Research On Oriented Object Detection Method On The Visible Optical Remote Sensing Image

With the continuous exploration and practice of human beings,in the past decades,the related research of remote sensing image has been further mature with the development of space technology.As a classical remote sensing research field,object detection in visible remote sensing images plays an important role in civil and military fields,such as urban planning,population flow and military strike.In recent years,with the emergence of deep learning technology,general object detection technology has developed rapidly.However,due to the characteristics of large scale variation and variable direction of objects in visible remote sensing images,the application of general object detection methods in visible remote sensing images has become a challenging topic.Therefore,it is of great value and significance to study the rapid and accurate object detection technology of visible light remote sensing images,both for the national defense field and the improvement of people’s livelihood.Therefore,this paper takes oriented object detection,which is a difficult point in the field of remote sensing object detection,as the research background,based on deep learning technology,and aiming at the problems and difficulties of oriented object detection in the field of visible remote sensing image,proposes a variety of real-time remote sensing object detection methods.The main research contents of this paper are as follows:Aiming at the problem that remote sensing oriented object detection algorithm is generally inefficient and difficult to achieve real-time while ensuring high accuracy,this paper proposes a real-time oriented object detection algorithm based on slenderness ratio information.In this method,a low complexity backbone is used to improve the real-time performance of the algorithm,and deformable convolution and transposed convolution are used to compensate for the accuracy reduction caused by the decrease of backbone complexity.The central region of Gaussian distribution is used as the anchor point of object localization,which balances the number of positive and negative samples and enhances the stability of object classification.In order to improve the prediction accuracy of the object’s long side,a prediction method of the slenderness ratio of objects is proposed,and the detection accuracy of some categories is improved by combining the short side and the slenderness ratio instead of the long side which is difficult to predict.The proposed method is tested on two classical remote sensing datasets: DOTA and HRSC2016,and achieves good performance in both detection efficiency and detection accuracy.Aiming at the problem of critical ambiguity caused by periodicity in the angle representation of oriented objects,a dual-angle oriented object detection algorithm guided by geometric features is proposed in this paper.The method reduces the influence of the inaccurate angle prediction on the detection result and improves the stability of the object direction prediction by the form of dual-angle prediction.In order to solve the problem that the intersection area of prediction box and truth box decreases sharply due to the unrestricted direction prediction in the process of double angle regression.A selfsupervised guidance loss based on geometric features is proposed,which establishes the coupling relationship between two angles and side lengths through geometric relations,guides the direction of Angle regression,and ensures that the dual-angle representation can stably improve the detection accuracy.The proposed method is validated on DOTA and HRSC2016 datasets,and compared with other mainstream methods on more datasets.The results show that the proposed method can improve the accuracy obviously and has better performance than other mainstream methods in precision and speed.Aiming at the problem of endpoint alignment caused by corner order and matching relationship in the endpoint representation of oriented objects,a oriented object detection algorithm based on direction vector is proposed in this paper.In order to solve the loss fluctuation caused by the sequence of endpoints,the method uses two-dimensional graph to predict endpoints.In order to solve the problem of difficult matching between object and endpoint caused by excessive number of endpoints,a new guide point selection rule and prediction architecture are proposed in this method,and the corresponding matching method between center point and guide point is proposed.In this method,the direction vector is constructed by the center point and the guide point to represent the object.Each center point only needs to match one guide point,which reduces the matching difficulty and improves the detection accuracy.In this paper,the effectiveness is verified on two datasets,and the parameters in the object decoding method are compared and the optimal solution is obtained.Compared with other algorithms on several datasets,the proposed method is proved to be superior in speed and accuracy.Aiming at the problem that it is difficult to detect small objects in remote sensing images,and the detection algorithm is difficult to give consideration to both real-time and accuracy in small objects detection,this paper proposes a oriented object detection algorithm fused with supervised attentional map.By fusing the feature map with the saliency map of the supervised object,the feature of the object is strengthened and the detection accuracy of the small object is improved.In addition,for the direction vector based oriented object detection algorithm with center point and guide point as the core,a center feature enhancement method based on Gaussian center-line attentional map is proposed to improve the positioning accuracy of center point and guide point.The results on multiple datasets demonstrate the effectiveness of the two methods,especially for small targets.The comparison with other similar methods also shows that the two enhanced methods can reach the level of the highest accuracy algorithm on the premise of ensuring real-time.To sum up,this paper,according to the theory of visible remote sensing image has the object detection problem study,proposes the real-time oriented object detection algorithm based on slenderness ratio information fusion,the dual-angle oriented object detection algorithm based on geometric feature guidance,the oriented object detection algorithm based on direction vector,the oriented object detection algorithm fused with supervised attentional map.The research in this paper has enlightening and important theoretical significance and application value for oriented object detection in visible remote sensing images.