Building Extraction From Remote Sensing Images Combining Saliency And Deep Learning

In recent years,with the continuous development of remote sensing technology,the difficulty of acquiring remote sensing images is gradually reduced,the resolution of images is constantly improving,and the scope of application is constantly expanding,including land change detection,urban data updating,disaster prevention and emergency response and other aspects.The images processed by computer are all digital images.As one of the most prominent features of remote sensing images,buildings have attracted wide attention.The traditional way of building extraction is labeling manually,which consumes a lot of manpower,financial resources and time.Therefore,automatic building extraction has always been a difficult problem in scientific research.Saliency detection algorithm is mainly based on the principle of visual attention mechanism to quickly capture salient areas in images.The automatic extraction of salient areas in images is bottom-up visual attention model.It is a data-driven attention model,which does not rely on prior knowledge and expectations of human beings and extracts the basic features of image such as color,direction,brightness and texture to obtain salient image.In this paper,three algorithms are analyzed: Itti model based on biological characteristics,FT model based on global contrast and SR model based on information theory and frequency domain.The advantages and disadvantages of the three models are compared.The effects of the three algorithms on building extraction from remote sensing images are analyzed through experiments.The experimental results are analyzed and discussed through precision P,recall R and comprehensive evaluation F value.In the past 10 years,AI has shown blowout-like development.Artificial intelligence is based on a large number of training data.Remote sensing image has a large amount of data resources,which provides support for the development of artificial intelligence.The core of artificial intelligence is the deep learning algorithm,which can obtain high accuracy by using the deep learning method to extract buildings.In this paper,the algorithm principle and hierarchical structure of deep learning network are analyzed,and an improved U-net network based on image segmentation is proposed.In the experiment,the input,structure,edge filling method and convolution core size of the classical U-net network are modified.The CCF data set is used to test the network performance,and the intersection-merge ratio IOU isused to evaluate the accuracy.Based on the above two methods,this paper proposes a new idea to realize automatic building extraction,that is,saliency map extracted by saliency algorithm is trained as training set of improved U-net network.The main programming languages used in this experiment are Python and Matlab.The deep learning framework is keras.The experimental data are captured from Baidu,Google Earth and Sky Maps,and labelme software is used for data label.In the training process,the network parameters,including learning rate,batch size,number and size of convolution kernels,model layers and so on,are continuously fine-tuned.The control variable method is used to analyze the different parameters.The experimental results show that compared with Itti,FT,SR and improved U-net network,the accuracy of the experimental methods in this paper is improved in precision P,recall R,comprehensive evaluation F value,intersection and merging compared with IOU standards.