Deep Learning Algorithm For Oriented Object Detection In Optical Remote Sensing Image

Object dectection in optical remote sensing images is widely used in military reconnaissance,traffic monitoring,search and rescue.In remote sensing images,the axis direction of objects such as aircrafts,ships and vehicles is arbitrary,and there are characteristics of large aspect ratio and dense arrangement of objects.Horizontal bounding box can introduce more background information when detecting objects,which makes it difficult to reflect the real position of the objects,while the oriented bounding box can locate the objects more accurately.In practical applications,oriented object detection in optical remote sensing images needs to meet the requirements of both speed and accuracy.Therefore,based on oriented bounding box,this paper carried out the following research for the problems of slow detection speed and missed detection of objects.Aiming at the shortcomings of the existing oriented bounding box representation methods,this paper proposes a novel representation method based on the Point-Boundary distance Vectors（P-BV）and IOU Angle loss function.P-BV takes a point inside the object as the origin of coordinates to establish Cartesian coordinate system,and uses the vertical vectors from this point to the four sides of the object bounding box to represent the target.The IOU Angle loss uses the intersection of union between the ground truth box and the bounding box to reflect the prediction deviation of the angle.This representation method can adapt to the change of the objects’ aspect ratio more flexibly,solve the problems of discontinuous angle loss and inconsistent parameter regression,and speed up the convergence of the model..To address the problems of low accuracy for horizontal object detection and slow speed in existing oriented object detection algorithms,this paper proposes a fast object detection algorithm（P-BV+ algorithm）,which is based on the Point-Boundary distance Vectors and is anchor-free.In this method,feature maps of different scales are fused by up-sampling to obtain a single high-resolution feature map,on which categorys,distance vectors and centerness are directly predicted intensively.On this basis,the horizontal bounding box prediction branch is added to improve the detection accuracy of the horizontal object.The experimental results show that the method proposed in this paper can realize the fast object detection.To address the problems of miss detection and inaccurate location in current single-stage oriented object detection algorithms when the objects are densely distributed.The P-BV++ algorithm adds a multi-scale spatial attention mechanism and a feature refinement module on the basis of the P-BV+ algorithm.The multi-scale spatial attention mechanism generates a spatial attention heat map that distinguishes objects and backgrounds on a densely connected multi-scale feature map to enhance target features and suppress background interference.The feature refinement module uses a coarse-to-fine cascade regression strategy to generate more accurate localization results by using the initial predicted directed box location information to guide the deformable convolution to aggregate context and boundary information around the target and recode the classification and regression features.Experimental results on multiple datasets demonstrate the improved detection performance of the P-BV++ algorithm proposed in this paper.