Research On Building Detection Algorithms For High Resolution Remote Sensing Images Based On Deep Learning

With the development of Aerospace Technology,remote sensing satellite image resources are widely used in many fields such as urban planning,military investigation,and land resources.High-resolution remote sensing image target detection has become an important research direction in the field of remote sensing image processing.Buildings are an important landmark feature of urban or suburban grounds,so the detection and identification of buildings in the field of remote sensing has become an important issue.The feature representation of the traditional target detection method is designed based on experience,which is time consuming and cannot explore the feature information of the higher dimension of the image,and the image processing based on deep learning can quickly and effectively learn the image feature.At present,Faster R-CNN is the mainstream algorithm of target detection based on deep learning.Therefore,this paper applies Faster R-CNN algorithm to building detection tasks.The algorithm achieves detection accuracy of 53.7% and 51.7% on the feature set of Google and Gaofen2 images respectively.In view of the low detection accuracy of buildings,this paper improves the detection accuracy from the following three aspects.Firstly,this paper uses the Online Hard Example Mining(OHEM)to improve the Faster R-CNN algorithm,and proposes the OHEM+Faster R-CNN algorithm.OHEM is a kind of screening and retraining difficult samples that are difficult to train.Algorithm,this improvement makes the building detection accuracy increase to 58.2% and 63.3% in the feature set of Google and Gaofen2 image buildings respectively.Secondly,in order to further improve the building detection accuracy,the VGG16 network in the Faster R-CNN algorithm is replaced by the residual network ResNet based on the OHEM algorithm.ResNet solves the problem of performance degradation after the depth of the network model becomes deeper.The content is to add an identity mapping layer on the basis of shallow network.This improvement makes the building detection accuracy increase to 60.2% and 65.6% in the image set of Google and Gaofen2 image respectively.Finally,this paper uses the R-FCN(Region-based Fully Convolutional Networks)algorithm to apply to remote sensing image building detection.R-FCN designed a position sensitive score map to encode position information,which can further improve the detection accuracy.The OHEM+R-FCN algorithm achieved 74.1% and 78.5% detection accuracy in the feature set of Google and Gaofen2 images respectively.