Research On Building Detection Method Of Remote Sensing Image Based On Neural Network

With the continuous progress of China’s space science and technology,the research on various topics based on remote sensing technology is getting deeper and deeper.The target detection task based on remote sensing image has extremely important research value for aviation,navigation,ground exploration and other fields,and some researches have been successfully deployed to practical applications.Such as ocean ship inspection,urban and rural planning,uninhabited areas exploration,etc.The building is one of the important objects for human living activities.The detection task of the building in remote sensing image has always been a hot and difficult topic in the field of remote sensing.Based on the classical one-stage detection method Yolov3 network model,this paper makes lightweight improvement on it,and combines the features of high resolution remote sensing image,large data volume and small target scale to optimize it.In this paper,the structure of the Yolov3 network is improved firstly.The lightweight neural network Efficient Net is used to replace the original backbone network Darknet-53 for feature extraction,and the scale of the feature map is modified.A multi-scale void convolution module is introduced to expand the sensitivity field of the network.Then,the K-Means ++algorithm was used to conduct cluster analysis on the data set again,and the new Anchor accelerated model convergence was used to improve the detection accuracy.Then CIo U boundary regression loss function is used to replace Io U loss function.Experimental verification shows that the improved Yolov3 algorithm model is smaller with fewer parameters,and effectively retains the feature information of small targets,thus improving the missing detection of small targets.In summary,combining with the characteristics of buildings in remote sensing image,this paper studies the problems of large model,slow convergence and low detection accuracy in remote sensing image target detection domain.By improving the network structure of Yolov3,the effective detection of buildings in remote sensing images is realized.On the basis of preserving the detection speed and meeting the real-time requirement,the detection accuracy is improved compared with the original algorithm.