Research Of Rotating Target Detection Algorithm In Remote Sensing Image Based On Group Equivariant Convolution Neural Network

In recent years,with the increasing convenience of remote sensing image acquisition,the use of remote sensing image scene understanding to quickly obtain ground information has been widely used in many fields such as military,electric power,transportation,disaster rescue,agriculture,urban planning,etc.On the one hand,remote sensing image scene understanding requires a large amount of human cost.On the other hand,deep learning,represented by target detection,has achieved great success in natural image understanding.Using deep learning to understand the information of remote sensing image scenes has become a hot research topic.Among them,rotating target detection of remote sensing images is an important branch of this research direction.Unlike the targets in natural scenes,the targets in remote sensing images have characteristics such as drastic scale changes,large aspect ratios,and variable directions.Therefore,there are two problems in the detection of rotating targets in remote sensing images: the difficulty of target feature extraction and the difficulty of Io U loss function design for rotating target frames.To address these two problems,this thesis investigates three aspects,mainly including.（1）For the characteristics of the above remote sensing image targets,this thesis designs a baseline network based on the anchorless framework.For the acquisition of target localization information,a vector-based rotating frame labeling method combined with smoothed L1 loss is used to effectively improve the target localization performance of the baseline network.（2）To address the above problem of difficult extraction of rotating target features,the E（2）group equal-difference convolution method is used on the basis of the baseline network to make the convolutional neural network achieve rotational equal-difference and alleviate the difficulty of rotating target feature extraction.First,the residual backbone network Re Res Net101 based on E（2）group equivariant convolution is established;then,Re Res Net101 is proved to have excellent rotational equivariance by comparing the similarity of output vectors of different convolutional neural networks under the case of multi-angle input of the same image.The ablation experimental results show that the baseline network m AP@0.5 of Re Res Net101 is improved by 20%,thus proving that Re Res Net101 has excellent feature extraction ability.（3）To address the above problem that the Io U loss function of rotating target frames is difficult to design,based on the baseline network,in order to approximate the Io U loss of arbitrary quadrilateral,the localization loss function of rotating target frames is split into the Io U loss of external horizontal frames and the internal Io U loss in this thesis,and the distance-Io U loss of external horizontal frames is adopted,while the Io U loss of internal quadrilateral is designed by the Gliding＿Vertex＿Io U Loss is implemented.It is experimentally demonstrated that the Distance-Io U Loss and Gliding＿Vertex＿Io U Loss proposed in this thesis improve the baseline network m AP@0.5 by 15.46%.