Research On Oriented Object Detection Method In Remote Sensing Image

With the development of remote sensing information technology,remote sensing image object detection has a wider application space in the fields of geological monitoring,military reconnaissance,urban planning and post-disaster investigation,and has also become one of the important research branches of computer vision.Due to the increasing improvement of information collection technology,the spatial resolution of remote sensing images is increasing,and the detailed texture features are more abundant,which can meet the application needs of remote sensing targets in many fields.Therefore,the application of object detection to remote sensing images has sufficient application value and research significance.Due to the complex background,various categories,multi-scale and multi-angle characteristics of remote sensing images,the traditional object detection technology does not have good generalization and robustness,and also cannot meet the current detection performance requirements.The object detection technology has been further improved after combining with the current development of deep learning,and has a more accurate detection effect on dense multi-scale targets and rotating targets.Based on the target detection based on the deep convolutional neural network,this paper designs a scheme for the difficulties of remote sensing images such as multi-category,multi-scale,and arbitrary orientation,and completes the performance improvement of remote sensing target detection in terms of accuracy and real-time performance.This paper proposes a remote sensing rotating object detection network based on multi-level feature selection(MFS-Det)and an anchor-free remote sensing rotating object detection algorithm based on center point prediction(CPDet).The main work of this paper includes:(1)In response to the challenges of various scales and arbitrary directions of remote sensing images,this paper proposes a remote sensing rotating object detection network based on multi-level feature selection(MFS-Det).By introducing the PA-FPN module to aggregate the feature information of high and low dimensions,the detection effect of the network on multi-scale targets is improved.Through the feature selection module,the Ro I output by the RPN network is mapped to the most suitable feature level.In order to solve the problem of inaccurate positioning accuracy caused by two quantizations of Ro I Pooling,the Ro I Align module of bilinear interpolation is introduced.In this paper,various boundary problems of rotating box representation are analyzed,and a new rotating box representation,the sliding vertex method,is introduced,and a targeted loss function is designed to solve the boundary problem of rotating boxes.On the datasets DOTA and HRSC2016,a variety of cutting-edge algorithms were compared to verify the detection performance of this network model and achieved good results.(2)In order to deal with the problem of reducing the model inference speed caused by the flooding of anchor parameters,this paper proposes an anchor-free remote sensing rotating object detection algorithm based on center point prediction(CPDet).By generating a heat map,CPDet leads to four prediction branches: the coordinates of the prediction center point and its offset,width and height,and rotation angle.In order to better predict the position of the center point,the method of shrinking Gaussian map is adopted,and the sampling is concentrated in the center ellipse area.By introducing the circular smooth label method,the angle regression problem is optimized into a classification problem,thereby solving the boundary problem of angle regression.By increasing the weight deviation function,the detection effect of square-like objects is optimized.While ensuring the accuracy of network detection,the real-time performance of model detection is improved,and the effect is verified on the datasets DOTA and HRSC2016.