| As the main carrier of human activities,buildings are of great significance in urban and rural planning,city construction,terrain monitoring and other applications.The rapid development of remote sensing has promoted the mass popularization of remote sensing images,efficient building extraction has therefore become an important research topic.Traditional methods require artificially designed rules or empirical information,which is not suitable for fast building extraction of massive remote sensing data.Deep learning-based methods can learn the high-dimensional features automatically to achieve end-to-end inference,which has been widely applied in building extraction and has made great success.However,the model will reduce the resolution when extracting high-dimensional features,which may give rise to miss detection of small buildings and blurring building edges.Meanwhile,well-performing models often imply a huge number of parameters and computation,the operational efficiency needs to be improved.Therefore,focusing on building extraction of remote sensing images,the following researches are carried out in terms of improving the extraction accuracy of details and accelerating the inference of building extraction models.(1)For the present,most public building dataset pay attention mainly to single buildings in high-resolution remote sensing images,ignoring the presence of building complex.Focusing on this issue,a brand-new building dataset is constructed on the basis of GF-1 and GF-2 data,where small buildings and building complex in cultivated land are carefully labelled.During partitioning the dataset,a relatively balanced building ratio is ensured for the sake of fairness.(2)The performance of common building extraction methods in details such as small buildings and building edges shall be improved.In this regard,a dual-stream detail-concerned network is proposed to optimize the detail representation of building extraction.ResNet-34 without the first maxpool layer is utilized as encoder to amplify the resolution of the overall features.In the decoder,features are divided into semantic features and detail-concerned features.Then detail refinement modules are used where high-resolution detail-concerned features can compensate for the lack of detail information and building context modules can explicitly enhance building features in semantic features,meanwhile,high-dimensional semantic features can enhance the semantic continuity of detail-concerned features.Detail-oriented hybrid loss are applied to balance the loss weight between building details and non-details using multi-task learning to achieve better performance of building extraction.Comparative experiments and ablations experiments are carried out on WHU aerial building dataset,Inria building dataset as well as self-made building dataset which prove the superiority of the proposed method.(3)To alleviate the problem of low inference efficiency of complicated building extraction models,a new lightweight method for building extraction models based on structural re-parameterization and knowledge distillation is proposed to improve model efficiency.First,the idea of structural re-parameterization is introduced to demonstrate the feasibility of multi-branch merging.Specific modules are therefore able to be losslessly transformed from multi-branch structure at training time to single-branch structure at inference time,ensuring the accuracy and accelerating the inference at the same time.Second,the encoder of the model is simplified and the model is further compressed using probabilistic distillation with regard to edges and multi-level feature distillation,which improves the feature extraction ability and feature utilization ability of the small model.Ablation experiments are conducted on WHU aerial building dataset and self-made building dataset,proving that the proposed method can significantly reduce the number of parameters and computation of the model while retaining high accuracy. |
PDF Full Text Request