融合背景信息和方向信息的遥感图像旋转目标检测OA

Aimingat the problem of insufficient detection accuracy caused by large differences in target scale and complex direction distribution in remote sensing image target detection,a detection network com-bining background information and direction information was proposed.Firstly,aimed at the problem of large differences in target scale,a Receptive Field Extending(RFE)module was designed.Different from the traditional fixed receptive field or complex multi-branch structure,this module integrated multi-scale background information through large kernel decomposition,dilated convolution,and parallel branch struc-ture design,and solved the problem of different background information requirements for different scale targets without significantly increasing the amount of calculation.Secondly,aimed at the problem of com-plex distribution of target directions,an Orientation Aware Cross Attention(OACA)module was de-signed.Different from the existing attention mechanism convolution kernel shape,the module extracted di-rectional texture information through horizontal and vertical separable convolution to prevent feature loss and fracture;at the same time,a cross-attention mechanism was designed to suppress background noise and enhance the interactivity of directional information.The experimental results show that the detection accuracy of the proposed method on DOTA,HRSC2016 and DIOR-R datasets reaches 76.88%,98.43%and 65.06%,respectively,which is 1.01%,0.83%and 0.76%higher than that of the Orient-ed R-CNN method,respectively,which further verifies the effectiveness of background information and di-rection information collaboration.