| Remote sensing image object detection aims at locating the position of the objects of interest in the remote sensing image through the bounding box.Accurately and quickly detecting the objects of interest in remote sensing images is one of the basic steps for subsequent analysis and utilization of remote sensing data;It is widely used in major fields such as urban planning,geographic information acquisition,military defense and so on.With the continuous development of remote sensing imaging technology,the resolution of remote sensing images becomes very high,highresolution remote sensing images contain richer detailed information and structural features.These new changes pose new challenges to traditional object detection tasks.Traditional remote sensing image object detection algorithms mainly rely on feature description operators set based on empirical rules.This type of operator needs to manually set some empirical parameters for objects with different characteristics,and when faced with massive amounts of remote sensing image data,both accuracy and robustness are low.At present,object detection methods based on deep neural networks represented by R-CNN series algorithms and YOLO series algorithms are the mainstream of research.But these algorithms are mainly for object detection tasks in conventional images.The direct application of these algorithms to remote sensing image object detection has achieved unsatisfactory results.The reason is that compared with ordinary images,the objects in remote sensing images show arbitrariness in direction and large differences in scales.Therefore,according to the characteristics of remote sensing images,this article conducts related research on how to improve the accuracy of oriented object detection in remote sensing images.The specific work is as follows.First,we designed a two-stage multi-category rotation detector for rotated objects.This method first uses semantic features to guide the anchoring,predicts the locations where the center of objects of interest are likely to exist as well as the scales and aspect ratios at different locations,thus,get the horizontal high-quality Anchor as proposal.Secondly,the Horizontal Region of Interest（HRo I）is converted into a Rotated Region of Interest（RRo I）through the RRo I learner.At last,the rotation-invariant features are extracted from RRo I for boosting subsequent classification and regression tasks.We conducted Extensive experiments on remote sensing public datasets DOTA,and the detection accuracy reached 78.17（m AP）,which was 5.1 points higher than the benchmark.Our method performed more prominently in the detection performance of extreme-shaped objects.Second,on the basis of the previous method,we designed a multi-stage rotation detector for rotated objects combined with an attention mechanism.The multi-stage detector consists of three detection branches with Io U thresholds of 0.5,0.6,and 0.7,and the output of a detector is a good distribution for training the next higher quality detector.The resampling of progressively improved hypotheses guarantees that all detectors have a positive set of examples,reducing the overfitting problem.The same cascading process is applied at inference,so that the quality of the candidate area increases with the increasing Io U threshold of the detection branch,and the proposal with higher quality can match the detection branch with a higher Io U threshold.Experiments show that the detection accuracy of our method has been further improved on DOTA,reaching 79.98（mAP）. |
