Research Of Rice Plants Localization Technology Based On Deep Convolutional Neural Networks

At the end of the 20 th century,intelligent agricultural machinery has gradually become the main research direction of agricultural development with the rapid development of artificial intelligence technology,The positioning of rice is the prerequisite for solving the intelligent weeding and precision spraying in the field.Moreover,the dynamic planning of the mechanical travel path can be implemented by means of the position information of the rice plant,and the mechanical field is guided to realize the intelligent production of agricultural machinery.Positioning of rice in the paddy field is usually influenced by complex factors,such as the complex lighting background,field water reflection,intertwining between rice plants,blurred imaging and so on.This makes the traditional positioning algorithm that is purely relied on manually designed features of difficult to achieve accurate positioning of the strain under the complex factors.Deep Convolutional Neural Networks(DCNNs)is a new type of neural network inspired by human brain vision processing mechanism.It has the ability of self-learning feature representation,which can overcome the limitatio.This method is very suitable for the rice plant positioning under the complex factors.In this thesis,the accuracy and real-time performance of the rice plant localization algorithm are studied by using the convolution neural network.The main contents and contributions are as follows:(1)Construction of rice plant database.A complete database is a basis for the study of localization algorithm.In this paper,we construct a database with rich diversity,large scale and complete-annotated.The database is in the simulated mec hanical weeding under the real environment of image collection,covering the field of rice strains facing the complex light background,water reflection,imaging blur and other complex factors interfere with the sample.In addition,we took photos of different growth period of rice in spring and summer.In order to simulate the shooting angle change for the sample which caused by the instability of shooting device on the weeding machine.the algorithm is also adaptive,we consciously adjust the shooting angle for collection,greatly increasing the diversity of the sample.In the end,a total of 8381 samples were cleaned and marked,and the sorting and labeling of the database was of great value in the study of rice loci in the field of agricultural intelligent machinery.(2)The study solve the problem of rice plant positioning accuracy basing on the localization of rice plants,There are different size of rice plants due to different growth conditions.There are also occlusion in different camera angles,lead ing in localization algorithm can't position very accurately for small target.In this paper,a multi-scale region extractionrice plant localization algorithm is proposed.The multi-scale anchor frame clustering algorithm is used to study the candidate regions corresponding to the size distribution of rice plants,and the small target omission is avoided,and further improved on the basis of Faster R-CNN rice plant localization algorithm The accuracy of rice plant positioning.The experimental results show that the accuracy of the method is improved from 85.47% to 92.38% in the rice plant database.(3)Basing on the regression of one-stage rice plant localization algorithm the study solved the problem of real-time positioning of rice plants.Agricultural machinery needs to work in the field constantly,there is a real-time demand for rice plant positioning algorithm,whether it is mechanical intelligent weeding,automatic navigation,or other ways.so enhance the positioning speed is very important.The paper offered a new way to greatly enhance the positioning speed basing on SSD rice plant localization algorithm,through a network which can directly return to the location and category information of the rice,and avoid the time cost caused by the combination of regional extraction and classification.The experimental results show that in the rice plant database,the speed of 84.10% is 46 frames per second,which has reached the real-time demand of mechanical weeding and agricultural production.