Research On Rotating Object Detection Algorithm Of Remote Sensing Image Based On Key Points

With the rapid development of remote sensing technology,remote sensing images contain more and more information.High-resolution remote sensing images contain information about urban buildings,urban transportation,natural resources and other information.The detection and recognition of remote sensing image objects is the basis of remote sensing information processing.Compared with traditional object detection algorithms,object detection based on deep learning has become the mainstream object detection algorithm because of its high detection accuracy and fast detection speed.The research on object detection in natural images has achieved considerable results.However,compared with natural images,the objects in remote sensing images have the characteristics of close arrangement,large scale change and arbitrary direction.These characteristics require the object detection network to have high performance.Therefore,this poses a new challenge to the object detection of remote sensing images.Based on the Centernet network of key point detection,this paper studies the object detection of remote sensing images through attention model,feature fusion module and the expression of rotating objects.The main research contents of this paper are as follows:（1）Due to the great differences between natural images and remote sensing images,according to the characteristics of objects in remote sensing images,this paper proposes an R-Centernet remote sensing image object detection algorithm.On the problem of object directionality,the angle information providing the rotation change of remote sensing object is added.Because the remote sensing object has scale change,the multi-scale feature fusion module is added to the network to fuse the multi-scale pooling information of remote sensing object and attention feature information.The experimental results show that R-Centernet model has a good effect in remote sensing image object detection.（2）This paper analyzes the limitations of using angle information to process rotating objects in remote sensing images.Based on the BBAvectors method,a rotating object detection algorithm guided by Spatial-Coordinate Attention module and multi-path residual block is proposed in this paper.Firstly,modify the Coordinate Attention module through multi branches,and then fuse with the Spatial Attention module to generate the Spatial-Coordinate Attention module,the module integrates a variety of attention feature information and can learn and fuse the feature information of the object from different aspects.Secondly,expand the original mapping relationship in the Residual Block and design a more comprehensive information transmission and sharing structure of multi-path residual block,A new feature fusion module is introduced in the sampling process on the network,which reduces the loss of information and improves the utilization of feature information.The experimental data show that the m AP value of our network is improved on a variety of remote sensing image datasets,and can maintain good real-time performance.（3）Aiming at the problem that BBAVectors method uses more parameters in determining the directivity of remote sensing objects,a rotating object detection method based on coordinate system projection is proposed in this paper.This method uses coordinate system projection to determine the rotation boundary box of remote sensing object.Firstly,the remote sensing image object is placed in the same coordinate system,and the four vertices of the object are projected to the x-axis and y-axis in the coordinate system to obtain the projection vector of the object based on the coordinate system.This method requires less parameters,and the parameters of network learning can be directly used to describe the rotation frame of remote sensing object.Through DOTAv1.0 and HRSC2016 remote sensing datasets,the coordinate system projection method shows better performance in accuracy and speed.