Research And Implementation Of Ear Counting Method Of Maize At Heading Stage Based On Deep Learning

Corn is the world’s total output and the highest average yield of grain crops,is one of the world’s most important food crops,and heading stage is the most important period in life cycle of corn,and of the corn is directly related to the growth of maize and now the final production,so the heading stage of corn in the number of monitoring is forecast the key indicators of maize yield.At present,most corn ear counting tasks are realized based on target detection,and its model architecture is constructed by deep learning neural network.One of the main purposes of this method is to classify targets by sliding candidate boxes on feature maps.However,in the process of neural network training,it is extremely difficult to detect the whole area of the image due to the huge computational cost when facing the scene with high density and serious overlap of the target.However,high density distribution and overlapping of maize plants are common in the field,so target detection is difficult to become the optimal solution of ear counting.In addition,the current ear counting model based on deep convolutional neural network faces a common problem,that is,the counting performance of ear corn will be different in different regions.Therefore,the ear counting method of maize faces many challenges.In view of the above problems,the following aspects are studied in this thesis:(1)A plant counting method(TNet)based on local regression network was proposed.Local regression was used to solve the problem of crop overlap,and regression was more specific for counting than detection.The method takes the network with similar structure to Le Net as the backbone,and adds counting module and merging normalization module to the back-end network.In the module,regression counting is performed on all feature graphs of locally intensive sampling,and then all local counts are mapped back to the original image to generate the final count graph.The experimental results show that the MAE/MSE of Maize Tassels count method is 5.5/8.5 on public Maize Tassels data sets,and the MAE/MSE of Maize Tassels count method is the second best on other two data sets with comprehensive background(different density grades).(2)On the basis of the above TNet ear counting model,DATNet,a model optimization scheme based on density classification,was proposed to effectively alleviate the overestimation of ear counting in high-density scenario and underestimation in low-density scenario.The optimization scheme of particular way is,first of all to TNet as backbone network generation has a certain accuracy of two kinds of density figure,then let the original image through two branch network respectively to generate the corresponding density figure attention mask and scaling factor,and then let them multiply by single density diagram based on attention,finally will get the final density figure density figure together.The experimental results on three data sets prove that the DATNet proposed in this thesis can alleviate the difference of counting performance in different regions and further improve the counting performance.(3)An ear counting system was designed and implemented,and the ear counting model was applied to production practice.The system has the function of automatic counting of corn images,which can automatically count the uploaded corn images through the background model and display the results on the front page.Firstly,the application significance of the system is discussed.Secondly,according to the development cycle,the requirements of the system are analyzed to the specific implementation of the development process.Finally,some functions of the system are displayed in a visual form.Users can upload images to be processed in the front-end browser,or enter valid data such as a date,and the system immediately invokes the plant count model as needed,or invokes a database query to return predicted count results or historical information data in real time.