Multi-scale Remote Sensing Object Detection Based On Attention Mechanism

Object detection in remote sensing images has important application value in various fields such as weather forecasting,ocean monitoring,environmental monitoring,disaster monitoring and evaluation,and military.However,there are many objects in remote sensing images with various scales,and it is difficult to extract the characteristics of objects at multiple scales at the same time.In addition,the distribution of object locations in remote sensing images is usually dense and difficult to locate.The method based on deep learning has become the mainstream in the field of object detection.Among them,the Faster R-CNN network performs better,but the network does not perform well in the above remote sensing scenes.This paper proposes the MRSAM Net based on the Faster R-CNN network.Aiming at the problem of information loss caused by the down-sampling operation of the original residual module in the network,this paper improves the original residual module so that the network can extract more objects Feature information to improve the utilization of remote sensing image information;In response to the problem of multiple scales of remote sensing images,it is difficult to extract features of multiple scale objects at the same time,the attention mechanism and feature pyramid structure are introduced to build a multi-scale dual channel with an attention mechanism.The feature extraction network combines deep features with shallow features,and improves the detection effect of remote sensing images without basically increasing the amount of calculation of the original model;In view of the dense distribution of object positions in remote sensing images,they are usually misaligned during positioning and difficult to carry out subsequent processing.The introduction of Transformer to convert the horizontal detection frame into a rotating detection frame can better locate densely distributed objects.This paper conducts experimental verification on the DOTA data set,and the results show that the MRSAM Net has improved the detection performance of multiple scale targets,and can better deal with densely distributed targets.The m AP of the MRSAM Net on this data set is 75.12%,which achieves good results.