| With the development of the aviation industry and the advancement of observation technology,aerial remote sensing images are more and more widely used in military and life,and remote sensing image target detection is one of its applications with important research value,which has received extensive attention from scholars at home and abroad.The traditional target detection method detects images from a horizontal perspective,while aerial images have the characteristics of arbitrary orientation,complex background,large scale changes and dense distribution of target instances compared with traditional target detection images due to the special bird’s-eye imaging perspective.If the traditional target detection method is directly used for aerial remote sensing images,there will be problems that the detection results do not match and the detection accuracy is low.This paper uses BBAVectors as the basic method of remote sensing target detection,and improves and optimizes the problems existing in BBAVectors in remote sensing target detection.The main work is divided into the following two points:(1)Aiming at the problem of unbalanced positive and negative samples in heatmap regression,a heatmap regression loss function combined with confidence is proposed,and the vector loss function is optimized to improve the quality of vector generation.BBAVectors locates the center point of the target by heat map regression,and then returns the boundary vector of the bounding box at the center point to complete remote sensing target detection.In view of the insufficiency of the BBAVectors loss function,we optimize from two aspects.First,in heatmap regression,in order to alleviate the imbalance of positive and negative samples and improve the probability that the heatmap predicts key points at positive samples,a heatmap regression function combined with confidence(With Confidence Focal Loss,WCFL)is proposed.The classification and regression are combined to alleviate the problem of unbalanced positive and negative samples in the thermal layer and improve the accuracy of model positioning;secondly,in the process of boundary vector regression,the predicted boundary vector of BBAVectors is not vertical.For this problem,this paper introduces the vertical condition of the vector into the vector regression function to promote the verticality of the predicted vector,improve the quality of the vector generation,and thus improve the detection accuracy.Through experiments and analysis,the effectiveness of the improvement of the loss function in this paper is verified.(2)Introduce position attention module and channel attention module into BBAVectors,and propose a dual-Attention for Object Vectors(DAOV)model based on attention mechanism.Remote sensing images have the characteristics of complex picture background,large variation of target scale and dense target distribution.In order to detect remote sensing targets more accurately,the model needs to obtain more comprehensive and effective image feature information.In order to further improve the feature extraction ability of the model,based on the BBAVectors method,this paper introduces the position attention module and the channel attention module,and proposes a remote sensing target detection model DAOV based on the attention mechanism.Establish global dependencies from different dimensions to obtain more comprehensive contextual information.Through experimental analysis,the effectiveness of the DAOV model is verified.In order to verify the improvement of the loss function of BBAVectors and the effectiveness of the proposed remote sensing target detection model based on the attention mechanism,this paper uses the DOTA dataset and HRSC2016 dataset for training and verification,and compares with the detection accuracy of BBAVectors to verify the improvement.After experimental analysis,the methods proposed in this paper are all effective. |
