Rotated Object Detection For Remote Sensing Image Using Deep Learning

Due to the continuous development of remote sensing image technology and the richer application scenarios,object detection algorithms for remote sensing images are constantly receiving attention.The vertical rectangular object frame output by the traditional object detection algorithm has various problems such as inability to effectively separate the objects when facing dense objects and large aspect ratio objects in remote sensing images.In this context,the niche area of rotated object detection has emerged.Rotated object detection represents the object as a rectangular box with a rotation angle.This representation is more suitable for object detection in remote sensing images as it can better separate different objects.In the background of the rapid development of deep learning,the application of deep learning for object detection has become a trend.Therefore,this thesis chooses to conduct a study on rotated object detection based on deep learning,which is applicable to remote sensing images.In this thesis,we firstly introduce and analyze the rotated object box representation,network for rotated object detection and related datasets from the practical significance of rotated object detection.After that,this thesis analyzes and summarizes the existing object detection networks and the network training strategies in terms of object detection algorithms using deep learning.To address the problems of complex and time-consuming training of anchor-based rotated object detection and time-consuming inference of twostage detection network,this thesis designs a one-stage anchor-free rotated object detection network as baseline with a dynamic label assignment strategy of "Optimal Transport Assignment" in the training stage and demonstrates the effectiveness of the network through experiments.To address the problems of omission detection and low accuracy of rotated box regression in the baseline,this thesis improves the network from four aspects: backbone network,multi-scale features,data enhancement,and loss function.Feature extraction is performed using a backbone network with stronger feature extraction capability and higher operational efficiency;more sophisticated feature fusion methods are used;and data enhancements such as Mix-up and Mosaic enhancement are used to improve network generalization.In this thesis,we propose a hybrid regression loss including IoU-loss to enhance the regression frame accuracy and use the activation function to limit the angular output range to avoid misguiding the network regression due to the periodicity of the angle.Through the above improved method,this thesis proposes an efficient rotating target detection network,which has certain advantages over existing methods in terms of accuracy and speed.Aiming at the complex and time-consuming problem of rotating frame NonMaximum Suppression（NMS）,a rotating target detection algorithm without NMS postprocessing is proposed in this thesis.The algorithm uses a two-stage training method to train the network: in the first stage,the basic model is trained by using one-to-many label allocation strategy,and in the second stage,some parameters of the model are fixed for one-to-one label allocation training,The detection model without NMS is obtained.Experiments show that this method greatly improves the running speed of the network at the expense of small accuracy.