Deep Learning-based Building Extraction From Remote Sensing Images With Attention And Multi-Scale Feature Enhancement

With the continuous progress of urbanization in China,the number of urban buildings has increased rapidly.Using high-resolution remote sensing images for fast and automated building extraction is of great significance for urban planning and land management.Due to the large intra-class variance of the foreground and small inter-class variance between foreground and background in remote sensing image building extraction tasks,traditional feature-based building extraction methods cannot achieve end-to-end automation and are influenced by background information,resulting in unsatisfactory extraction results.Building extraction methods based on deep learning can achieve high-precision end-to-end automatic extraction and have gradually become a popular method for building extraction.In this thesis,we conducted related research on deep learning-based remote sensing image building extraction,and the main research work is as follows:(1)To address the problems of large-scale changes,complex backgrounds,and diverse appearances of buildings in remote sensing images,we propose a building extraction network called MAEU-CNN based on non-local and multi-scale feature enhancement.Based on the UNet network,this model uses different scales of image inputs to obtain richer multi-scale image information,and uses parallel irregular dilated convolutions to extract features of different sizes.At the same time,more non-local image information is obtained through dual attention mechanisms,and the optimization of the building contour edge smoothness phenomenon is assisted by a building boundary distance classification auxiliary task.Experimental results show that MAEU-CNN achieves higher extraction accuracy than classical semantic segmentation networks on the WHU Aerial Image dataset and the Space Net Building Extraction dataset,and the extraction results have better contour edge effects.The ablation experiments show that each module in the network has different degrees of improvement in accuracy.Compared with the baseline network UNet,the intersection-over-union(Io U)metric for the extraction results improved by 3.39% and 4.36%on the two datasets,respectively.(2)To address the problem of the large difference in semantic level and semantic information between deep and shallow features in convolutional neural networks,we propose a multi-scale and multi-path network called MMB-Net with boundary enhancement.This network increases the semantic level of shallow features by adding parallel convolutional layers to shallow branches of the network and eliminates the problem of poor feature fusion due to semantic differences through dual attention.At the output stage of the network,deep features are used to improve the semantic level of rough features and remove noise information contained in the rough features.Experimental results demonstrate that MMB-Net has higher extraction accuracy than classical semantic segmentation networks,proving the effectiveness of eliminating semantic differences for building extraction.The ablation experiments show that each module of MMB-Net brings different degrees of improvement in accuracy.Compared with UNet,the Io U metric for the extraction results improved by 2.99% and 4.05% on the two datasets,respectively.