Building Extraction Network for Remote Sensing Images Combining Large Convolutional Kernel and Optimizer
Buildings are of different sizes and shapes with high background similarity,which leads to poor overall recognition ability of the building extraction network for remote sensing images,low multi-scale information extraction ability,and fuzzy boundaries.A 2 U-Net network combining a large convolutional kernel and an optimizer is proposed to improve the building extraction accuracy more effectively.Firstly,a large kernel deep convolutional module is used to construct a U-Net encoder in the feature extraction part,and a larger convolutional kernel is used to improve the sensory field,which solves the problem of overall recognition and multi-scale information extraction.Secondly,to address the low attention to the overall semantic information of the building,a parameter-free attention mechanism is added to the output of the decoder to increase the attention to the building through the weighting function,and to inhibit the expression of unwanted background information.Finally,to avoid direct output of rough feature maps with fuzzy building boundary extraction,a U-shaped optimizer is constructed to improve the building boundary extraction accuracy and optimize the edge detail information.On Satellite dataset Ⅱ(East Asia)and WHU dataset,the evaluation index IoUBuilding reaches 72.04％,90.71％and MIoU reaches 85.19％,94.75％,which improves 2.54％,2.64％and 1.34％,1.51％,respectively,comparing with U-Net.The proposed method also outperforms the existing mainstream methods.The experimental results show that the extraction effect of 2U-Net network is more accurate,which has certain reference value for practical applications.
万方数据
维普
