Large Transportation Detection Of Remote Sensing Image Based On Rotating Region Proposal Network

With the rapid development of space remote sensing technology,the number of Earth observation satellites launched in recent years has been increasing.At present,the amount of accessible remote sensing data is increasing,data types are enriched,data quality is improved,and image resolution is increased.Remote sensing data can be widely used in various fields such as military,agriculture,surveying and mapping,geology,hydrology,environmental monitoring,etc.The application of increasingly abundant remote sensing data in various fields has ushered in new opportunities and also posed new challenges to remote sensing image interpretation.The detection and recognition of objects in remote sensing images is one of the important ways of remote sensing image interpretation.Remote sensing objects usually have complex background,arbitrary direction,blurred edges and other characteristics,which are significantly different from natural images.These characteristics make the detection and recognition of remote sensing objects more difficult,and also put forward higher requirements on the robustness of relevant algorithms.Traditional detection methods are mainly implemented based on manually designed features,which can show better performance and achieve higher accuracy rate when detecting objects in specific scenes,but often lack robustness for detecting objects in complex scenes.In contrast,the currently popular deep learning methods can learn deep image features from a large amount of training data through multilayer neural networks.Although the deep learning-based detection method is not as interpretable as the traditional method,it can make full use of the existing massive remote sensing data to extract the deep features of images,so it can often achieve better results than the traditional method in complex scenes.The rapid detection of transportation tools in remote sensing images can effectively enhance the capability of space-based remote sensing applications,which is of great significance to transportation monitoring,maritime vessel search and rescue,wrecked aircraft localization,military reconnaissance,etc.The current mainstream object detection methods generally use area region proposal network to locate objects,the predicted horizontal bounding boxes are usually represented by four position parameters,including the coordinates of the object’s center point and the length and width of the object.In contrast,for objects in remote sensing images taken from the bird’s-eye view,their orientation is arbitrary,and in some cases,using a horizontal bounding box to represent it will lead to inaccurate or serious overlap.In view of this,rotating region proposal network is proposed for locating directed objects,and their predicted rotational bounding boxes are usually represented by five positional parameters,including the center point coordinates of the object as well as the length,width,and angle of the object.To address the problems of existing rotating object detectors,such as large number of parameters,complex computation,learning difficulties,and inaccurate prediction,the following work is done in this thesis:1.A new rotation-decoupled bounding box representation method is designed.The method combines the advantages of easy learning of horizontal bounding box and more accurate representation of object by rotating bounding box,and avoids the interference of angular periodicity problem on model learning.2.A rotation-decoupled region proposal network is designed,and a new single-stage rotation-decoupled object detection model is further constructed.The model uses a unified framework to complete the detection of horizontal and oriented object,and for the detection of oriented object,the model requires less number of parameters and can keep the training process more stable.3.The detection of transportation in remote sensing images is investigated.The proposed rotation-decoupled detector is used to perform experiments on several publicly available remote sensing datasets.The higher accuracy detection of common categories of transportation in remote sensing images,such as aircraft,ships,vehicles,etc.,is achieved.