| With the development of space technology and remote sensing technology,remote sensing image interpretation plays an important role in promoting the development of missile warning,city planning,land use,sea situation monitoring and so on.As a key component of remote sensing image interpretation,remote sensing image target detection technology has become a research hotspot.However,remote sensing images usually contain rich background information,and the types and quantities of targets are large,and the angles of targets are changeable.So far,the multi-scale and multi-angle target detection technology is still lacking.Therefore,how to quickly and accurately detect multi-scale objects in remote sensing image under complex background becomes the key to remote sensing image interpretation.In order to solve the above problems,this paper studies the improved remote sensing image target detection method based on deep learning technology,and proposes a remote sensing image target detection method with target feature enhancement and rotation Angle guidance based on multi-scale feature extraction.The main research contents are as follows.Firstly,based on YOLOX improved algorithm(YOLOX-OD),this paper proposes the design of object association module(ORM)to improve the difficulty of detecting small and medium-sized objects in remote sensing images.Aiming at multi-scale channel features,this module extracts the target features with large receptive field and small receptive field in the corresponding feature layer in parallel.Then,multi-head self-attention mechanism is used to learn the correlation degree between small receptive field features and global contour information,and the correlation degree between large receptive field features and global detail information respectively.The weighted average is used to balance the influence of two kinds of relational weights on the target features,and finally the enhanced target features are obtained.In order to further optimize the detection effect of remote sensing image target detector,a mutual attention mechanism(DHAI)was proposed for the non-interaction of the decoupling task of detection head.By calculating the weight matrix of the shared feature layer of two subtasks in the space and channel direction,the weight matrix of the two subtasks was weighted respectively,and the detection effect was improved.Secondly,to solve the problem of variable target Angle,a rotating detector(YOLOX-ODK)was proposed in this paper based on YOLOXOD.Angle parameters were added to the detection head of the neural network.Relative divergence was used as the loss function of the regression part of the model.In this paper,a multi-scale target detection model of remote sensing images based on feature enhancement,YOLOX-OD,and a multi-angle target detection model of remote sensing images based on rotation guidance,are proposed to improve the capture ability and detection performance of remote sensing image target detector on target features.Experiments show that the YOLOX-OD model reaches 69.60% average accuracy on the DOTA level detection data set.The YOLOX-ODK model achieves an average mean accuracy of 72.68% on the DOTA rotation detection data set.Compared with YOLOX,the YOLOX-OD model improved the small target detection performance by 4.9% and the overall detection performance by 4.3%.In addition,the performance of YOLOX-ODK is further improved by 3.08% on the basis of YOLOX-OD,realizing more accurate remote sensing target detection. |
