A Study On Remote Sensing Image Object Detection Based On Atrous Feature Pyramid Network

With the development of optical technology,information technology and computer technology,remote sensing satellite technology has made remarkable progress in recent years.Among them,by identifying and locating specific targets in remote sensing images,object detection tasks in remote sensing images has broad application prospects in urban planning,environmental management and military applications.Different from object detection in natural images,the background of optical remote sensing images is complex and the object instances in images change greatly in scale,direction and shape.Therefore,the detection of multi-scale targets and small targets is a challenging problem for object detection tasks in optical remote sensing images.In order to solve the above problems,this paper studies how to improve the detection accuracy of multi-scale targets in optical remote sensing images,and proposes a object detection method suitable for remote sensing images.Aiming at the problem of multi-scale targets and small targets detection in remote sensing images,Atrous Feature Pyramid Network(AFPN)is proposed based on the two-stage object detection method.The AFPN network improves the detection accuracy of multi-scale targets by constructing a feature pyramid network with powerful feature expression capabilities.In the AFPN network,Atrous Feature Fusing Module(AFFM)is proposed,which can merge deep features with rich semantic information and shallow features with high spatial resolution for many times,and significantly enhances the multi-scale expression capability of network.In the AFFM module,Atrous Lateral Connection Block(ALCB)is proposed,which is a multi-branch atrous convolution module.The ALCB module can capture multi-scale context information while expanding the receptive field of feature maps,so that it can make full use of image feature information.Finally,the multi-level features generated by the AFFM will be fed to the region proposal network and detection head respectively for predicting the detection results.The proposed network can not only detect multi-scale targets more accurately in optical remote sensing images,but also improve the detection accuracy in natural images.