Research On Building Extraction Of High-Resolution Remote Sensing Images Based On CNN

As the sources of high-resolution remote sensing images continue to increase,image acquisition is becoming easier.The feature information contained in massive data has been widely used in many fields such as map mapping,resource exploration,environmental monitoring,land and resources investigation,change detection and disaster assessment.Buildings are an important part of urban feature information.Manually extracting and updating buildings wastes a lot of human,material and financial resources.Therefore,how to accurately,quickly and automatically extract building information from satellite images has become one of key research directions in the field of remote sensing.In the past,the methods for extracting buildings from remote sensing images were mostly based on artificially designed features,such as texture,spectrum,shadows,and shapes.However,due to different roof coverings,inconsistent structural trends,and different spatial distributions of buildings,which makes these extraction methods unsuitable.In recent years,the rapid development of deep learning has led many scholars to apply it to remote sensing image processing,and has achieved certain results.Therefore,this paper makes deep research and analysis on related algorithms in the field of deep learning semantic segmentation.The main research work is as follows:(1)This article introduces the basic principles and training methods of convolutional neural networks.In order to solve the problem of lack of high-resolution remote sensing image data,the related operations of building a remote sensing image building data set are introduced in detail.It is mainly divided into image preprocessing and image augmentation,including image denoising and image Enhancements,data expansion,and data set partitioning enhance the richness and diversity of the dataset.(2)In this paper,the characteristics of several existing popular network structures are studied and explored.Based on this,a multi-resolution feature fusion semantic segmentation network named MRNet is designed.The network consists of a parallel multi-resolution subnet structure and a multi-scale feature fusion structure.These two structures can achieve parallel training of features at different levels,focusing on buildings of different scales,making the information fusion more diverse.The information flow of feature maps with different resolutions is strengthened,which is beneficial to the reconstruction of lost information and is more efficient than the pyramid structure.Finally,a boundary loss function is proposed,which enables the model to pay more attention to the building boundary during the training process,which effectively improves the boundary jagged problem.(3)Based on the idea of conditional generative adversarial network,this paper designs a generative-segmentation adversarial network named ResUNet-GAN to achieve the segmentation task in an adversarial manner.Generating network ResUNet uses the residual module in the encoder stage,which effectively solves the problem of gradient disappearance during the training process and deepens the number of network layers at the same time,which is conducive to feature extraction.Adding skip connections in the middle is beneficial to multi-scale information fusion.Finally,adding the discriminative network SimNet is used for alternate training.Experiments show that the discriminator has a certain effect on the optimization of the segmentation network,and the segmentation adversarial architecture can improve the accuracy of building extraction to a certain extent.