanced.(2)Use MSE and MAE as the loss function and evaluation function of the regression model,and use the correlation coefficient r and the determination coefficient R ~2 As an evaluation indicator for the generalization performance of the model on the test set.(3)Conduct grouping experiments on maize leaf phenotype data,construct a convolutional neural network regression model,propose an RGB three channel separation structure,optimize the model by adding channel attention mechanism and RGB three channel separation structure,and detect different phenotypic combinations based on maize leaf phenotype grouping.Through comparative experiments on the test set,select the optimal model.The correlation coefficient r of the phenotypic detection model for maize leaves reached over 0.94 on the test set,which determines the coefficient R~2 Reached above 0.90.(4)Multi perspective image collection was conducted on the maize plant phenotype collection task,and multi perspective and single perspective experimental analysis was conducted.B-MSE and B-MAE were proposed to improve data imbalance.The feature map size was greatly reduced by using large step convolution,and the fresh weight,dry weight,and leaf area phenotype detection of maize plants were detected using convolutional neural networks,The phenotypic detection of maize plant height and leaf number adds global convolution to the basic network structure.Comparative experiments were conducted and the optimal model was selected to achieve optimal detection of maize plant phenotype.Finally,the correlation coefficient r of the maize plant phenotype detection model on the test set reached over 0.92,which determines the coefficient R~2 Reached above 0.85,with an average absolute percentage error of MAPE below 4.97%.(5)Selecting an FPGA platform for model deployment experiments,using HLS tools to implement convolutional and pooling in C++language and convert them into RTL implementation,the design of a maize phenotype detection lower machine based on FPGA was achieved.Design a visualization interface for the phenotype detection system on the upper computer,using data transmission between the upper computer and FPGA to achieve visualization of the detection system.