Research On Object Detection In Port Based On Optical Remote Sensing Image

Port is an important part of international trade and an important basis for measuring regional economic development.It is of great practical significance to analyze the current port operation situation indirectly by detecting the quantity changes of various object types in the port,so as to further forecast the current economic form.However,due to commercial interests and national security and other factors involved,the specific operation data of ports are often unable to be directly obtained.Remote sensing satellites can obtain global remote sensing images periodically without being affected by geographical restrictions and artificial subjectivity.With the continuous progress of remote sensing technology,remote sensing satellite images with spatial resolution as high as 0.1m can be obtained at present,and surface object information can be clearly seen.Therefore,high-resolution remote sensing images can be used for fine-grained detection and recognition of ground object information.However,compared with natural image,optical remote sensing image has many difficulties,such as large size difference of each object,large angle change,dense and small object distribution,complex background,etc.,which brings greater difficulty to the object detection task.False detection and missing detection will be more serious than natural image object detection.To address the above difficulties,this thesis proposes a two-branch object detection network TBT-MYOLO that integrates Transformer and MLP based on YOLOv5,and develops an AI service platform for object detection in port based on the research results.The main research contents of this thesis are as follows:First,the remote sensing object detection dataset is constructed.According to the research content of the project,a port remote sensing dataset supporting rotating object detection was constructed by integrating different public datasets,including four categories of containers,oil tanks,trucks and ships.Moreover,an algorithm model supporting rotating object detection was constructed based on the YOLOv5 algorithm.Secondly,a two-branch object detection network TBT-MYOLO is proposed,which integrates Transformer and MLP.In order to enhance the detection performance of the model,a two-branch network architecture is designed to enhance the learning ability of the model.Moreover,Transformer and MLP are integrated into the model design,and the powerful modeling ability of Transformer and the powerful learning ability of MLP are utilized to further improve the model detection performance.Finally,the port object detection AI service platform is designed and developed.The system adopts the mainstream design mode of separating the frontend and backend.The frontend is based on Vue.js,and the backend is based on Django.It integrates the functions such as datasets management,model training,model testing and model evaluation.