Semantic Segmentation Of Buildings From High-resolution Remote Sensing Images Of Complex Scenes

Buildings are the main places for human production and life.Spatial information such as the location and distribution of buildings plays an important role in the real-time update of basic geographic databases,urban dynamic monitoring,and urban planning applications.With the development of satellite observation technology,the spatial resolution of remote sensing images has reached the sub-meter level,and the information of building characteristics contained in unit images is more abundant.The buildings in the high-resolution images of complex scenes are of various sizes and have multi-scale characteristics,and the layout of the buildings is irregular and the spatial relationship is complex.Compared with traditional low-resolution remote sensing images,it is more difficult to extract the feature information of buildings in high-resolution remote sensing images.In terms of building feature extraction,traditional methods have poor accuracy,low efficiency,and weak generalization ability.In recent years,the rapid development of deep learning technology has provided a new solution for the segmentation of buildings in complex scene remote sensing images.Using the powerful learning ability of deep neural network,different levels of features can be automatically extracted from sample data,and the hierarchical representation of building features in remote sensing images can be realized.However,existing building semantic segmentation models mostly retain global semantic information,lose local underlying features such as edges and textures,do not take into account the multi-scale characteristics of high-resolution remote sensing images in complex scenes,and have complex forward feature extraction modules.This paper constructs a lightweight end-to-end building segmentation network based on deep neural network combined with multi-scale features and local semantics,which greatly simplifies the steps of building semantic segmentation and achieves accurate segmentation of high-resolution remote sensing image buildings.It also has good robustness and generalization ability in the application of complex scenes.The main work is as follows:(1)In order to enhance the multi-scale characteristics of the network,a multi-scale feature fusion module is introduced between the encoder and decoder to increase the receptive field and extract the salient and average features in the high-order feature map.(2)In order to fully highlight the key local semantic information,an attention-weighted semantic enhancement module is introduced into the decoder of the model to assist the upsampling process by using the shallow features output by each stage of the encoder.(3)The multi-scale feature fusion module and the attention-weighted semantic enhancement module are introduced into the construction of the encoder-decoder model,so that the network model based on the encoder-decoder structure has both multi-scale features and local key semantic features.In order to prove the effectiveness of the overall model,we conduct comparative experiments based on two aerial image datasets,WHU and Inria,and excellent building semantic segmentation models,such as FCN,Deep Lab,etc.The experimental results show that our proposed model is effective on the two datasets.Both have good performance and are ahead of other models in multiple evaluation indicators;on the other hand,we conduct ablation experiments based on the multi-scale feature fusion module and dual attention pooling weighting module based on the WHU dataset,and the experiments show that these two modules are included.Compared with the original model,the proposed model has improved the integrity and continuity of the segmentation of large buildings and the accuracy of the segmentation of small buildings,and is better than the original model in each evaluation index.