Research On High Resolution Remote Sensing Object Detection Based On Anchor-free Method

Object detection on high-resolution remote sensing images are widely used in military and civilian fields,and are playing an increasingly important role.Remote sensing object detection methods based on deep learning mainly include anchor-based and anchor-free methods.Anchor-based methods using preset anchors to cover the region proposals as much as possible,but there are some limitations: First,because default anchors need to be specially preset for unique object distribution of the dataset,this kind of method is very hard to be migrated to different datasets.Second,due to the limited size and aspect ratio of the default anchors,the solution space of the bounding box will be relatively limited thus the object with extreme aspect ratio cannot be well detected.Third,since the anchors need to cover the potential region on the image as much as possible,a large number of anchors will be generated,which will consume a lot of time in the detection process.The anchor-free methods do not need to set the default anchors of a fixed size,removes the limitation on the aspect ratio,and do not need to generate a large number of anchors before detection.Compared with anchor-based detector,it is more flexible and relatively faster.Therefore,in recent years many deep learning models for anchor-free detector have emerged.Nevertheless,the existing anchor-free methods have some shortcomings in the following three aspects when processing remote sensing images: First,because the distribution of some small objects in remote sensing images is relatively dense,the distance between these targets is further reduced after the original image is down-sampled.Because of weakened distinguishability,it is easy to detect multiple objects as one object.Second,in the detection of oriented bounding box of remote sensing objects,anchor-free methods mostly use key points and rotation angles to predict,and slight offset of the rotation angle may cause a large deviation of the bounding box,finally resulting in a decrease in the m AP(mean average precision).Third,the existing anchor-free methods are mostly one-stage detection,which only rely on the confidence of the key-points on the heat map to locate the target,and there is no combination of semantic features in Ro I(Regions of Interest)to distinguish foreground and background,so the false alarm rate is higher than the two-stage detection methods.Aiming at these problems,this paper designs three improved anchor-free detectors.The main research contents are as follows:Aiming at the misdetections caused by the dense distribution of some objects in remote sensing images,a remote sensing anchor-free detector(Grouped Dense Sampling Detector,GDSDet)based on grouped dense sampling is designed.In the center-point heat map branch,a grouped dense sampling strategy is introduced to increase the number of candidates for the object center and at the same time enhance the category confidence of the center point.On this basis,an inter-class attention mechanism is added to the center heat map after grouped dense sampling to overcome the imbalance of different classes of objects.By comparing our detector with a variety of currently popular anchor-based and anchor-free detection models on UCAS-AOD and NWPU VHR-10,the experimental results show that the m AP of our method is 0.5～8.4% higher than those of the other methods.There is an improvement of3.4～35.8 FPS in the detection speed.Aiming at the problem that rotated bounding box is sensitive to small deviation of rotation angle in oriented box target detection,an oriented anchor-free detector(Corner-Vector Detector,CVDet)based on four corner vectors is designed.Firstly,the vectors from the center point to the four corner points is used to define an oriented bounding box so as to overcome the sensitive problem of rotation angle offset.Secondly,the orientation branch is used to select objects in the non-horizontal direction,so that the objects in the horizontal direction can be detected by the horizontal bounding box so as to further improve the m AP.On the DOTA-v1.0 and HRSC2016 datasets,our detector is compared with the currently popular anchor-based and anchor-free detection models.The experimental results show that the m AP of our method on the DOTA-v1.0 is improved by 1.3～63.3% compared with the other models.The m AP of our method on the HRSC2016 is improved by 1.4～31.9%compared with the other models.Aiming at the high false alarm rate of the one-stage anchor-free detection methods,a twostage anchor-free remote sensing detector(Corner Regions of Interest Detector,CRDet)based on candidate region recommendation is designed.First,the candidate boxes are generated by extracting the corner points of the bounding box,and the candidate boxes are initially filtered by the foreground classifier in the first stage to obtain the foreground candidate boxes with higher confidence so as to reduce false alarms.Then in the second stage,the secondary classifier finally achieves the precise category of the object.Due to the filtering of foreground candidate boxes in the first stage and the fusion of the semantic features of Ro I in the second stage,CRDet can obtain candidate boxes with a higher foreground probability,richer semantic features and a balanced ratio of positive and negative samples before the final classification,thereby reducing the false alarm rate.Comparison experiments were carried out on the NWPU VHR-10 and UCAS-AOD datasets with the currently popular anchor-based and anchor-free detection models.The experimental results show that the detection m AP of CRDet in the UCAS-AOD is increased by 0.4～15.2%compared with the other models.The detection speed is 5.5～12.0 FPS higher than that of other two-stage models.The m AP of CRDet in the NWPU VHR-10 is 1.2～7.8% higher than that of the other models.The detection speed is 4.5～14.4 FPS higher than that of other twostage models.