Object Detection In Optical Remote Sensing Images Beased On Deep Convolutional Neural Networks

Object detection in remote sensing images is one of the research hotspots in computer vision.It has been widely applied in the field of military surveillance,ocean monitoring,urban planning and disaster evaluation and rescue recently.In recent years,with the continuous development of commercial remote sensing technology and sensor technology,the spatial resolution of remote sensing images has gradually increased.As there is more recognizable detailed information in remote sensing images,multi-class object detection in remote sensing images becomes feasible.Remote sensing technology has covered multi-purpose and multiresolution applications,and gradually meets the needs for different application fields.Therefore,the research of object detection in remote sensing images has important theoretical value as well as practical merits.Thanks to the development of deep convolutional neural networks(DCNN),the research on generic object detection has made significant progress.The development of generic object detection has also guided the research for object detection in remote sensing images.Since remote sensing images have complex backgrounds,large variations in object scale and aspect ratio,arbitrary orientation of objects and non-uniform spatial distribution,the traditional object detection methods face performance bottlenecks when processing detection tasks in remote sensing images.The existing research work on object detection in remote sensing images still has some limitations.Detection methods based on shallow learning models can only detect a few types of objects in remote sensing images and the detection results are not robust.Generic object detection methods based on DCNN cannot handle the object particularity of multiple scales,multiple aspect ratios,multiple orientations and clustering in remote sensing images.Based on the existing generic object detection methods and remote sensing technologies,we analyze and then summarize the limitations of the existing work and propose some improvements to address the detection of objects at multi-scales and multi-orientations in remote sensing images.The proposed improvements are in two-folds:1.Aiming at the difficulty of multi-scale object detection in remote sensing images,we first analyze the pros and cons of the existing multi-scale object detection methods.Then,following the idea of path aggregation,a bottom-up branch structure is introduced based on the Feature Pyramid Network(FPN)to improve the feature representation capabilities of multiscale objects.Furthermore,a remote sensing object detection network based on path aggregated feature pyramid(PA-FPN)is designed to enhance the detection ability of multi-scale remote sensing objects under complex backgrounds.PA-FPN has three important components,namely the robust multi-scale feature extraction sub-network,the adaptive feature fusion module and the more accurate ROI Align pooling module.PA-FPN can accurately extract the multi-scale features of objects in remote sensing images,and has better robustness to multi-scale objects.Therefore,it can accurately localize and classify multi-scale objects in remote sensing images.Experiments on DOTA dataset and NWPU＿VHR-10 dataset show that our proposed method has higher detection performance on multi-scale object detection task in remote sensing images.2.Aiming at the problem of detecting rotated objects in remote sensing images,we first analyze different representation methods of the rotated bounding box,and then design a detection method for rotated objects using generic object detection framework.Our proposed method improves the model’s ability to detect rotated objects.In order to improve the detection efficiency for rotated objects,we improve the traditional object detection pipeline and develop the rotated object detection network-RRDet.Our proposed RRDet includes the modules of generating rotated anchor boxes,prediction of rotation candidate regions,RROI pooling,inclined box Io U,and inclined boxes NMS.Compared with object detection framework based on horizontal bounding boxes,our proposed method has a stronger ability to detect rotated objects,and solves the problem of missing detection of dense rotated objects caused by horizontal bounding boxes.Experiments of multi-level detection tasks on HRSC2016 dataset reveal that RRDet achieves higher detection accuracy for rotated objects in remote sensing images compared with other methods.