Research And Implementation Of Remote Sensing Imagery Building Extraction Model Based On Non-local Attention Mechanism

Extracting buildings from remote images is a significant and fundamental task for various practical applications,such as urban planning,land-use statistics,mapping,and national defense security.With the development of aerospace technology,more and more remote sensing data could be accessible easily.Manually extracting buildings from remote sensing images will consume a lot of human resources and cannot meet real-time requirements,limiting the scope of use of building information on remote sensing images.Traditional automatic building extraction methods extremely rely on manual designed features,such as color,shape,texture,etc.However,the shape,size,location,and color of buildings are variable in remote sensing images,meanwhile,the remote sensing image has a complex background and contains a lot of detailed noise information,which decrease the building extraction precision.Therefore,building extraction algorithms based on traditional methods generally do not work well.With the development of artificial intelligence and deep learning,the convolution neural networks,which is a widely used deep learning method,have reached many excellent achievements in computer vision.Convolution neural networks can automatically extract hierarchical features from images and use them for image recognition and semantic segmentation.The successful application of convolutional neural networks in the field of machine vision has attracted many researchers to apply it to the task of extracting buildings from remote sensing images,and there have been excellent results.However,the established building extraction methods based on convolution neural networks use continuous local operations to expand the receptive field,have a weak ability to capture global information,which decrease the precision of building extraction to a certain extent.To solve the above problems,this paper proposes a new encoder-decoder structure of remote sensing image building extraction network and uses non-local blocks to introduce global information to enhance the precision of building extraction.The main highlights of this paper can be summed as follows:(1)Investigate and analyze the current research status in image semantic segmentation and building extraction,introduce the theory and training process of convolution neural networks.(2)An end-to-end encoder-decoder structure building extraction model based on the convolution neural network is proposed.In this paper,a convolution neural network is used to extract features on remote sensing images and produce pixel-level building extraction results.In order to introduce multi-scale and global context information,an atrous spatial pyramid pooling block and a dual non-local block are inserted into the model,which increases the accuracy of the model to extract buildings.(3)To measure the performance of the model,two public remote building extraction datasets,the Massachusetts building dataset and WHU dataset,are used in this paper as train dataset and test dataset.Some established images segmentation models such as FCN,U-Net,Seg Net,and Deeplab v3 are also used as comparison.Experiments demonstrate that the building extraction model proposed in this paper could achieve higher accuracies on the two datasets at a relative higher efficiency.Compared with the existing extraction models,the improvement of F_1-Score on the two datasets is 0.26%and 1.41%respectively,which proves the outstanding building extraction ability of the model proposed in this paper.(4)Explore the impact of different non-local blocks on the model proposed in this paper,the spatial non-local block and the channel non-local block are inserted into the model respectively,and train and test these models on the Massachusetts building dataset.