Identification Of Weeds In Maize Field Based On Convolution Neural Network

Weeds are self-evident to agricultural production,and the cost of weed control is an important part of agricultural production practice.Therefore,in order to effectively avoid environmental pollution,waste of production materials,and even threaten human health,the problems such as how to quickly and accurately identify weed species and to spray them on a planned basis need to be solved.In order to solve these problems,this paper takes convolutional neural network algorithm as the core technology in artificial neural network,and applies it to weed identification research.The experiment designed in this paper takes maize weed in dry land in North China as the research object.The convolutional neural network in deep learning is used as the core algorithm.The common models are compared and analyzed,and the parameter adjustment optimization is optimized and wechat program is developed based on the optimal framework.The main research contents and innovation points are as follows:1.In view of the problems of incomplete research types,weak locality and no open data set,this paper collected 6678 samples of corn,Portulaca,cuspida,Portulaca,thistle,dishing flower,cowgluten,barnyard,Humulus,quinoa and longanea in the dry land of northern China,The experimental data is more characteristic and unique.At the same time,the collection of data by manual photographing makes the species selection of weed identification more extensive,which makes up for the limitation of data in this field.2.To find a model that can quickly identify weeds.In this paper,the support vector machine method of machine learning is used to extract the gray level symbiosis matrix and hog feature of weed data.The model is built by using grid search method,but the training accuracy is only 61% and can not meet the basic requirements of wechat program development.The reason is that SVM can not meet the requirements when facing large-scale data and multi classification problems.3.Aiming at the limitation of SVM,the multi model of deep learning is more optimal.In convolutional neural network,three frameworks are used,namely,AlexNet、GoogLeNet、ResNet.The accuracy of the three frameworks training has reached 85% and meets the basic requirements of commercial development procedure.The results show that deep learning has a natural advantage in weed identification,and can be more suitable for the classification of weeds under complex conditions.4.According to the frame structure characteristics of each model,the influence on the increase of model depth and convolution decomposition is discussed.In this paper,we set up two groups of tests,the deepening test of layers: ResNet18,ResNet50,ResNet101 contrast;Convolution decomposition,small convolution kernel instead of large convolution kernel test: comparison of GoogLeNet and Inception Net-V3.From the comparative analysis of experiments,it is found that convolution decomposition does not improve the recognition effect significantly,so the in-depth discussion of convolution decomposition is abandoned;Due to the special structure of residual block,the training effect of residual network will be better with the increase of depth,but the results in the experiment are quite different.In order to solve the problem,four methods of random color,contrast enhancement,brightness enhancement and color enhancement are used to enhance the data,and the sample size is expanded to33390.Due to the limitation of original data samples,ResNet101 has over fitting phenomenon,so the number of network layers is reduced.In order to find a balance between accuracy and training cost,ResNet50 was the best in the experiment,reaching 95.16%.In order to increase the actual detection effect and generalization of the model,we set up several groups of comparison between the super parameter learning rate and the optimizer.Finally,we get the best result by using the multiple decreasing learning rate and the Adam optimizer,reaching 99.7%..5.In view of the practical application of agriculture and the waste of weed medicinal resources,in order to show the results of this study more intuitively.In this paper,we use guide to design a simple software interface;At the same time,considering the portability of the actual operation,this paper also develops wechat applet which can be used across platforms to realize the application of two different end programs.Write code,establish visual interface,implement simple instructions,clearly show the results of weed identification on the interface,and attach the control measures and medicinal value of all kinds of weeds,so that farmers can clearly understand and take corresponding measures.From another point of view,understanding the advantages and treatment of weeds plays an important role in agricultural production practice.