Precise Object Detection In High Resolution Optical Remote Sensing Images

Object detection is an important part of remote sensing image analysis,and it is also a key part of transforming remote sensing data into application.In recent years,Convolutional Neural Network(CNN)based deep learning has made a breakthrough in Computer Vision field and has shown great potential in object detection.However,in the precise object detection of optical remote sensing images,it still faces the following challenges: 1)Remote sensing image imaging platform has low spatial resolution and a wide field of view,resulting in objects’ small absolute size and relative size,and it is also an important and difficult problem to precisely detect small and tiny objects.2)The remote sensing image imaging platform observes ground objects from the aerial perspective,resulting in objects’ arbitrary orientation,and it is a difficult problem to precisely detect object in arbitrary orientation.3)Ground objects are acquired with Off-nadir imaging in most cases,resulting in the boundaries of many buildings’ footprint partly occluded by the building bodies,and it is also a challenging problem to precisely detect the building footprint in Off-nadir images.In order to solve the above problems,this research aims at precise detection of objects in high-resolution optical remote sensing image and is based on deep Convolutional Neural Network(CNN).This research is conducted from three typical problems: small object detection,oriented object detection,and building footprint detection.And indepth explorations are made in the aspects of dataset construction,object characteristics analysis,method model design and experimental results analysis.The research contents and main contributions are summarized as follows:(1)For small object detection,in order to solve the problem that the object scales of the existing remote sensing object detection datasets are generally too large,and it is difficult to use them to effectively evaluate the detection performance of small objects,a remote sensing small object dataset AI-TOD(Tiny Object Detection Dataset in Aerial Images)with an average size of only 12.8 × 12.8 pixels is established in the aspect of dataset.In the aspect of method,firstly,the reason for the low accuracy of the general object detectors when detecting small objects is analyzed,that is,the intersection over union metric is sensitive to the location error of the small objects.Secondly,the Normalized Wasserstein Distance(NWD)metric is proposed for anchor-based detectors,and applied to the module of assigning positive and negative samples,the module of non-maximum suppression and the design of loss function.The multiple center point design is proposed for anchor-free detectors.Extensive experiments on the AI-TOD dataset show that the proposed methods effectively improve the small object location performance of detectors.Compared with the baseline method,the accuracy of small object detection is increased from11.1% to 17.8% by adding NWD to multiple modules,which is relatively increased by 60.4%.Besides,the accuracy of small object detection is increased from 13.4% to14.5%(relatively increased by 8.2%)by using multiple center design in anchor-free detectors,the highest accuracy of small object detection is increased by 60.4%.(2)For object detection in arbitrary orientations,the problem of true value confusion and background pixel confusion in the representation of object in arbitrary orientations is firstly analyzed,and then Center Map OBB,which is an object representation method in arbitrary direction based on center probability map,is proposed.This method makes the center point value of oriented rectangle 1 and the edge value 0,and makes the value of other location decrease gradually from 1 to 0.This method effectively reduces the impact of background pixels on the network.Center Map OBB simultaneously ameliorates the problem of true value confusion and background pixel confusion,because its essence is to perform pixel-wise soft classification of the pixels in the bounding box and it is naturally unique.On this basis,an attention network guided by a weighted pseudo-segmentation map is further proposed based on Center Map OBB,to suppress the interference of complex background pixels in remote sensing images and generate instance-level features to enhance the features of objects in arbitrary orientations.Our proposed method achieves the accuracy of 76.03% on the most representative oriented object detection dataset DOTA,and achieves the SOTA when compared with the same period methods.(3)For the detection of the building footprint,BONAI(Buildings in Off-Nadir Aerial Images)which is a dataset of building footprint detection in Off-nadir remote sensing images,is established from the aspect of the dataset.And the dataset includes images from six Chinese cities,3,300 remote sensing images of 1024 × 1024 pixels,268,958 buildings with annotated roofs,footprints and offsets between footprints and roofs.In the aspect of the method,based on the prior information that most of the building’s roof and footprint shapes in urban scenes are almost the same,a fully supervised offset learning method trained with fully annotated data is proposed.Secondly,based on the characteristic that the shape and location of the footprint in the multi-temporal remote sensing image of the same building remains unchanged,a semi-supervised offset learning method trained with a large amount of unlabeled data is proposed to further improve the detection performance.Finally,based on the characteristic that the angular distribution of the offset between roof and footprint obeys the uniform distribution,a feature-level offset augmentation method is proposed,which significantly improves the detection accuracy without increasing the cost of training and annotation.A series of experiments on the BONAI dataset verify the effectiveness of the above methods in the precise detection task of the building footprint in the Off-nadir image.Compared with the traditional method,the detection accuracy of the building footprint has been significantly improved.