Research On Target Detection And Weed Classification And Identification Of Wheat Ears

Purpose —In the development of agriculture,improving the quality,yield and economic benefits of crops is the focus of work,and weed control and removal is a must.After weeding,it is beneficial for crops to obtain sufficient nutrients,water and growth space,which ensures the growth of crops and maintains high yields of crops.How to accurately detect the wheat ears of crops in large-scale image data sets provides convenience for subsequent wheat breeding and wheat yield,this is of great significance to the development of agriculture.Therefore,this paper will use image recognition technology to study the classification of different types of crops and weeds and the target detection of wheat ears,and develop a small plant classification and detection system for the follow-up study and research work.Method —In the classification stage of different crops and weeds,the use of image recognition and classification technology was proposed to extract the characteristics of different plants.In this paper,four pre-trained models,VGG16,VGG19,Inception V3 and Res Net50 are firstly studied by transfer,and then a convolutional neural network with a depth of 9 layers is built to compare with the above four models.For the experimental effect,and to solve the problem of imbalanced image datasets of different crops and weeds in the dataset,the method of data enhancement is mainly used.In addition,a validation set as well as a test set will be used to evaluate the ability to recognize the classification model.In the crop wheat ear target detection stage,in order to detect the wheat ear in the image faster and more accurately,the feature extraction of the crop wheat ear in the image mainly uses the convolutional neural network Res Net101 and VGG16 algorithms,using Faster R-CNN algorithm detects the wheat ears of crops,thus forming the Res Net101+Faster R-CNN model and the VGG16+Faster R-CNN model are compared.During the detection process,the augmented image data set mainly adopts the method of data enhancement,and the 3-Fold cross-validation method is used in the experiment process,which greatly reduces the experimental chance caused by the random division of data.In addition,experiments were conducted on YOLOv5,the latest representative model of the "onestep" method of target detection,and a comparative study was conducted with the above two models.Finally,the experimental results are displayed,compared and analyzed.Findings —In the classification stage of different crops and weeds,four network models,VGG16,VGG19,Inception V3 and Res Net50 are trained on the established training set.The better network model is VGG19.Then combined with the VGG19 network model,this paper designs a 9-layer network model to improve the classification effect of crops and weeds.Experimental results show that the network model designed in this paper is better than VGG19.Finally,based on the network framework designed in this paper and the four models mentioned above,the accuracy and loss rate of each model in the set are discussed,and it is concluded that the performance of the 9-layer network model constructed in this paper is better.In the target detection stage of crop wheat ears,the Faster R-CNN model based on Res Net101 and VGG16 and the YOLOv5 model are used to compare the target detection accuracy after training and the results of target positioning loss,and Res Net101+ Faster R-CNN can be obtained.The training result of the CNN model is significantly higher than that of the VGG16 based Faster R-CNN model.Finally,compare it with YOLOv5.Through comparison,it is found that the m AP value of YOLOv5 is not much different from the Faster R-CNN algorithm,but the target loss is much lower than the Faster R-CNN.Therefore,YOLOv5 algorithm has a relatively outstanding effect on small target detection,and can well solve the problem of mutual occlusion of crops,wheat and ears.Limitations of research — In the classification stage of different crops and weeds,the feature learning of different crops and weeds using image recognition classification technology is relatively shallow,and the model in this paper is more similar to the plant seedlings in the subtleties.There are still gaps in the feature learning of,and further research is needed.Moreover,the research targets mainly include five crops and five weeds,and the scope of the research targets is small,and the data set needs to be further expanded.In the target detection stage of crop wheat ears,due to the overlapping images of wheat ears,the wind direction will blur the image,resulting in a low detection accuracy,which needs to be further improved.Practical implications —Compared with traditional methods,the application of image recognition and classification technology to crop cultivation and production can effectively control the loss of crops caused by weeds,reduce the use of herbicides and the cost of weed control.time,is conducive to the sustainable development of the ecological environment.During the detection of wheat ears,various traits of wheat,such as density,health status and number of ears,can be efficiently estimated,making it important in wheat yield estimation,wheat breeding and crop management.A visual system has been developed,which is very helpful for follow-up research and learning.Value —It is proposed to use image recognition and classification technology to classify different crops and weeds,so that weeds can be effectively removed,which makes up for the defects of traditional field weed control methods that are cumbersome,low in control accuracy,time-consuming and labor-intensive.The deep learning method is used to effectively distinguish different plant seedlings and effectively remove weeds.The image recognition and classification technology is used to extract and learn the characteristics of wheat ears in the image,and the Faster R-CNN algorithm and the YOLOv5 algorithm target detection algorithm are used to detect the target of the wheat ears of crops,so as to locate the position of the wheat ears in the image.Conducive to accurate production estimates,etc.A plant classification detection system has been developed,which is helpful for different groups of people to study and research this aspect.